WorksComputer Science
Pathology Randomization for Synthetic Priors and Zero-Shot Detection
Abstract
Hospital brain imaging archives hold a diverse array of pathology at sample sizes beyond prospective studies, particularly for underrepresented populations and diseases. To unlock uncurated clinical data and scale analysis, methods are needed to detect abnormalities across contrasts and resolutions. Synthetic data has provided robust, contrast-agnostic models for a variety of human neuroimaging tasks in healthy anatomy, but pathology detection and other tasks in clinical settings still require real patient data, contrast-specific models, or disease-specific generators. We propose PathoDraw, a generative model of brain MRI with pathology. A conditional distribution of lesions given the healthy anatomy is drawn into the label map before contrast-agnostic rendering, and a single network estimates healthy anatomy with exact signed intensity deviation of any input scan. The contrast-agnostic 3D UNet learns the healthy labels under the lesion and the signed deviation directly and identifies pathology via robust tissue statistics and learned features, despite no training on real pathology. We show zero-shot detection on public stroke, tumor, and white matter hyperintensity datasets as well as archived clinical data, spanning w, FLAIR and w at various resolutions under a recent unsupervised anomaly detection benchmark protocol. Our single model exceeds or performs comparably to contrast-specific anomaly detectors trained on real scans and keeps healthy anatomy intact. The same generator trains networks for super-resolution and for registration in the presence of pathology. Our model, robust across acquisition and disease, enables the use of uncurated, archived data for clinical analysis. Our code is available at https://github.com/agr78/Pathology-Randomizer.
Code: github.com/agr78/Pathology-Randomizer detection benchmark protocol. Our single model exceeds or performs comparably to contrast-specific anomaly detectors trained on real scans and keeps healthy anatomy intact. The same generator trains networks for super-resolution and for registration in the presence of pathology. Our model, robust across acquisition and disease, enables the use of uncurated, archived data for clinical analysis. Our code is available at https://github.com/agr78/Pathology-Randomizer.
Introduction
Hospital archives hold orders of magnitude more brain MRI than any research cohort from a vast array of scanners, protocol, and diseases. Yet these data are uncurated, spanning a number of possible contrasts, resolutions and pathologies. Analyzing these data requires robustness to all three. Domain randomization over healthy label maps has addressed contrast and resolution variation with contrast- and resolution-agnostic segmentation, super-resolution, and registration of healthy MRI [13, 23, 37, 42, 67]. Pathology remains a critical gap, particularly when the disease population is not predefined. Synthetic pathology generators require real lesion masks or real pathological scans [17, 48, 52, 57, 68, 97] or disease-specific models [16, 34, 72], and state-of-the-art anomaly detectors are contrast specific [5, 29, 30, 46, 58, 59, 61, 62] or require real healthy or patient data [22, 49, 78].
We present a generative model of structural brain MRI with pathology. We model pathology as a conditional distribution of lesions given the healthy anatomy, drawn into the label map and rendered via domain randomization [84, 85]. With such “pathology randomization” we train a single contrast-agnostic network to estimate the healthy anatomy of any scan and where the scan deviates from it, zero-shot, without a real image, lesion mask or patient-derived threshold (Figure 1). We demonstrate robust inference across contrasts, resolutions, and disease status. Our contributions are:

Figure 1. One generator for brain MRI with pathology. (i) Pathology randomization draws lesions into healthy label maps (pink) and renders synthetic scans; (ii) these alone train one network per task; (iii) at inference, real scans with a stroke on w, a glioma on w and white matter lesions on FLAIR (green: ground truth) are processed zero-shot: detection (blue box) of the lesion as a signed deviation from the healthy anatomy predicted beneath it, super-resolution (green box) from 3 mm slices (6 mm for the white matter lesions) to 1 mm, and registration (pink box) onto the MNI152 atlas (cyan: atlas brain) with the lesion preserved. No real image, lesion mask or threshold is used in training.
a disease-agnostic generative model of structural pathology for domain randomization: lesions drawn into healthy label maps, conditioned on the anatomy, with no real image or mask of any disease, and rendered contrast-agnostically (§ 2.1) and exact supervision (§ 2.6);
zero-shot lesion detection on four datasets, three contrasts and 37 sites under the unsupervised anomaly detection (UAD) benchmark protocol (§ 4);
generalization to zero-shot super-resolution and registration in the presence of pathology.
Related work
The use of synthetic priors for healthy data is well-established. Segmentation [13, 67], super-resolution [42] and registration [37] networks have been trained on renderings of healthy label maps, with extensions requiring real pathological data [13, 42]. Synthetic data has also enabled feature learning [23, 24], yet the label prior is healthy or requires real lesion segmentations [12, 48]. PathoDraw, in contrast, keeps the prior synthetic and adds the pathology to it.
Segmentations from real pathology have informed synthetic data, as masks pasted into healthy label maps, labels under target-sequence priors, patient-fitted intensities, labeled susceptibility maps, or seeds for randomization [17, 48, 52, 56, 57, 68, 74]. PathoDraw, in contrast, needs no segmentation of real pathology: shape, placement and intensity are drawn from the prior.
UNA generates anomalies on label maps by advection–diffusion from real stroke annotations and noise seeds, and trains on real stroke scans among its synthetic ones. PathoDraw draws the lesion into the label map, conditioned on the anatomy and with mass effect, before rendering; supervises the exact signed deviation between the two renderings of one draw and the healthy labels beneath; and admits no real image, mask or threshold. Additionally, PathoDraw is evaluated on stroke, tumors, and white matter hyperintensities, enabling super-resolution and registration from the same generator.
Disease-specific generators model one process, such as corpus-callosum malformations (PIDR [34]), biomechanical hydrocephalus (SynHydro [72]) or procedural hemorrhage [16], needing no real pathology but limited to the disease of interest. PathoDraw is tied to no disease process; one prior serves stroke, glioma and white matter hyperintensities.
Like our method, anomaly detection does not require real pathology data. Feature-space reconstruction (FAE [59]), cold-diffusion restoration of painted anomalies (DISYRE [61, 62]), diffusion-latent scoring [58], patch-location prediction [5], disentangled anatomy [95] and training-free foundation-model scoring [50] generalize across pathologies, training on real healthy images and primarily using one model per contrast [9, 29, 46]. That painted-anomaly line, from patch interpolation [79, 81, 82] to pasted or textured patches [8, 51, 96] and DISYRE, corrupts real healthy images in pixel space and binds each model to one contrast; scoring voxels as outliers of a per-class tissue model goes back to [88]. Diffusion priors over real healthy images solve detection as a per-scan inverse problem (dmprior [49]) or, in 2D with one prior per contrast, with a jointly sampled latent mask [78], healthy counterfactuals produce anomaly maps when subtracted.
None of these methods accept arbitrary clinical scans. The anomaly detectors generally require one model per contrast. Synthetic-pathology generators either require real pathology data or model one disease process. To our knowledge, no prior generator expresses entirely synthetic, disease-agnostic structural pathology, formulated as a domain-randomization problem.
Method
Our framework consists of a generative model of a structural brain MRI scan with pathology. We begin with a model of healthy anatomy, a conditional distribution of pathology given that anatomy, and a renderer that turns the combined label map into an image under a randomized acquisition. We define the model (§ 2.1 to § 2.4), the anatomy factorization the network estimates (§ 2.5), and a network fit to recover it from samples of the model, since the posterior itself is intractable (§ 2.6).
Generative model
A scan is generated from a healthy label map , a pathology and an acquisition , with joint distribution
The factorization follows from two independence assumptions, and (Appendix § 6.§ 1). Here is the empirical distribution of healthy segmentations under spatial augmentation (§ 2.2). The distribution places a lesion, with its own labels and a mass-effect deformation, on that anatomy (§ 2.3). The prior is uniform over wide ranges of tissue contrast, resolution, bias field and noise (domain randomization [13]). Simulated renders the label map with the pathology injected (§ 2.4). Rendering the same under the same with and without gives a pathological image and its control, the supervision of § 2.5. Every probability and range below is fixed a priori (Table S1, Appendix § 6).
| Component | Setting | Provenance |
| Anatomy | 1,033 healthy label maps (HCP 821, Chinese HCP 212)†, 56 classes*; crop at a random center‡; rotation , shear , scale per axis, nonlinear field with standard deviation voxels on a 3–6% grid‡; left/right flip with label swap, ‡; extracerebral block removed, | SynthSeg [13]; HCP [87], Chinese HCP [31] |
| Modes of | primary mode: local 0.23, vascular-aligned 0.385, tract-aligned 0.385* (diffuse and atrophy modes off); each other mode adds a second pathology with * (28 % of draws carry modes; co-occurrence†) | [21, 90] |
| Mode labels | local: any brain label present (uniform over labels)†, ROI bias *, confined to that label *, 1–3 foci with weights 0.85 / 0.12 / 0.03†; vascular-aligned: perforator branch (thalamus 7/24, caudate 8/25, putamen 9/26, pallidum 10/27) or cortical branch (cerebral white matter 1/18, cortex 2/19)†, * each; tract-aligned: one of lateral and inferior-horn (3, 4 / 20, 21) or third and fourth (11, 12) ventricles (*) or all of them, with cerebral white matter (1/18)† | [6, 25, 26, 47]; foci [66] |
| Geometry | excursion sets of a Gaussian random field with a spectral-mixture power spectrum†: 1–9 zero-centered Gaussian components, bandwidths cycles/voxel, (local) or (vascular, tract)*; anisotropic warp, elongation per axis, rotation *; gated by a focal prior (local: Gaussian envelope, voxels; vascular: blob for the perforator branch, wedge or column for the cortical branch; tract: periventricular white matter band, tangent slab or Gaussian envelope)*; the field’s top (local, vascular) or (tract) of the prior becomes lesion* (thresholded noise†); core/shell ring always for local and tract, never for vascular*; three GMM intensity zones per lesion, cut by a second field with components centered at and scale voxels ( lesion half-width), one zone when * | [1, 94]; thresholded noise [57, 96] |
| Mass effect | smoothed, normalized gradient of the lesion potential (blur voxels, displacement smoothed with *); per mode (tumor and infarct mass effect†): local , magnitude , random sign; vascular , ; tract , , random sign*; total clamped at 26.25 (fold-free) | tumor [19, 69]; infarct [7, 73] |
| Rose criterion | criterion† with *; standard deviation of the control rendering over a 3-voxel-wide band starting 3 voxels outside the lesion (see § 2.4; the estimated noise standard deviation when the band has voxels*); a lesion whose mean rendered change is below is rebuilt to it (floor), one above is scaled down to it (ceiling), so every drawn lesion satisfies the criterion against its surround; the geometric lesion label is kept in full* | [77] |
| Sign | lesions rebuilt at the floor: where white matter renders brighter than cortex, 0.85 otherwise† (flipped when only the other side has of headroom)*; lesions above the floor as drawn: multi-zone lesions: one sign per lesion, *, every zone mean re-anchored to that side of the host class mean with its drawn magnitude, zone standard deviation host standard deviation* | [15, 32] |
| Composition | with (vascular / local / tract)* each zone renders at with the host standard deviation (host = the label the lesion overwrites, modulating class = CSF or GM)†: vascular core CSF , shell GM ; local core CSF , shell GM ; tract GM (classes†, ranges*); core = erosion to 15–50%*; applied only if , otherwise the Rose-gated zone above*; exempt from floor and ceiling* | [2, 86, 92] |
| Rendering | class mean , standard deviation , background mean zeroed with ‡; dark CSF on half the draws*, †; partial-volume ramps†; synthesis in white matter units*; gamma expo *; multiplicative bias field, log-standard deviation on a grid‡; resolution isotropic mm or one axis mm (equiprobable), slice thickness = resolution, blur standard deviation thickness‡; acquisition noise standard deviation *; per-image max normalization‡; each draw trains three steps‡: as rendered, then twice re-rendered with extra noise standard deviation and gamma * | SynthSeg [13]; partial volume [89]; slice profile [11]; bias [80]; noise [35]; dark CSF [36, 60, 76, 91]; three steps per draw [18] |
| Shared per step | subject, crop, warps, flip, class table, synthesis noise, gamma, bias, resolution, noise; identical for the control and the pathological rendering; only the label map (lesion, mass effect) and the visibility correction differ (common random numbers†) | [33] |
| Target | , of the pathological rendering†, on the lesion support with sign-matched floor *, zero elsewhere and zero on the control rendering* | § 2; intensity frame [13, 37, 42] |
Table S1. Table S1. Generator summary: the conditional pathology distribution and the acquisition of the reported model (§ 2). Every entry is a fixed range or probability; no value is set on a real image. Modes are drawn as Bernoulli trials: one primary mode, then each other mode adds a second pathology independently with probability 0.15. Label ids follow the 0–56 scheme of the training label maps; is the logistic function. Marks on the values: † literature value; $‡ SynthSeg default.
Model of healthy anatomy
We assume a dataset of 3D segmentations of healthy subjects into labels (§ 3). A draw is one of them under a random spatial transform,
where composes a random affine transform, a smooth nonlinear field and a left–right flip with a crop at a random center, as in [13] prior to intensity draws.
Pathology randomization
The conditional distribution is a mixture of three disease modes chosen to span the geometry of structural brain pathology rather than any one disease [92]. We introduce tract-aligned lesions in the periventricular white matter (small-vessel disease), vascular-aligned lesions in the deep gray nuclei or along a cortical territory (infarcts), and local lesions in any brain label (tumors, focal lesions), amenable to mode additions.
to mode additions. A pathology is constructed by a sequence of draws from random variables. We term this pathology randomization method PathoRoaD. A primary mode is drawn; each remaining mode then adds a second pathology, independently, with probability .
with the quantile that keeps a random fraction of the envelope. Eroded into a core and a shell, each a new label , the pathology is injected by “overwriting,”
a hard replacement of labels rather than a blend of images, since every intensity is drawn afterwards (§ 2.4). Mass effect [69] displaces the surrounding anatomy along the smoothed, normalized gradient of the lesion field, , with a mode-specific probability setting a maximum queue depth, clipped to a fold-free range. Thus , with support the lesion label of § 2.5. The control label map is the anatomy . Figure 2 shows one draw with details in Appendix § 6.§ 2.

Figure 2. Pathology randomization and the network, one draw of the generative model. The top row shows the healthy labels, the drawn mode and its eligible region, the random field whose excursion set is the lesion, the injected label map (pink) and the rendering (Table S1, Appendix § 6). Below, the rendering enters the contrast-agnostic 3D UNet backbone, whose anatomy head predicts the healthy labels and deviation head the signed deviation. The same draw passes through the task networks, super-resolved from 6 mm slices and registered to another draw’s healthy rendering, with the warped lesion (pink), target brain (cyan) and Jacobian determinant.
Rendering
The acquisition holds one mean and one standard deviation per label, a bias field, a resolution and a noise level, each drawn uniformly over a wide range [13]. With the pathological label map, the image is rendered one stage at a time,
so that is a Gaussian mixture conditioned on the labels alone, with the bounds , , , of SynthSeg [10, 13] for every label . is the 1 mm rendering with partial-volume ramps at label boundaries and a smooth multiplicative bias field . The operator blurs with the slice profile of resolution (isotropic, or thick slices along one axis), subsamples it to and resamples back to 1 mm, and is acquisition noise. The control image at 1 mm is rendered from with the same , and , so the two renderings differ only where the labels differ. Lesion labels obey the following conventions. (i) They are rendered by composition between the host tissue they overwrite and a modulating class, CSF or gray matter, with uniform in a per-mode range, so that lesion polarity follows the drawn contrast, dark where white matter renders bright (-like draws) and bright otherwise (FLAIR-like draws). (ii) Every lesion satisfies the Rose criterion [77], in that its mean rendered change against the control over is held between and , , with the control’s standard deviation in a band around the lesion, so that no in-visible lesion becomes a training target (Appendix § 7). (iii) Dark-CSF contrast [36, 60, 76, 91] on half of the draws (Appendix § 8).
Anatomy factorization
The network factorizes a scan into the healthy anatomy as if the pathology were absent and a deviation map of where and how the scan departs from it. Lesion detection is posed as a global thresholding rule on that map. The anatomy target is , the healthy labels under the lesion. The deviation target is built from the two renderings of one draw, , in the scan’s own intensity frame , the percentile range synthetic-data methods rescale their inputs by [13, 37, 42]. On the lesion support , , and outside it, with signal floor from the Rose criterion (Appendix § 7). The magnitude is a soft lesion mask that never vanishes inside a lesion drawn as visible, its sign the lesion’s polarity, and both are zero on the control image.
Inference with a neural network
The posterior of anatomy and pathology under Eq. (1), , is intractable, since the acquisition must be marginalized. Sampling the forward model, however, is tractable, and every draw comes with respective latent variables: the healthy labels and deviation of § 2.5, the 1 mm rendering, and the deformation between two draws, which are the targets of detection, super-resolution and registration. Tasks like detection, super-resolution, and registration, whose targets are such draw-derived quantities, each with a loss , can therefore be learned by amortized inference, a network fit on draws,
which at the optimum is the Bayes estimator of under [20]. The generative model is the same for every task, only the pair changes, and no real image is used. The targets are of § 2.5, the 1 mm rendering, and the label overlap between two draws, with the losses , and given in § 2.7.
Network. is a 3D UNet (Figure 2) after [13] with identity-initialized attention gates on the skips [64, 65]. The same backbone serves every task behind a task-specific first convolution and head. For detection the head is dual. The anatomy head ( classes, softmax) predicts the healthy label map, and the deviation head is a two-layer MLP on the 24 decoder features plus , the input rescaled by the median and median absolute deviation (MAD) of its anatomy-head class (Appendix § 9). The two task gradients on the backbone are balanced per step [54]. The detection readout is the deviation head’s magnitude . The super-resolution head predicts a residual on the input and a per-voxel amplitude in that scales it, and the registration head a diffeomorphic deformation field, each with its final convolution zeroed to start at the identity.
Controls. Every task has a control whose backbone was trained without pathology in the draws, so that pathology enters only through the training data (Appendix § 10).
Task losses
Detection. Over the batch holding the control and the pathological image of one draw, , with [13, 43, 48] on the anatomy target and , an loss on the deviation head: on the pathological rendering, weighted to balance the lesion against the background and to stress the border band where a detector oversegments, and on the control, , where the deviation must vanish. The cross-entropy is weighted likewise inside the ROI, the healthy structures the lesion overlaps; the weights are fixed a priori after [41, 75] and written out in Appendix § 11.
Super-resolution. For the network predicts a residual on the generator’s own degraded image using operator , with the 1 mm rendering before the acquisition stage, and is fit with an loss, over the brain mask of , where and is scaled by the same percentiles of the input, as in § 2.5. At test time every scan is resampled to 1 mm before the network, so any voxel size is accepted.
Registration. For the network takes the moving and fixed image for two channels and predicts a stationary velocity field , integrated by scaling and squaring into a diffeomorphic field . It is fit with SynthMorph’s loss [37, 38], with , the soft Dice taken over the brain labels present in the pair and the forward-difference gradient of the displacement in voxel units over the volume . The pair is two independent draws under unique spatial augmentation (Appendix § 13). The moving image is the pathological rendering with probability and the fixed image is the control rendering. The control trains likewise on healthy renderings only.
Experimental setup
Anatomy source
We draw healthy anatomy uniformly from 1,033 label maps of young adults of the Human Connectome Project (HCP-YA [87]) and the Chinese HCP [31], segmented with FreeSurfer [27] under manual quality control, augmented with the claustrum and limbic structures by the label pipeline of [83] and with extracerebral labels from SAMSEG [70] and CHARM [71], merged to the 56-label scheme of the anatomy head. No image of these subjects enters training.
Test data
We infer on the following datasets: ATLAS [53] (acute to chronic stroke, ), the benchmark’s unseen-scanner subset ATLAS-O (, nine scanner cohorts). BraTS [3] ( from 24 sites [4], , FLAIR, co-registered) is the 2023 adult glioma pool, scored as whole tumor on every modality [29]. WMH [44] () has three sites, with masks delineated on FLAIR and reported on by the benchmark [29]. PRES (posterior reversible encephalopathy syndrome, ) is a hospital-archive cohort, clinical FLAIR at acquired mm, with manual masks. These scans were acquired during routine clinical care and analyzed retrospectively as secondary data under Institutional Review Board approval at our institution, with informed consent obtained. ADNI3 [93] ( per contrast, and FLAIR; 132 cognitively normal, 199 impaired) is a neurodegenerative cohort without focal lesions.
Evaluation
Performance metrics. The published benchmark’s primary metric [29] is the ceiling Dice (Table 1), a global threshold per dataset chosen on the scored cases (an oracle, hence an upper bound), 4 mm median filter, per-case Dice averaged over every case inside a SynthStrip brain mask [40], as FAE, DISYRE and dmprior are reported. Specificity is measured on healthy subjects (§ 4.2).
| Dataset | PathoDraw | Control (no pathology) | dmprior (no pathology) | FAE | DISYRE |
| Single model | ✓ | ✓ | ✓ | × | × |
| ATLAS-O | 0.369 | 0.184 | 0.107 | 0.167 | 0.386 |
| BraTS | 0.537 | 0.228 | 0.129 | 0.437 | 0.274 |
| BraTS FLAIR | 0.622 | 0.494 | 0.214 | – | – |
| BraTS | 0.657 | 0.462 | 0.185 | 0.467 | 0.646 |
| WMH FLAIR | 0.259 | 0.255 | 0.039 | – | – |
| WMH | 0.229 | 0.072 | 0.021 | 0.032 | 0.130 |
| PRES FLAIR | 0.225 | 0.205 | 0.123 | – | – |
Table 1. Zero-shot detection on the UAD benchmark [29], with (§ 3), “–” when not evaluated. PathoDraw, the control (trained without pathology randomization, Appendix § 8) and dmprior on the benchmark (Appendix § 16), PRES on every case; paired -values in Table S8. Bold: the row maximum and any value inside PathoDraw’s 95 % bootstrap interval of it (Appendix § 17).
Downstream evaluation. Super-resolution is scored on ATLAS, acquired-isotropic BraTS and and WMH FLAIR, degraded threefold (Table 2). Registration is scored on ATLAS, BraTS and WMH to MNI152 [28] (Table 3; matched-displacement control in Table S6, Appendix § 13).
| Tricubic interpolation | Control (no pathology) | dmprior (no pathology) | ||
| ATLAS-O (210) | Brain | +0.96±0.03 | +0.32±0.18 | +4.74±0.03 |
| Lesion | +0.25±0.03 | +0.60±0.28 | +5.76±0.05 | |
| BraTS (690) | Brain | +0.79±0.17 | +0.47±0.47 | +5.20±0.16 |
| Lesion | +0.24±0.03 | +0.74±0.49 | +7.47±0.03 | |
| BraTS (358) | Brain | +1.09±0.08 | −0.18±0.10 | +3.52±0.08 |
| Lesion | +0.44±0.07 | +0.10±0.22 | +7.55±0.07 | |
| WMH FLAIR (60) | Brain | +0.62±0.02 | +0.12±0.06 | +2.83±0.02 |
| Lesion | +1.48±0.22 | −0.02±0.41 | +8.55±0.22 |
Table 2. Super-resolution. Mean PSNR margin (dB) of detector-modulated PathoDraw over the tricubic input, the control trained without pathology randomization, and dmprior; mean ± standard deviation over seeds; modulation on degraded inputs (BraTS: FLAIR); dmprior on its completed cases (170 / 550 / 358 / 55); bold: significant gain (Table S8).
| ATLAS-O | BraTS | WMH | |
| PathoDraw | 0.757±0.019 | 0.925±0.011 | 0.564±0.037 |
| Control (no pathology) | 0.710±0.025 | 0.911±0.013 | 0.502±0.021 |
| SynthMorph | 0.644 | 0.782 | 0.384 |
Table 3. Lesion Dice through the warp, mean ± standard deviation over two seeds.
Table S6. Pathology in registration, against its control. Ours: the pathology-trained backbone fine-tuned for 100k steps with the pathological rendering on half of the steps. Control: the control model without pathology randomization under the identical fine-tune without pathology randomization, its velocity scaled by 0.97 to match our mean displacement (1.53 against 1.56 voxels). Lesion Dice as in Table S7, with the percentage of cases in which ours is higher; brain Dice, and folds as there. Bold: the better of the pair; neither where the two are within 0.002. One seed, 1,964 cases.
| Lesion Dice | Brain Dice | Folds (%) | |||||||||
| Ours | Control | (%) | Affine | Ours | Control | Ours | Control | Ours | Control | ||
| ATLAS | 655 | 0.693 | 0.688 | +0.005 (48) | 0.921 | 0.925 | 0.920 | 0.263 | 0.189 | 6.4 | 23.1 |
| BraTS | 1,249 | 0.917 | 0.922 | −0.005 (26) | 0.937 | 0.940 | 0.939 | 0.223 | 0.180 | 5.6 | 35.1 |
| WMH FLAIR | 60 | 0.537 | 0.525 | +0.012 (57) | 0.921 | 0.924 | 0.920 | 0.248 | 0.183 | 20.0 | 40.0 |
| All | 1,964 | 0.831 | 0.832 | −0.001 (34) | 0.931 | 0.934 | 0.932 | 0.236 | 0.183 | 6.3 | 31.3 |
Table S6.
Baselines
Detection. We compare to FAE and DISYRE published by the benchmark [29] as contrast-specific models. The benchmark-trained no-FAE model (“–” in Table 1). UNA [57] and PEPSI [56] train on real pathological data and are not compared. No model in Table 1 sees real pathology in training. The control, the same UNet trained on healthy label maps only, with the tissue-class statistic as readout, is the ablation of the pathology itself (Table 1). The data regime is ablated likewise: the same network trained on DISYRE’s pasted anomalies or on FAE’s healthy scans (Appendix § 8). Contrast-agnostic dmprior [49] runs with its multicontrast prior at default , input minus refinement as the deviation map, with joint super-resolution and refinement on anisotropic scans.
Super-resolution. dmprior’s super-resolution task with a contrast-specific prior [49] serves as a reference trained on healthy data only.
Registration. SynthMorph denotes the published pre-trained model [37, 38], registered from the same affine. The same network trained under its protocol on our pool without pathology is the ablation of Appendix § 10.
Training
Detection. Trained from random weights for 500k steps with Adam at learning rate , two seeds; the last 100k steps decay the learning rate by cosine to and keep an exponential moving average of the weights (decay 0.9999), from which the reported model is read. A step holds the control rendering of one draw and, on half of the steps, its pathological rendering in the same batch (§ 2.7); the pathology-free control omits the second rendering and is otherwise identical. Super-resolution. The pathology-trained backbone at 300k steps, fine-tuned for 200k steps on healthy renderings with the calibrated residual head of § 2.7 on the tricubic interpolation of the degraded draw; its control is the same backbone trained without pathology, under the identical fine-tune. Registration. The same backbone fine-tuned for 100k steps with the velocity head of § 2.7, the moving image pathological on half of the steps; its control is the same backbone trained without pathology under the identical fine-tune without pathology randomization. Every reported model is one set of weights: the EMA weights at 500k for detection and the final checkpoint of each fine-tune. Hyperparameters and ablations are in Appendices C and I.
Results
Zero-shot detection benchmark
Our model exceeds or matches benchmarked approaches on four datasets (Table 1, Figure 3). is the benchmark’s oracle quantity, the same for every column; operating points are set on healthy scans in § 4.2. Lesion size and contrast bound absolute Dice with median volumes of 5 mL on ATLAS, 15 mL on WMH and 89 mL on BraTS. WMH masks are drawn on FLAIR; on those lesions are difficult to visualize. On ADNI3, a neurodegenerative cohort, the anatomy head agrees with the control’s for cognitively normal and impaired subjects alike (Appendix § 15).

Figure 3. Exemplar cases. One model on test scans from four diseases, three contrasts and 37 sites; one column per dataset (ATLAS , BraTS /FLAIR/, WMH FLAIR/, clinical posterior reversible encephalopathy syndrome (PRES) FLAIR), with the input image, ground truth (green) and our segmentation (pink), and the signed deviation map, at protocols from 1 mm isotropic to 6 mm slices.
False-positive analysis with healthy subjects
We measure the deviation map on healthy scans, with thresholds set on HCP-Aging [14] [29]. PathoDraw is comparable to the published models at matched specificity (Table S9, Appendix § 18.§ 1). We analyze false positives in pathological datasets in Appendix § 19.
| Published [min, max] | At FAE’s rate (0.82%) PathoDraw | Level from | At DISYRE’s rate (0.12%) PathoDraw | Level from | |
| ATLAS-O | [0.17, 0.38] | 0.36 [0.33, 0.40] | 0.04% | 0.27 [0.24, 0.30] | 0.29% |
| BraTS | [0.26, 0.44] | 0.42 [0.40, 0.44] | 0.78% | 0.16 [0.15, 0.18] | 0.21% |
| WMH | [0.02, 0.11] | 0.23 [0.18, 0.28] | 0.03% | 0.10 [0.07, 0.13] | 0.10% |
| Mean | [0.15, 0.31] | 0.34 | 0.18 |
Table S9.
Task generalization
Super-resolution. Each acquisition is degraded along its slice axis, super-resolved by PathoDraw and dmprior (Figure 4) and scored with peak signal-to-noise ratio (PSNR). The residual is scaled by , the deviation complement on the degraded input (Appendix § 12). PathoDraw improves tricubic input over the brain and lesion. Over the control trained without pathology randomization it is ahead on and level in the lesion on and FLAIR (Table 2). Pathology does not conform to the piecewise-constant label model the renderer assumes, so the correction is scaled according to the deviation map, which bounds the margins. Despite learning exclusively from synthetic tissue rendered from piecewise-constant labels, the network still learns a mapping from interpolated low-resolution to high-resolution images.

Figure 4. Super-resolution. Coronal slices with lesion outlines in green. Each row shows the interpolated input, PathoDraw (detector-modulated, as in Table 2), the healthy diffusion prior dmprior [49] at its contrast-specific super-resolution task, and a callout (yellow) on the lesion; the retrospective half (left of the dashed line, degraded slices, the clinical regime of the native thick-slice acquisitions on the right) adds the acquisition as target. Arrows point inside the lesion at the detail under comparison, where the healthy prior erodes or obscures pathology. The white matter hyperintensity row is degraded from its own 3 mm acquisition, which is its target.
Registration. Against pretrained SynthMorph at its default regularization our field preserves the lesion, measured as its overlap with its affine source (Table 3; Table S7, Appendix § 13.§ 1). The control (no pathology) preserves less
Table S7. Pretrained SynthMorph [37, 38] (mri_synthmorph, FreeSurfer 8, default ) against our field on every ATLAS, BraTS and WMH case with a lesion mask (1,877 pairs), registered to MNI152 from the same affine. Lesion Dice: overlap of the lesion carried through the field with its affine-only position, mean, with paired difference and Wilcoxon . Brain Dice: overlap of the warped brain mask with the template’s, mean; Affine is the same before the deformable step. : mean displacement in the brain (voxels). : first percentile of the Jacobian determinant in the brain, median over cases. Folds: cases with a negative determinant. Bold: the better of the two.
| Lesion Dice | Brain Dice | Folds (%) | ||||||||||||
| Ours | SM | Affine | Ours | SM | Ours | SM | Ours | SM | Ours | SM | ||||
| ATLAS | 639 | 0.693 | 0.588 | +0.105 | 0.921 | 0.925 | 0.925 | 1.46 | 2.70 | 0.261 | 0.420 | 6.4 | 0.0 | |
| BraTS | 1,179 | 0.915 | 0.782 | +0.133 | 0.940 | 0.943 | 0.963 | 1.61 | 2.77 | 0.225 | 0.447 | 5.5 | 0.0 | |
| WMH FLAIR | 59 | 0.537 | 0.384 | +0.153 | 0.922 | 0.924 | 0.916 | 1.48 | 2.62 | 0.247 | 0.432 | 20.3 | 0.0 | |
| All | 1,877 | 0.827 | 0.703 | +0.124 | 0.933 | 0.936 | 0.949 | 1.55 | 2.74 | 0.238 | 0.437 | 6.3 | 0.0 |
Table S7.
of the lesion on every dataset. As smaller fields favor lesion Dice, Appendix § 13 repeats the comparison at matched mean displacement. When comparable, PathoDraw offers a better-conditioned Jacobian. Figure 5 shows lesion preservation where SynthMorph distorts it.

Figure 5. Registration to MNI152. Four cases at axial lesion slice with the template, source, warps and Jacobian determinants of our field and SynthMorph at default regularization (Table 3). Template brain mask in cyan, yellow arrows show lesion distortion in SynthMorph.
Discussion
Our generative model of pathology randomization over healthy label maps suffices for a single contrast-agnostic network robust to a variety of diseases zero-shot. Despite training purely on synthetic data, real lesions lie within the generator’s shape manifold at high recall 0.86–0.93 (Appendix § 20.§ 1). The modes of pathology used in randomization may be broadened to improve the estimated manifold.
Limitations. Synthetic pathology is randomized over shape, placement and intensity rather than disease, so detectable conditions must lie on this manifold.
Conclusion. We demonstrate our model trained with pathology randomization can detect stroke, glioma and white matter hyperintensities zero-shot, on public data and archived clinical scans. Our backbone extends to super-resolution and registration, and to any task whose target the forward model supplies. We release PathoDraw to expand the utility of uncurated clinical data and accelerate disease decoding in neuroimaging.
Generator
Table S1 contains every range of the generator parameters and their sources, and the mode label mapping. Below we describe the Rose criterion sampler and mass effects.
Derivation of the joint distribution
The factorization of Eq. (1) follows from the chain rule and two independence assumptions that the generator satisfies by construction. By the definition of conditional probability,
and . The acquisition is drawn without reference to the anatomy, , so and Eq. (8) becomes
Substituting Eq. (9) into Eq. (7),
The pathology is drawn from the anatomy alone, before any intensity exists (§ 2.3), so and , which gives
Eq. (1). Every lesion quantity that does depend on , the composed class means and the Rose-criterion rebuild of § 2.4, is part of the rendering and therefore of , not of .
Pathology randomization
Modes. Drawn modes consist of tract-aligned (periventricular: seeded from the ventricles into the white matter), vascular-aligned (deep gray nuclei, or cerebral white matter and cortex), and local (any brain label) [92]. Within these modes, pathology geometry is conditioned on the anatomy through each mode’s eligible region and envelope (Eq. (3), mode labels row of Table S1), shown per mode in Figure S1. Mass effects are computed by a smoothed gradient of the lesion potential. Each draw is a chain of independent trials, beginning with a primary mode over (local, vascular-aligned, tract-aligned). Every other mode is added by an independent draw. There may be one to three foci in the local mode, a perforator or cortical branch ( each) for the vascular-aligned mode and a periventricular band seeded from one ventricle group for the tract-aligned mode, the lesion inside any envelope being the excursion set of Eq. (3), which can fall into several components in every mode, so no mode is limited to one focus. Each mode adds mass effect with probability 0.8 / 0.4 / 0.2 per mode (local / vascular-aligned / tract-aligned). The composition rendering below is applied with probability 0.5 / 0.5 / 0.3 per mode. Table S1 lists every probability and range.

Figure S1. Pathology modes of : one generator draw per column (local, vascular-aligned, tract-aligned, and their combinations) and one stage per row (healthy labels , the mode’s eligible region, the injected draw in pink, the control and pathological renderings under one acquisition draw, and their difference).
Intensity. Every label renders as a Gaussian (Eq. (5)), including lesion labels. With its mode’s composition probability (0.5 / 0.5 / 0.3 above) a lesion is rendered by composition rather than with a label mean of its own. Otherwise, a lesion renders as an ordinary label under the Rose criterion below. A zone of the composed lesion, its core or its shell, retains the host’s and takes the mean , where the host is the healthy label the zone overwrites and the modulating class a second healthy label whose mean modulates the zone’s. The acquisition draw already holds one mean per label (Eq. (5)), and the zone reads the host’s and the modulating label’s from it. The modulating label is fixed by the zone, CSF for a core (cavitation, encephalomalacia, necrosis) [2, 86, 92] and gray matter for a shell or a tract-aligned lesion (gliosis, WMH) [92]. Only the fraction is drawn, per zone, from the ranges in Table S1. A zone whose composition is invisible against its host () falls back to the ordinary path. As both means belong to the same , the zone is darker or brighter than its host exactly as CSF or gray matter is in that draw, so the lesion polarity follows the drawn contrast, and the composition is contrast-agnostic by construction.
Rose criterion [77]. Every drawn lesion must be visually perceptible with respect to its surrounding tissue as defined by contrast-to-noise ratio , where is the lesion’s mean rendered contrast and the standard deviation of the control rendering in a 3-voxel band around it. To randomize contrast sufficiently the required CNR itself varies, , and the renderer raises a lesion below to it and lowers one above to that ceiling (Table S1). Composition lesions are exempt from this floor and instead gated at against their host class. Figure S1 shows one draw per mode and combination and Figure 2 follows one draw through every stage of pathology randomization.
Region-target floor
The Rose criterion admits a lesion whose mean contrast is at least with (§ 2.4). The smallest deviation the renderer calls visible is therefore (), which in target units (§ 2.5) is . Its median over 200 pregenerated draws is 0.184 on one set and 0.182 on an independent second set, hence
A lesion voxel whose rendered contrast falls below this floor is still labeled lesion by the pathology model, so its target is raised to the region target with the voxel’s own sign, or the lesion’s mean sign where .
Ablation details
Development protocol. Attention gates on the skip connections [64, 65] and dark-CSF rendering [36, 60, 76, 91] were adopted using robustness to distribution shift in synthetic data alone. Distribution shift was simulated on a validation set re-rendered with every appearance and nuisance range of the generator widened to twice the training ranges. The model with both components is above the control model in both tested seeds (shell AUC +0.016 [+0.005, +0.026] and +0.059 [+0.047, +0.072] over 566 cases with a visible lesion). Neither component alone separates. Table S3 gives the per-dataset gains on the real data.
Table S3. Ablation of dark CSF and attention on six datasets from a common backbone at 300k to 400k following distribution shift robustness in synthetic validation data. Reported as [Dice] (median-filtered, the benchmark’s convention) as mean standard deviation over the two seeds.
| ATLAS | BraTS | BraTS FLAIR | BraTS | WMH FLAIR | WMH | |
| + attention gates | +0.018 0.019 | +0.012 0.002 | +0.071 0.032 | +0.019 0.016 | +0.056 0.064 | +0.054 0.037 |
| + dark-CSF rendering | +0.029 0.020 | +0.087 0.024 | +0.095 0.047 | +0.062 0.054 | +0.127 0.050 | +0.005 0.036 |
| + both (reported) | +0.039 0.026 | +0.029 0.009 | +0.147 0.059 | +0.063 0.036 | +0.067 0.046 | +0.040 0.049 |
Table S3.
Data-regime ablation. As the pathology randomization generator is our contribution, the data regime is ablated on the same architecture, loss, healthy pool and step budget, every model trained to 500k and read out with batch-normalization statistics of the input scan (Table S2). This ablation contains (a) synthetic renderings with pathology randomization, the reported model, (b) the DISYRE analog, real healthy scans with a foreign patch of another real scan pasted in as the anomaly [29], read through the deviation head, (c) the FAE analog, real healthy scans with no anomaly, read through the tissue-class statistic, the readout of the control of Table 1, and through its own head, which has no deviation to learn. The two real-image regimes keep each scan’s acquired contrast and skip contrast randomization and dark-CSF rendering; only the acquisition stage after rendering still applies, so neither sees a FLAIR scan, the pool’s real contrasts being and . Pathology randomization leads on ATLAS and WMH against both real regimes and is within 0.02 of the healthy-scan statistic on the two FLAIR datasets, ahead on BraTS and behind on WMH. The pasted patches lead on BraTS and . We expect this to follow from tumors consisting of abnormal growth of the tissue around them. On and contrasts a tumor presents largely as near-normal tissue in the wrong place with mass effect, within the distribution of a pasted patch of real brain. Conversely, a stroke cavity, FLAIR edema and white matter hyperintensities are intensity deviations outside the healthy tissue distribution. Removing pathology randomization from the synthetic regime gives the control model of Table 1 when read through the robust tissue-class statistic; read through its deviation head it is near chance, as regime (c) is, since there is no deviation to learn.
Table S2. Data-regime ablation. Ceiling Dice on every case, mean of two seeds (500k steps, EMA weights, 4 mm median filter): our network and loss trained on (a) synthetic renderings with pathology randomization, the reported model, (b) the DISYRE analog, real healthy scans with pasted foreign patches, and (c) the FAE analog, real healthy scans alone, read through the tissue-class statistic or its own deviation head. Bold: column maximum.
| Regime | ATLAS | BraTS | BraTS FLAIR | BraTS | WMH FLAIR | WMH |
| (a) Synthetic, pathology randomization | 0.336 | 0.539 | 0.654 | 0.653 | 0.268 | 0.239 |
| (b) Real scans, pasted patches | 0.261 | 0.585 | 0.547 | 0.676 | 0.115 | 0.075 |
| (c) Real healthy scans, statistic readout | 0.165 | 0.246 | 0.631 | 0.544 | 0.288 | 0.118 |
| (c) Real healthy scans, head readout | 0.032 | 0.128 | 0.120 | 0.120 | 0.038 | 0.036 |
Table S2.
Robust tissue-class statistic
Robust tissue-class statistic. The intensity of voxel is rescaled to according to robust statistics within its anatomy-head class . The scaling requires median and median absolute deviation (MAD) over the class’s voxels in the acquisition. The robust tissue-class statistic requires no parameters and no gradients. For models trained without pathology randomization, the robust tissue-class statistic serves as a substitute deviation map. The anatomy head never predicts a lesion class (its target is the healthy map under the lesion), so a lesion voxel is scored against its host tissue.
Downstream heads
Controls. The control of each task is the control model without pathology randomization, a detector trained for the same 300k steps with pathology randomization switched off, fine-tuned for the same steps on healthy renderings only. The anatomy pool, steps, seed and renderings are fixed so the only difference is lack of pathology randomization in backbone training data. For registration the reported arm itself adds the pathological rendering to the fine-tune on half of the steps (Table S6). We claim robustness under pathology, rather than accuracy over the pretrained SynthMorph model or amp;rior.
Loss
Deviation loss. Over the batch holding the control and the pathological image of one draw (§ 2.7),
with . The first sum runs over the pathological rendering and the second over the control, where the deviation must vanish. The weights emphasise regions relevant in the detection task. The ratio balances the lesion against the far larger background, and , the 5-voxel dilation of outside it, is the border band. The cross-entropy is likewise weighted 5 inside the ROI (the healthy structures the lesion overlaps), the border-band and class weighting of [41, 75], fixed a priori.
Super-resolution
Scoring (Table 2). Each input is degraded along its slice axis: blurred with a Gaussian of standard deviation times the slice thickness, the slice-profile convention of Brain-ID, PEPSI and UNA generator code [55, 56, 57] (in training the same acquisition stage draws the degraded axis at random), sampled at three times the native slice spacing and restored to the native grid by tricubic interpolation, which is the network’s input and the interpolation baseline. Input and target share one normalization constant and the input’s window, the frame the network was trained in. The head predicts a residual and, next to it, a per-voxel amplitude in that scales it, trained jointly under the same loss; nothing but the input enters at inference. PSNR is read within the SynthStrip [40] brain mask and inside the dataset’s lesion mask. The acquisition the output is scored against is never read, by the network or by the gate.
Acquisitions
ATLAS-O. The 210 scans of the benchmark’s unseen-scanner subset, 1 mm, degraded to 3 mm; the lesion mask is the dataset’s and the modulation map is the detector on the degraded itself.
BraTS. The template-registered 1 mm volumes are resampled from acquisitions of unknown geometry, so automatic anisotropy detection on the acquired volumes (Appendix § 12.§ 2; 690 of 1,249 near-isotropic) splits the super-resolution result in two: on the acquired-isotropic volumes the gain over the tricubic input is +1.0 dB over the brain and −0.1 inside the tumor (unmodulated, one seed); on the 559 resampled volumes it is −1.3 and −5.7 dB, where the tricubic input already scores 37.6 dB inside the tumor because the “1 mm” target is itself an interpolation of a thick acquisition and every restored edge counts as error. The tables report the acquired-isotropic subset, 690 , 358 and 316 FLAIR volumes by the same two-component split; the exclusion is by acquisition, stated once. For the BraTS and columns of Table 2 the modulation map is the detector on the session’s co-registered FLAIR, degraded exactly like the input (3 mm along the same axis), where the tumor is faint on (Table 1); in Table S4 it is the detector on each contrast’s own degraded input.
| ATLAS-O (210) | BraTS (690) | BraTS FLAIR (316) | BraTS (358) | WMH FLAIR (60) | WMH (60) | WMH 3D (59) | ||
| Over tricubic | Brain | +0.96±0.03 | +0.76±0.23 | +1.01±0.10 | +0.96±0.06 | +0.68±0.02 | +0.55±0.03 | +0.89±0.00 |
| Lesion | +0.25±0.03 | −0.22±0.23 | +0.52±0.14 | +0.51±0.18 | +1.48±0.22 | −1.11±0.21 | −0.88±0.03 | |
| Over control | Brain | +0.32±0.18 | +0.44±0.53 | +0.34±0.28 | −0.31±0.08 | −0.12±0.06 | +0.19±0.08 | +0.22±0.01 |
| Lesion | +0.60±0.28 | +0.28±0.69 | +0.88±0.34 | +0.17±0.33 | −0.02±0.41 | +0.57±0.19 | +0.66±0.57 |
Table S4. Super-resolution per contrast, single-input modulation. Mean PSNR margin (dB) of detector-modulated PathoDraw over the tricubic input and over the control (no pathology, unmodulated), mean ± standard deviation over seeds; the modulation map is the detector on the degraded input itself, . BraTS on the acquired-isotropic volumes of each contrast; WMH FLAIR and 2D degraded 3 to 9 mm, WMH 3D 1 to 3 mm. Bold: significant gain, paired over cases.
WMH. Every site’s FLAIR is a 3 mm axial acquisition, so it is degraded threefold in voxels (3 to 9 mm) along its slice axis and scored against the acquisition on its own grid, never an interpolated target; the manual mask is drawn on that grid and the modulation map is the detector on the degraded FLAIR itself. The 2D of the same subjects shares that grid. The 1 mm 3D (59 subjects; one Utrecht subject at 0.89 mm in-plane excluded) is degraded 1 to 3 mm with the FLAIR mask rigidly resampled onto it (mri_coreg, nearest).
Automatic anisotropy detection
Template registration to 1 mm rotates the original slice direction off the grid axes and interpolates along all three, so axis-aligned tests such as per-axis gradient energy or spectra cannot detect a thick acquisition. The anisotropic point-spread function along the original slice direction, is unchanged however, which we measure rotation-invariantly. The high-pass residual of the volume (the image minus a Gaussian blur of 2 voxels) is taken inside the foreground eroded away from its edge, its gradient structure tensor is accumulated over that region, and the statistic is the eigenvalue ratio , which is 1 for an isotropic point-spread function and falls as the slices thicken. The residual is used rather than the image because anatomy has its own anisotropy, while noise does not, and interpolation correlates noise along the interpolated axis. The ratio is bimodal over the dataset (modes near 0.2 and 0.8, Figure S2). A two-component Gaussian mixture fit to all 1,249 values, with no reference to labels or scores, assigns a volume to the near-isotropic component where its posterior exceeds one half: 690 , 358 and 316 FLAIR volumes.

Figure S2. Acquired-isotropic versus resampled BraTS volumes (Appendix G.2). Left three columns: the volume in the three orthogonal planes; right three: its high-pass residual. Top: an acquired near-isotropic volume (ratio 0.81); bottom: a volume resampled from thick slices (0.15), smooth through-plane in the coronal and sagittal views.
Detector modulation
Detector modulation. The residual is the high-frequency content the healthy prior restores. Inside pathology, the true content differs from a healthy prior’s, and the residual is unreliable. The detector’s deviation captures voxels where the healthy prior does not apply, and scales the residual toward the interpolant there. At inference the output is with , the detector’s signed deviation at the voxel and , the value fixed a priori. The residual is scaled down linearly with the deviation: whole where the detector sees healthy tissue, halved at and nulled only at or more.
The deviation map is computed on the degraded input itself or a second degraded acquisition of the same session as listed above. The acquisition the output is scored against is never read. The modulation moves the brain by at most 0.07 dB and lifts the lesion from at or below the input to +0.25 (ATLAS-O) and +0.24 dB (BraTS ) in the seed mean, and paired over cases (Table 2). The control model without pathology randomization, under the same fine-tune, has no detector and is read without modulation.
Per contrast inference
Single input. Table S4 scores super-resolution on the contrasts of Table 1 under one rule: the modulation map is the detector’s deviation on the network’s own degraded input, nothing else of the session is read. Within the brain the modulated output is above the tricubic input on every contrast. Inside the lesion it is above the input wherever the lesion is visible on the input (ATLAS-O , BraTS FLAIR and , WMH FLAIR) and below it where the input hides the lesion (BraTS , the WMH ), where the detector cannot modulate what it cannot see; reading the session’s FLAIR, degraded like the input, instead lifts BraTS to +0.24 dB (Table 2). Inside the lesion the margin over the control is positive on every contrast except WMH FLAIR, on which both models see the lesion.
Deep generative brain prior (dmprior)
The healthy diffusion prior’s contrast-specific super-resolution, run on the same degraded pools and scored in the same frame, is below the tricubic input on all but one of the cases it has completed, by 3.7 dB within the brain and 5.5 inside the lesion on ATLAS-O (170 of 210 cases), 4.4 and 7.2 dB on BraTS (550 of 690), 2.4 and 7.1 dB on BraTS (all 358), 1.3 and 7.0 dB on WMH FLAIR (55 of 60; its forward-model fit fails on five Utrecht scans) and 2.7 and 4.0 dB on WMH 3D (50 of 59), means over cases as in Table 2.
Registration
The whole-pool comparison of Table 3 and Table S7 (cases in Figure 5) results from PathoDraw and the published pretrained SynthMorph model, an outside reference (Appendix § 13.§ 1). Additionally, the control backbone without pathology randomization is fine-tuned under SynthMorph’s loss in Table 3 and these are equal field size (Table S6), since a smaller field favors every lesion score.
Registration augmentation. Both draws of a pair, two different subjects, are augmented independently before rendering. A residual affine per draw with rotation , scaling and shear per axis, and a smooth random deformation are drawn on a coarse control grid. The registration models double the generator’s own setting, a 0.03–0.06 grid at up to 4 voxels standard deviation, to a 0.04–0.12 grid at up to 8 voxels, so that the two draws differ more in shape. The generator’s nonlinear field is a Gaussian displacement drawn on the coarse grid and linearly interpolated, a displacement rather than an integrated velocity, so it is not diffeomorphic by construction.
Folds. The network predicts a stationary velocity field integrated by scaling and squaring, so its field is diffeomorphic up to discretization. A fold is a negative smallest in-brain Jacobian determinant on a case, which the reported field has on 6.3% of cases, on fewer than 0.005% of the voxels of such a case (median smallest determinant −0.006). Evaluation pairs are affinely pre-aligned, and the fine-tune pre-aligns its training pairs with the same routine (300 iterations inside every step). The learned correction is a local field of 1.55 voxels mean displacement against 2.74 for the pretrained model at its default weight. The gradient penalty is small on a field of this size, and overlap is maximized with sharp local corrections at structure boundaries, which is where the first-percentile determinant falls (0.24 against 0.44). Pathology in training reduces the folds relative to the control (Table S6), and penalizes them directly. Table S5 sweeps our fine-tune (100k steps, one seed) over SynthMorph’s own weights translated into our loss, each column on the cases SynthMorph completed at that weight. SynthMorph weights the smoothness term by and the overlap term by , with as the default, and ours has overlap at weight 1 and smoothness at , so and the default is . Our field is fold-free from upward (and already at , a rung not shown), and SynthMorph’s at every weight. Lesion Dice rises as the field shrinks, while brain alignment falls to the affine level at the stiffest setting. In the cases where our field folds, the lesion-Dice margin over the control model trained without pathology randomization is +0.008 and +0.043 (seeds 1 and 2), and in the fold-free cases +0.003 and +0.055, with the Jacobian determinant inside the lesion at 1.02 in both groups.
Table S5. Regularization sweep. Our fine-tune (seed 1) and the published SynthMorph model across the smoothness weight (ours at ; default 0.5). Metrics as in Table S7; both scored on the cases SynthMorph completed (1,877 / 1,910 / 1,950 / 1,217 pairs). Bold: the better of the two at the same weight.
| 0.5 | 0.7 | 0.9 | 0.95 | ||
| PathoDraw | Lesion Dice | 0.827 | 0.869 | 0.926 | 0.885 |
| Brain Dice | 0.936 | 0.935 | 0.932 | 0.927 | |
| 1.55 | 1.19 | 0.80 | 0.97 | ||
| Folds (%) | 6.3 | 0.7 | 0.0 | 0.0 | |
| SynthMorph | Lesion Dice | 0.703 | 0.749 | 0.852 | 0.873 |
| Brain Dice | 0.949 | 0.946 | 0.939 | 0.934 | |
| 2.74 | 2.39 | 1.38 | 0.98 | ||
| Folds (%) | 0.0 | 0.0 | 0.0 | 0.0 |
Table S5.
Per-structure anatomy. Pathology randomization does not degrade anatomical accuracy. Carried through each field to MNI152 and scored as Dice of the anatomy head’s 56-class labels against the template’s, 23 large structures (brainstem, and per hemisphere cerebral white matter, cortex, lateral ventricle, cerebellar white matter and cortex, thalamus, caudate, putamen, pallidum, hippocampus, amygdala) score 0.743 under our field and 0.743 under the matched control’s over 1,964 cases (affine 0.671; paired difference −0.000, interval −0.001 to +0.000), within 0.002 on every dataset. On the 35-structure set that adds twelve small structures, the means are 0.688 against 0.691, within 0.003 on every dataset at a single seed.
Matched displacement. Lesion Dice through the warp rewards a smaller field, so a comparison at equal mean displacement can be met by stiffening the pretrained model. Its lesion Dice rises monotonically as its field shrinks (Table S5). Matching our 1.55 voxels falls between (1.74 voxels) and 0.9 (1.38 voxels), and at 0.95 the warp is sub-voxel. Such a match tests stiffness rather than registration, so we compare the pretrained model at its published default and across the sweep. The control, trained without pathology randomization, is matched to our displacement instead (Table S6). The stiffer weight is chosen after the fact against the lesion score on the evaluation pool, unavailable a priori. Our field arrives at its size from training alone, at the same published default. At every weight the two were run at, and at equal field size at , our field keeps more of the lesion without imposed stiffness (Table S5).
Field conditioning. At matched displacement (Table S6) the two fields keep the lesion and align the brain alike. Pathology randomization improves the conditioning of the field: the first-percentile determinant is 0.24 against 0.18 and the field folds on 6.3% of cases against 31.3%.
Pretrained SynthMorph
The published SynthMorph model is pretrained (Table S7) and inferred on the test datasets at its default regularization. Its deformable model is a hypernetwork whose regularization weight is a test-time parameter ( in its notation, the weight of the smoothness term against on the label overlap [38]; default 0.5). The HyperMorph extension [39] features amortized selection that could choose automatically, but only against a lesion criterion on annotated held-out scans, which a zero-shot setting lacks. The SynthMorph field at the default weight is larger than ours (2.74 against 1.55 voxels). At it halves its field (1.38 voxels) and reaches lesion Dice 0.85, above our 0.83 at the default weight but below our 0.93 at the same weight (Table S5), and its field is better conditioned at the default weight (Table S7) and fold-free at every weight (Table S5).
Training details
Detection. The reported model is the recipe of § 3.5 trained from random weights to 400k steps (two seeds), then continued for 100k steps with the cosine decay and the exponential moving average (EMA) to 500k. The component ablation of Table S3 branches at 300k from the seed-matched routine runs and trains each branch to 400k, so its rows are paired forks of one ancestor and are read from raw weights at 400k. The data-regime ablation of Table S2 is read at the 500k budget from the EMA weights, as Table 1. The control models differ from the reported model only by the absence of pathology randomization.
Fine-tuning. Super-resolution and registration fine-tune at constant learning rate . Half of the super-resolution draws are skull-stripped. For registration the two draws of a pair receive independent residual affines and the generator’s smooth deformation at twice its default amplitude and are affinely pre-aligned before the network, as at test time (Appendix § 13).
Neurodegeneration
ADNI3 [93] (331 scans per contrast, and FLAIR; 132 cognitively normal, 199 impaired) is a neurodegenerative cohort without focal lesions. Dice between the anatomy head and the control’s, mean over classes, median scan, mean over seeds, is 0.857 / 0.858 () and 0.826 / 0.829 (FLAIR) for cognitively normal / impaired subjects.
Benchmark context
The published FAE and DISYRE values were scored on the benchmark’s case lists, the BraTS 2020 training subjects and 54 of the 60 WMH subjects, and Table 1 scores PathoDraw on those lists through the organizers’ name mapping (329 of the 332 BraTS subjects are in the 2023 pool with a histogram; 54 of 54 WMH), threshold reselected on the subset: 0.537 [0.516, 0.558] on BraTS (every case of BraTS 2023: 0.539), 0.622 [0.600, 0.642] on BraTS FLAIR (0.654), 0.657 [0.635, 0.678] on BraTS (0.653), 0.259 [0.206, 0.314] on WMH FLAIR (all 60: 0.268) and 0.229 [0.178, 0.281] on WMH (0.239); ATLAS-O is the benchmark’s set already. The comparison is unchanged: above FAE on BraTS and above DISYRE on WMH outside the interval, and level with DISYRE on BraTS inside it. DISYRE [61, 62] has reported Dice scores up to 0.45 on ATLAS, 0.51 on BraTS and 0.73 on BraTS . Those compute thresholds by a different method than the independent benchmark. The published evaluation pools every voxel of the test set before searching the threshold, so its ceiling Dice is dominated by the largest lesions, computed on 2D slices over all 655 ATLAS scans with no scanner split and with that method’s own preprocessing. The benchmark [29] instead averages a per-individual Dice at one global threshold, which is what every row of Table 1 reports. For reference only, the same pooled ceiling computed from our per-case histograms (4 mm median filter) is 0.53 on all ATLAS cases (seeds 0.55 and 0.51), 0.59 on BraTS (0.59 and 0.59) and 0.72 on BraTS (0.73 and 0.71), against a per-individual 0.34, 0.54 and 0.65.
Confidence intervals
95 % bootstrapped confidence intervals of the PathoDraw column of Table 1 (20,000 resamples of the cases, seed-averaged per-case Dice at the global threshold) are, on the benchmark’s case lists, ATLAS-O 0.34–0.40, BraTS 0.52–0.56, BraTS FLAIR 0.60–0.64, BraTS 0.64–0.68, WMH FLAIR 0.21–0.31 and WMH 0.18–0.28. The published methods are single point estimates, so a value is bolded where the published value lies below our interval (BraTS against FAE, WMH against DISYRE) or inside it (ATLAS-O and BraTS against DISYRE). No published value lies above our intervals.
Healthy cohorts
The healthy cohort is HCP-Aging [14], 2,092 scans per contrast, none of them in training. Dice between the anatomy head and the control’s, mean over classes, median scan, mean over seeds, is 0.906 / 0.898 ( / ).
Matched specificity
Tab. reads PathoDraw at each published model’s healthy flag rate, with the threshold set on HCP-Aging and Dice on the benchmark’s case lists. This invokes a comparison with FAE and DISYRE at matched specificity.
False positives
False positives and negatives are described in Appendix § 19.
Failure analysis
False positives on every dataset are CSF extremes plus mildly deviant white matter, dominant on ATLAS. Dark CSF voxels are the majority of false-positive volume on WMH FLAIR and PRES scans. We expect this is because synthetic CSF is modeled as one mean per label while a real ventricle is darker at its center than at its walls. Misses are intermediate-intensity rim and edema that the anatomy head labels cortex, so the robust tissue-class statistic scores them as normal primarily on BraTS and FLAIR.
Manifold enclosure
Independent descriptors include 3D Zernike moments of each 26-connected component scaled into the unit ball, rotation invariants for (36 values), and volume reported separately. For a component with voxel positions , centered at its center of mass and scaled so that , the canonical 3D Zernike functions [63] are
with the order, the degree with even, the azimuthal index, the radius and the direction of a position, the spherical harmonics, the complex conjugate, the volume of voxel normalized to sum to one over , and the moments of one pair. The
Table S8. Intervals and -values behind the main tables. Against a published value (one number, other cases): 95% bootstrap interval of the seed-averaged per-case Dice (20,000 resamples) and the one-sided bootstrap that PathoDraw’s mean is at or below the published value. Against a model scored on the same cases: paired per-case difference, Wilcoxon signed-rank . PathoDraw on the benchmark’s case lists for the detection rows. †Super-resolution rows list ATLAS-O / BraTS / BraTS / WMH FLAIR; a negative margin is tested in its own direction. Median-scan match: the cutoff at which the median healthy scan, rather than the mean, flags the published rate.
| Table | Comparison | PathoDraw [95% CI] | Other | ||
| Tab. 1 | ATLAS-O vs DISYRE | 210 | 0.369 [0.337, 0.400] | 0.386 | 0.86 |
| BraTS vs FAE | 329 | 0.537 [0.516, 0.558] | 0.437 | ||
| BraTS vs DISYRE | 329 | 0.657 [0.635, 0.678] | 0.646 | 0.17 | |
| WMH vs DISYRE | 54 | 0.229 [0.178, 0.281] | 0.130 | ||
| Control (no pathology), paired, six rows | 54–329 | As in Table 1 | As in Table 1 | on five rows; WMH FLAIR 0.45 | |
| dmprior (no pathology), paired, six rows | 54–329 | As in Table 1 | As in Table 1 | on every row | |
| Tab. S9 | ATLAS-O at FAE’s specificity | 210 | 0.364 [0.332, 0.397] | 0.166 | |
| BraTS at FAE’s | 329 | 0.423 [0.402, 0.444] | 0.437 | 0.91 | |
| WMH at FAE’s | 54 | 0.229 [0.178, 0.281] | 0.016 | ||
| ATLAS-O at DISYRE’s | 210 | 0.269 [0.244, 0.296] | 0.384 | 1.00 | |
| BraTS at DISYRE’s | 329 | 0.160 [0.145, 0.175] | 0.264 | 1.00 | |
| WMH at DISYRE’s | 54 | 0.102 [0.073, 0.134] | 0.109 | 0.67 | |
| Median-scan match: ATLAS-O at FAE’s | 210 | 0.363 [0.330, 0.396] | 0.166 | ||
| BraTS at FAE’s | 329 | 0.427 [0.407, 0.448] | 0.437 | 0.82 | |
| WMH at FAE’s | 54 | 0.229 [0.178, 0.281] | 0.016 | ||
| ATLAS-O at DISYRE’s | 210 | 0.316 [0.288, 0.344] | 0.384 | 1.00 | |
| BraTS at DISYRE’s | 329 | 0.214 [0.198, 0.231] | 0.264 | 1.00 | |
| WMH at DISYRE’s | 54 | 0.140 [0.103, 0.179] | 0.109 | 0.053 | |
| Tab. 2 | Against tricubic, brain† | 210/690/358/60 | +0.96/+0.79/+1.09/+0.68 | Paired | |
| Against tricubic, lesion | 210/690/358/60 | +0.25/+0.24/+0.44/+1.48 | Paired | ||
| Against control, brain | 210/690/358/60 | +0.32/+0.47/−0.18/−0.12 | Paired | ||
| Against control, lesion | 210/690/358/60 | +0.60/+0.74/+0.10/−0.02 | Paired | ||
| Against dmprior (no pathology), brain | 170/550/358/55 | +4.7/+5.2/+3.5/+2.0 | Paired | ||
| Against dmprior (no pathology), lesion | 170/550/358/55 | +5.8/+7.5/+7.6/+8.6 | Paired | ||
| Tab. 3 | Lesion Dice vs control | 210 / 1,249 / 60 | +0.046 / +0.014 / +0.063 | Paired | |
| Lesion Dice vs SynthMorph | 209 / 1,179 / 59 | +0.112 / +0.140 / +0.180 | Paired | ||
| Tab. S6 | Lesion Dice vs matched control | 655 / 1,249 / 60 | +0.005 / −0.005 / +0.012 | Paired | 0.42 / / 0.038 |
Table S8.
Table S9. Dice at matched healthy specificity. PathoDraw read at each published model’s healthy false-positive rate (cutoff set on HCP-Aging), on the benchmark’s case lists (ATLAS-O 210, BraTS 329, WMH 54), with 95% bootstrap intervals. Published: FAE and DISYRE Dice at their own healthy thresholds (median-filtered), not the ceiling Dice of Tab. 1; rows. Level from: the lowest healthy flag rate at which the published value lies inside or below PathoDraw’s interval. Bold: inside or above the published interval; -values in Table S8.
36 pairs with give the 36 invariants. The descriptors were -scored on the synthetic set. Components of at least 0.03 mL (30 voxels at 1 mm) enter the analysis, and any component larger than 6,000 voxels (6 mL at 1 mm) is subsampled to 6,000 voxels. To estimate manifold enclosure, -NN precision and recall [45] with were used. Synthetic masks were block-resampled to the real grid before description. The control, fixed a priori, consisted of 2,500 random blobs (unions of 1–4 random ellipsoids (axis ratios , random rotation, random offsets, largest connected component kept), volumes log-uniform in [0.03, 200] mL) described on each dataset’s grid like the draws (Figure S3 and Figure S4). Single ellipsoids were also tested and enclose nothing (recall ), as their descriptors lie on a two-parameter curve whose nearest-neighbor radii are too small to contain any real shape, so they are omitted.

Figure S3. Manifold enclosure in the Zernike shape space: the generator’s -NN manifold ( [45], translucent) and random blobs (second row), projected on the first three principal components of the 1 mm draws, against each dataset’s real lesion components on its matched voxel grid (filled when the full 36-D test encloses them, hollow when not; titles give the recall), ATLAS and BraTS sharing the 1 mm generator set by construction.

Figure S4. Descriptor space (first two principal components of the synthetic set) with the real lesion components of each dataset, on matched voxel grids; titles give the recall.
Coverage measurement
Every lesion component of the evaluation masks and of 1,000 generator draws is described by its 3D Zernike invariants [63], and a real component counts as enclosed when it lies within the -nearest-neighbor ball of a generator component [45] (Table S10): recall is 0.86–0.93 on all four datasets with precision 0.87–0.95.
| Dataset | Generator Recall | Generator Precision | Random blobs Recall | Random blobs Precision | |
| ATLAS | 448 | 0.93 | 0.88 | 0.49 | 0.57 |
| BraTS | 1,833 | 0.89 | 0.91 | 0.55 | 0.65 |
| WMH | 475 | 0.86 | 0.95 | 0.62 | 0.70 |
| PRES | 387 | 0.93 | 0.87 | 0.74 | 0.73 |
Table S10. Manifold enclosure of real lesion components in the 3D Zernike descriptor space (), one generator reference set per voxel grid (ATLAS and BraTS are both 1 mm and therefore share it): recall = real components inside the generator’s -NN manifold, precision = the converse; synthetic masks resampled to each dataset’s voxel grid (ATLAS, BraTS 1 mm; WMH mm; PRES mm), components mL. Control: 2,500 random blobs (unions of 1–4 random ellipsoids, volumes log-uniform in 0.03–200 mL; no parameter taken from the generator or the data).
Enclosure interpretation
A control shape distribution, specified without reference to the generator or to any dataset (volumes log-uniform in 0.03–200 mL; construction in Table S10), bounds the reading: random blobs, unions of one to four random ellipsoids (axis ratios in [1, 3]), enclose 0.49–0.74 of the real components with precision 0.57–0.73, against 0.86–0.93 and 0.87–0.95 for the generator. Shape enclosure is therefore necessary but not sufficient. A shape-only null covers half of the real shapes, the anatomy-conditioned modes account for the rest, and enclosure does not order the Dice results (ATLAS has the highest enclosure and a mid-range Dice). The residual loss is in appearance, which Appendix § 19 locates in the healthy rendering.
Effect of lesion scale
The compared published methods trained only on healthy priors do not corrupt pathology uniformly. We speculate that a prior fit to healthy anatomy represents any structure on the scale of that anatomy, a stroke cavity like a ventricle, whereas a punctate lesion has no healthy counterpart at its scale and is regressed to the surrounding tissue.
References
- [1]Robert J. Adler and Jonathan E. Taylor. Random Fields and Geometry. Springer, New York, 2007. 16DOI
- [2]Laura M. Allen, Anton N. Hasso, Jason Handwerker, and Hamed Farid. Sequence-specific MR imaging findings that are useful in dating ischemic stroke. RadioGraphics, 32(5): 1285–1297, 2012. 15, 16DOI
- [3]Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The RSNA-ASNR-MICCAI BraTS 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021. 6arxiv.org/abs/2107.02314
- [4]Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, et al. RSNA-ASNR-MICCAI-BraTS 2021. The Cancer Imaging Archive, https://doi.org/10.7937/tcia.8c3x-9874, 2023. Dataset, version 1. 6
- [5]Hassan Baker and Austin J. Brockmeier. Patch2Loc: Learning to localize patches for unsupervised brain lesion detection. In Proceedings of the 29th International Conference on Artificial Intelligence and Statistics (AISTATS), 2026. arXiv:2506.22504. 2arxiv.org/abs/2506.22504
- [6]J. Bamford, P. Sandercock, M. Dennis, J. Burn, and C. Warlow. Classification and natural history of clinically identifiable subtypes of cerebral infarction. The Lancet, 337(8756): 1521–1526, 1991. 16pubmed.ncbi.nlm.nih.gov/1675378
- [7]Thomas W. K. Batty, Mahima Karki, Aneesh B. Singhal, Ona Wu, Saloomеh Sadaghiani, Bruce C. V. Campbell, Stephen M. Davis, Geoffrey A. Donnan, Kevin N. Sheth, and W. Taylor Kimberly. Brain edema predicts outcome after nonlacunar ischemic stroke. Stroke, 45(12):3643–3648, 2014. 16DOI
- [8]Matthew Baugh, Jeremy Tan, Johanna P. Müller, Mischa Dombrowski, James Batten, and Bernhard Kainz. Many tasks make light work: Learning to localise medical anomalies from multiple synthetic tasks. In MICCAI, 2023. 2arxiv.org/abs/2307.00899
- [9]Finn Behrendt, Debayan Bhattacharya, Lennart Maack, Julia Krüger, Roland Opfer, and Alexander Schlaefer. A review of deep learning-based unsupervised anomaly detection in brain MRI. Medical Image Analysis, 112:104076, 2026. 2DOI
- [10]Benjamin Billot. SynthSeg. https://github.com/Bjillot/SynthSeg, 2024. GitHub repository. 3DOI
- [11]Benjamin Billot, Eleanor Robinson, Adrian V. Dalca, and Juan Eugenio Iglesias. Partial volume segmentation of brain MRI scans of any resolution and contrast. In Medical Image Computing and Computer Assisted Intervention (MICCAI), pages 177–187, 2020. 16DOI
- [12]Benjamin Billot, Stefano Cerri, Koen Van Leemput, Adrian V. Dalca, and Juan Eugenio Iglesias. Joint segmentation of multiple sclerosis lesions and brain anatomy in MRI scans of any contrast and resolution with CNNs. In IEEE International Symposium on Biomedical Imaging (ISBI), pages 1971–1974, 2021. 2DOI
- [13]Benjamin Billot, Douglas N. Greve, Oula Puonti, Axel Thielscher, Koen Van Leemput, Bruce Fischl, Adrian V. Dalca, and Juan Eugenio Iglesias. SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining. Medical Image Analysis, 86:102789, 2023. 2, 3, 4, 5, 16DOI
- [14]Susan Y. Bookheimer, David H. Salat, Melissa Terpstra, Beau M. Ances, Deanna M. Barch, Randy L. Buckner, Gregory C. Burgess, Sandra W. Curtiss, Mirella Diaz-Santos, Jennifer Stiegler, Bruce Fischl, Douglas N. Greve, Hannah A. Hagy, Michael P. Harms, Olivia M. Hatch, Trey Hedden, Cynthia Hodge, Kevin C. Japardi, Taylor P. Kuhn, Timothy K. Ly, Stephen M. Smith, Leah H. Somerville, Kâmil Uğurbil, Andre van der Kouwe, David Van Essen, Roger P. Woods, and Essa Yacoub. The Lifespan Human Connectome Project in Aging: An overview. NeuroImage, 185:335–348, 2019. doi:10.1016/j.neuroimage.2018.10.009. 7, 23DOI
- [15]P. A. Bottomley, C. J. Hardy, R. E. Arnsperger, and G. Allen-Moore. A review of 1H nuclear magnetic resonance relaxation in pathology: are T1 and T2 diagnostic? Medical Physics, 14(1):1–37, 1987. 16pubmed.ncbi.nlm.nih.gov/3031439
- [16]Yuan Cao, Sumeet Dash, Antonia Zachariadis, Stefanie Schreiber, Katja Neumann, and Jose Bernal. Synthetic training for long-tail haemorrhagic lesion segmentation in data-scarce settings. In Simulation and Synthesis in Medical Imaging (SASHIMI), MICCAI 2026 Workshops, 2026. arXiv:2610.01542. 2arxiv.org/abs/2610.01542
- [17]Liam Chalfort, Ioannis Pappas, Cathy J. Price, and John Ashburner. Synthetic data for robust stroke segmentation. Machine Learning for Biomedical Imaging (MELBA), 3, 2025. arXiv:2404.01946. 2
- [18]Dami Choi, Alexandre Passos, Christopher J. Shallue, and George E. Dahl. Faster neural network training with data echoing. arXiv:1907.05550, 2019. 16arxiv.org/abs/1907.05550
- [19]O. Clatz, M. Sermesant, P.-Y. Bondiau, H. Delingette, S. K. Warfield, G. Malandain, and N. Ayache. Realistic simulation of the 3-D growth of brain tumors in MR images coupling diffusion with biomechanical deformation. IEEE Transactions on Medical Imaging, 24(10):1334–1346, 2005. 16pubmed.ncbi.nlm.nih.gov/16229419
- [20]Kyle Cranmer, Johann Brehmer, and Gilles Louppe. The frontier of simulation-based inference. Proceedings of the National Academy of Sciences, 117(48):30055–30062, 2020. 5DOI
- [21]F.-E. de Leeuw, J. C. de Groot, E. Achten, M. Oudkerk, L. M. Ramos, R. Heijboer, A. Hofman, J. Jolles, J. van Gijn, and M. M. B. Breteler. Prevalence of cerebral white matter lesions in elderly people: a population based magnetic resonance imaging study. The Rotterdam Scan Study. Journal of Neurology, Neurosurgery & Psychiatry, 70(1):9–14, 2001. 16pubmed.ncbi.nlm.nih.gov/11118240
- [22]Zhiwei Deng and Yonggang Shi. FluxCut: Shifting from Global Reconstruction to Local Geometry for Unsupervised Small Lesion Segmentation. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2026. Springer Nature Switzerland, 2026. 2DOI
- [23]Neel Dey, Benjamin Billot, Hallee E. Wong, Clinton J. Wang, Mengwei Ren, P. Ellen Grant, Adrian V. Dalca, and Polina Golland. Learning general-purpose biomedical volume representations using randomized synthesis. In ICLR, 2025. arXiv:2411.02372. 2arxiv.org/abs/2411.02372
- [24]Sebo Diaz, Polina Golland, Elfar Adalsteinsson, and Neel Dey. Why invariance is not enough for biomedical domain generalization and how to fix it. arXiv:2604.02564, 2026. 2
- [25]Vuk Dujlejić, Slobodan Marinković, Biljana Georgievski, Lazar Stijak, Milan Aksić, Laslo Puškaš, and Ivan Milić. Clinical significance of blood supply to the internal capsule and basal ganglia. Journal of Clinical Neuroscience, 25:19–26, 2016. 16DOI
- [26]F. Fazekas, J. B. Chawluk, A. Alavi, H. I. Hurtig, and R. A. Zimmerman. MR signal abnormalities at 1.5 T in Alzheimer’s dementia and normal aging. American Journal of Roentgenology, 149(2):351–356, 1987. 16pubmed.ncbi.nlm.nih.gov/3496763
- [27]Bruce Fischl. FreeSurfer. NeuroImage, 62(2):774–781, 2012. 6DOI
- [28]Vladimir Fonov, Alan C. Evans, Kelly Botteron, C. Robert Almli, Robert C. McKinstry, and D. Louis Collins. Unbiased average age-appropriate atlases for pediatric studies. NeuroImage, 54(1):313–327, 2011. 6DOI
- [29]Alexander Frotscher, Christian F. Baumgartner, and Thomas Wolfers. Deep unsupervised anomaly detection in brain imaging: large-scale benchmarking and bias analysis. arXiv:2512.01534, 2025. 2, 6, 7, 18, 23arxiv.org/abs/2512.01534
- [30]Alexander Frotscher, Javadihan Kapoor, Thomas Wolfers, and Christian F. Baumgartner. Unsupervised anomaly detection in medical imaging using aggregated normative diffusion. Medical Image Analysis, 109:103895, 2026. arXiv:2312.01904. 2arxiv.org/abs/2312.01904
- [31]Jianqiao Ge, Guoyuan Yang, Meizhen Han, et al. Increasing diversity in connectomics with the Chinese Human Connectome Project. Nature Neuroscience, 26(1):163–172, 2023. 6, 16DOI
- [32]Daniel T. Ginat and Steven P. Meyers. Intracranial lesions with high signal intensity on T1-weighted MR images: differential diagnosis. RadioGraphics, 32(2):499–516, 2012. 16DOI
- [33]Paul Glasserman and David D. Yao. Some guidelines and guarantees for common random numbers. Management Science, 38(6):884–908, 1992. 16DOI
- [34]Marina Grifell i Plana, Vladyslav Zalevskyy, Léa Schmidt, Yvan Gomez, Thomas Sanchez, Vincent Dunet, Mériam Koob, Vanessa Siffredi, and Meritxell Bach Cuadra. Enhancing corpus callosum segmentation in fetal MRI via pathology-informed domain randomization. In PIPPI Workshop, MICCAI (LNCS 16118), 2025. arXiv:2508.20475. 2DOI
- [35]H. Gudbjartsson and S. Patz. The Rician distribution of noisy MRI data. Magnetic Resonance in Medicine, 34(6):910–914, 1995. 16pubmed.ncbi.nlm.nih.gov/8598820
- [36]J. V. Hajnal, D. J. Bryant, L. Kasuboski, P. M. Pattany, B. De Coene, P. D. Lewis, J. M. Pennock, A. Oatridge, I. R. Young, and G. M. Bydder. Use of fluid attenuated inversion recovery (FLAIR) pulse sequences in MRI of the brain. Journal of Computer Assisted Tomography, 16(6):841–844, 1992. 4, 16, 17pubmed.ncbi.nlm.nih.gov/1430427
- [37]Malte Hoffmann, Benjamin Billot, Douglas N. Greve, Juan Eugenio Iglesias, Bruce Fischl, and Adrian V. Dalca. SynthMorph: Learning contrast-invariant registration without acquired images. IEEE Transactions on Medical Imaging, 41(3):543–558, 2022. 2, 4, 5, 6, 16, 22DOI
- [38]Malte Hoffmann, Andrew Hoopes, Douglas N. Greve, Bruce Fischl, and Adrian V. Dalca. Anatomy-aware and acquisition-agnostic joint registration with SynthMorph. Imaging Neuroscience, 2024. doi:10.1162/imag_a_00197. 5, 6, 21, 22DOI
- [39]Andrew Hoopes, Malte Hoffmann, Bruce Fischl, John Guttag, and Adrian V. Dalca. HyperMorph: Amortized hyperparameter learning for image registration. In Information Processing in Medical Imaging (IPMI), pages 3–17, 2021. 21arxiv.org/abs/2101.01035
- [40]Andrew Hoopes, Jocelyn S. Mora, Adrian V. Dalca, Bruce Fischl, and Malte Hoffmann. SynthStrip: skull-stripping for any brain image. NeuroImage, 260:119474, 2022. 6, 18DOI
- [41]Seyed Mohsen Hosseini and Mahdieh Soleymani Baghshah. Dilated balanced cross entropy loss for medical image segmentation. arXiv:2412.06045, 2024. 5, 18arxiv.org/abs/2412.06045
- [42]Juan Eugenio Iglesias, Benjamin Billot, Yaël Balbastre, Colin Magdamo, Steven E. Arnold, Sudehna Das, Brian L. Edlow, Daniel C. Alexander, Polina Golland, and Bruce Fischl. SynthSR: A public AI tool to turn heterogeneous clinical brain scans into high-resolution T1-weighted images for 3D morphometry. Science Advances, 9(5):eadc3607, 2023. 2, 4, 16DOI
- [43]Fabian Isensee, Paul F. Jaeger, Simon A. A. Kohl, Jens Petersen, and Klaus H. Maier-Hein. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 18(2):203–211, 2021. 5DOI
- [44]Hugo J. Kuijf, J. Matthijs Biesbroek, Jeroen de Bresser, Rutger Heinen, Simon Andermatt, et al. Standardized assessment of automatic segmentation of white matter hyperintensities and results of the WMH segmentation challenge. IEEE Transactions on Medical Imaging, 38(11):2556–2568, 2019. 6DOI
- [45]Tuomas Kynkäänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In NeurIPS, pages 3927–3936, 2019. arXiv:1904.06991. 24, 26arxiv.org/abs/1904.06991
- [46]Ioannis Lagogiannis, Felix Meissen, Georgios Kaissis, and Daniel Rueckert. Unsupervised pathology detection: A deep dive into the state of the art. IEEE Transactions on Medical Imaging, 43(1):241–252, 2024. arXiv:2303.00609. 2DOI
- [47]Suvi Lavaara, Riitta Mäntylä, Tiina Salminen, Hannu Haapasalo, Jani Raitanen, Juha Jääskeläinen, and Anssi Auvinen. Incidence of gliomas by anatomic location. Neuro-Oncology, 9(3):319–325, 2007. 16
- [48]Pablo Laso, Stefano Cerri, Annabel Sorby-Adams, Jennifer Guo, Farrah Mateen, Philipp Goebl, Jiaming Wu, Peirong Liu, Hongwei Li, Sean I. Young, Benjamin Billot, Oula Puonti, Gordon Sze, Sam Payabvash, Adam DeHavenon, Kevin N. Sheth, Matthew S. Rosen, John Kirsch, Nicola Strisciuglio, Jelmer M. Wolterink, Arman Eshaghi, Frederik Barkhof, W. Taylor Kimberly, and Juan Eugenio Iglesias. Quantifying white matter hyperintensity and brain volumes in heterogeneous clinical and low-field portable MRI. In IEEE International Symposium on Biomedical Imaging (ISBI), 2024. arXiv:2312.05119. 2, 5arxiv.org/abs/2312.05119
- [49]Ana Lawry Aguila, Dina Zemlyanker, You Cheng, Sudeshna Das, Daniel C. Alexander, Oula Puonti, Annabel Sorby- Adams, W. Taylor Kimberly, and Juan Eugenio Iglesias. Deep generative priors for 3D brain analysis. arXiv:2510.15119, 2025. 2, 6, 7
- [50]Tai Le-Gia and Jaehyun Ahn. Training-free zero-shot anomaly detection in 3D brain MRI with 2D foundation models. In Proceedings of the 9th International Conference on Medical Imaging with Deep Learning (MIDL), pages 3069–3088, 2026. arXiv:2602.15315. 2arxiv.org/abs/2602.15315
- [51]Chun-Liang Li et al. CutPaste: Self-supervised learning for anomaly detection and localization. In CVPR, 2021. 2arxiv.org/abs/2104.04015
- [52]Shangkun Li, Jie Xu, Yi Guo, Zeju Li, and Yuanyuan Wang. Enhancing Brain MRI Anomaly Detection and Reasoning with ROI Rethink and Synthetic Data. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2026. Springer Nature Switzerland, 2026. 2arxiv.org/abs/2606.25894
- [53]Sook-Lei Liew, Bethany P Lo, Miranda R Donnelly, et al. A large, curated, open-source stroke neuroimaging dataset to improve lesion segmentation algorithms. Scientific Data, 9(1):320, 2022. 6DOI
- [54]Liyang Liu, Yi Li, Zhanghui Kuang, Jing-Hao Xue, Yimin Chen, Wenming Yang, Qingming Liao, and Wayne Zhang. Towards impartial multi-task learning. In International Conference on Learning Representations (ICLR), 2021. 5
- [55]Peirong Liu, Oula Puonti, Xiaoling Hu, Daniel C. Alexander, and Juan Eugenio Iglesias. Brain-ID: Learning contrast-agnostic anatomical representations for brain imaging. In ECCV, pages 322–340, 2024. 18arxiv.org/abs/2311.16914
- [56]Peirong Liu, Oula Puonti, Annabel Sorby-Adams, William T. Kimberly, and Juan Eugenio Iglesias. PEPSI: Pathology-enhanced pulse-sequence-invariant representations for brain MRI. In Medical Image Computing and Computer Assisted Intervention (MICCAI), 2024. arXiv:2403.06227. 2, 6arxiv.org/abs/2403.06227
- [57]Peirong Liu, Ana Lawry Aguila, and Juan Eugenio Iglesias. Unraveling normal anatomy via fluid-driven anomaly randomization. In CVPR, pages 10455–10465, 2025. arXiv:2501.13370. 2, 6, 18arxiv.org/abs/2501.13370
- [58]Xinyu Ma, Jinhui Ma, Shiqi He, Jian Pei, and Lingyang Chu. InvDetect: Unsupervised Medical Anomaly Detection in the Noise Latent Space of DDIM. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2026. Springer Nature Switzerland, 2026. 2DOI
- [59]Felix Meissen, Johannes C. Paetzold, Georgios Kaissis, and Daniel Rueckert. Unsupervised anomaly localization with structural feature-autoencoders. In MICCAI BrainLes Workshop 2022 (LNCS), pages 14–24, 2023. arXiv:2208.10992; doi:10.1007/978-3-031-33842-7_2. 2DOI
- [60]J. P. Mugler, III and J. R. Brookeman. Three-dimensional magnetization-prepared rapid gradient-echo imaging (3D MP RAGE). Magnetic Resonance in Medicine, 15(1):152–157, 1990. 4, 16, 17pubmed.ncbi.nlm.nih.gov/2374495
- [61]Sergio Naval Marimont, Matthew Baugh, Vasilis Siomos, Christos Tzelepis, Bernhard Kainz, and Giacomo Tarroni. DISYRE: diffusion-inspired synthetic restoration for unsupervised anomaly detection. In IEEE International Symposium on Biomedical Imaging (ISBI), 2024. arXiv:2311.15453. 2, 23arxiv.org/abs/2311.15453
- [62]Sergio Naval Marimont et al. Ensembled cold-diffusion restorations for unsupervised anomaly detection. In MICCAI, 2024. arXiv:2407.06635. 2, 23arxiv.org/abs/2407.06635
- [63]Marcin Novotni and Reinhard Klein. Shape retrieval using 3D Zernike descriptors. Computer-Aided Design, 36(11):1047–1062, 2004. 23, 24DOI
- [64]Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, Ben Glocker, and Daniel Rueckert. Attention U-Net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 5, 17
- [65]Mark D. Olshanyi, David R. Schreier, Jian Li, Chiara Maffei, Annabel Sorby-Adams, Hannah C. Kinney, Brian C. Healy, Holly J. Freeman, Jared Shless, Christophe Destrieux, Henry Tredgido, Juan Eugenio Iglesias, Emery N. Brown, and Brian L. Edlow. Probabilistic mapping and automated segmentation of human brainstem with matter bundles. Proceedings of the National Academy of Sciences, 123(6):e2509321123, 2026. 5, 17
- [66]Chirag G. Patil, Anthony Yi, Adam Erlamsiry, Jethro Hu, Debraj Mukherjee, Dwain K. Irvin, John S. Yu, Serguei I. Bannykh, Keith L. Black, and Miriam Nuño. Prognosis of patients with multifocal glioblastoma: a case-control study. Journal of Neurosurgery, 117(4):705–711, 2012. 16DOI
- [67]Matteo Pentussaglia, Xiaoming Zhang, Benjamin Billot, and Maria A. Zuluaga. Anatomy-Conditioned Domain Randomization for Zero-Shot 3D Cerebrovascular Segmentation. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2026. Springer Nature Switzerland, 2026. 2DOI
- [68]Jesse Phitidis, Antanas Kascenas, Maria Valdés Hernández, William N. Whiteley, Joanna M. Wardlaw, and Alison Q. O’Neil. Confounder-aware image synthesis for pathology segmentation in new magnetic resonance imaging sequences. In Medical Image Understanding and Analysis (MIUA), pages 227–241, 2024. doi:10.1007/978-3-031-66958-3_17. 2DOI
- [69]Marcel Prastawa, Elizabeth Bullitt, and Guido Gerig. Simulation of brain tumors in MR images for evaluation of segmentation efficacy. Medical Image Analysis, 13(2):297–311, 2009. 3, 16DOI
- [70]Oula Puonti, Juan Eugenio Iglesias, and Koen Van Leemput. Fast and sequence-adaptive whole-brain segmentation using parametric Bayesian modeling. NeuroImage, 143:235–249, 2016. 6DOI
- [71]Oula Puonti, Koen Van Leemput, Guilherme B. Saturnino, Hartwig R. Siebner, Kristoffer H. Madsen, and Axel Thielscher. Accurate and robust whole-head segmentation from magnetic resonance images for individualized head modeling. NeuroImage, 219:117044, 2020. 6DOI
- [72]Zehua Ren, Jingyan Zhai, Haifeng Wang, Yixuan Bo, Fan Wang, Cailei Zhao, and Chunfeng Lian. SynHydro: Biomechanics-Driven Domain Randomization for Hydrocephalus-Agnostic, Generalizable Tissue-Ventricle Segmentation Across Age, Modality, and Resolution. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2026. Springer Nature Switzerland, 2026. 2DOI
- [73]Afra Ritzl, Stephanie Meisel, Hans-Jörg Wittsack, Gereon R. Fink, Mario Siebler, Ulrich Mödder, and Rüdiger J. Seitz. Development of brain infarct volume as assessed by magnetic resonance imaging (MRI): follow-up of diffusion-weighted MRI lesions. Journal of Magnetic Resonance Imaging, 20(2):201–207, 2004.pubmed.ncbi.nlm.nih.gov/15269944
- [74]Alexandra G. Roberts, Ha Ma. Luu, Mert Şişman, Alexey V. Dimov, Ceren Tozlu, İlhami Kovanlıkaya, Susan A. Gauthier, Thanh D. Nguyen, and Yi Wang. Synthetic generation and latent projection denoising of rim lesions in multiple sclerosis. In CVPR Workshop on Synthetic Data for Computer Vision, 2025. arXiv:2505.23353.arxiv.org/abs/2505.23353
- [75]Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234–241, 2015.arxiv.org/abs/1505.04597
- [76]W. D. Rooney, G. Johnson, X. Li, E. R. Cohen, S.-G. Kim, K. Ugurbil, and C. S. Springer, Jr. Magnetic field and tissue dependencies of human brain longitudinal 1H2O relaxation in vivo. Magnetic Resonance in Medicine, 57(2):308–318, 2007.pubmed.ncbi.nlm.nih.gov/17260370
- [77]Albert Rose. The sensitivity performance of the human eye on an absolute scale. Journal of the Optical Society of America, 38(2):196–208, 1948.
- [78]Hugues Roy, Reuben Dorent, and Ninon Burgos. Unsupervised Brain Anomaly Detection as a Bayesian Inverse Problem with Diffusion Prior. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2026. Springer Nature Switzerland, 2026.arxiv.org/abs/2609.24265
- [79]Hannah M. Schlüter et al. Natural synthetic anomalies for self-supervised anomaly detection and localization. In ECCV, 2022.arxiv.org/abs/2109.15222
- [80]J. G. Sled, A. P. Zijdenbos, and A. C. Evans. A nonparametric method for automatic correction of intensity nonuniformity in MRI data. IEEE Transactions on Medical Imaging, 17(1):87–97, 1998.pubmed.ncbi.nlm.nih.gov/9617910
- [81]Jeremy Tan, Benjamin Hou, Thomas Day, John Simpson, Daniel Rueckert, and Bernhard Kainz. Detecting outliers with Poisson image interpolation. In MICCAI, 2021. arXiv:2107.02622.arxiv.org/abs/2107.02622
- [82]Jeremy Tan, Benjamin Hou, James Batten, Huaqi Qiu, and Bernhard Kainz. Detecting outliers with foreign patch interpolation. Machine Learning for Biomedical Imaging, 1, 2022.DOI
- [83]Lin Tian, Jonathan Williams-Ramirez, Dina Zemlyanker, Lucas J. Deden-Binder, Rogeny Herisse, Theresa R. Connors, Mark Montine, Istvan N. Huszar, Lilla Zöllei, Sean I. Young, Christine Mac Donald, C. Dirk Keene, Derek H. Oakley, Bradley T. Hyman, Oula Pounti, Matthew S. Rosen, and Juan Eugenio Iglesias. Reference-free 3D reconstruction of brain dissection slabs via learned atlas coordinates. arXiv:2503.09963, 2025.arxiv.org/abs/2503.09963
- [84]Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 23–30, 2017.arxiv.org/abs/1703.06907
- [85]Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In CVPR Workshops, pages 969–977, 2018.arxiv.org/abs/1804.06516
- [86]Neil Upadhyay and Adam D. Waldman. Conventional MRI evaluation of gliomas. The British Journal of Radiology, 84 (special issue 2):S107–S111, 2011.DOI
- [87]David C Van Essen, Stephen M Smith, Deanna M Barch, Timothy EJ Behrens, Essa Yacoub, and Kamil Ugurbil. The WU-Minn Human Connectome Project: An overview. NeuroImage, 80:62–79, 2013.DOI
- [88]Koen Van Leemput, Frederik Maes, Dirk Vandermeulen, Alan Colchester, and Paul Suetens. Automated segmentation of multiple sclerosis lesions by model outlier detection. IEEE Transactions on Medical Imaging, 20(8):677–688, 2001.pubmed.ncbi.nlm.nih.gov/11513020
- [89]Koen Van Leemput, Frederik Maes, Dirk Vandermeulen, and Paul Suetens. A unifying framework for partial volume segmentation of brain MR images. IEEE Transactions on Medical Imaging, 22(1):105–119, 2003.pubmed.ncbi.nlm.nih.gov/12703764
- [90]Sarah E. Vermeer, Peter J. Koudstaal, Matthijs Oudkerk, Albert Hofman, and Monique M. B. Breteler. Prevalence and risk factors of silent brain infarcts in the population-based Rotterdam Scan Study. Stroke, 33(1):21–25, 2002.pubmed.ncbi.nlm.nih.gov/11779883
- [91]J. P. Wansapura, S. K. Holland, R. S. Dunn, and W. S. Ball, Jr. NMR relaxation times in the human brain at 3.0 tesla. Journal of Magnetic Resonance Imaging, 9(4):531–538, 1999.pubmed.ncbi.nlm.nih.gov/10232510
- [92]Joanna M Wardlaw, Eric E Smith, Geert J Biessels, Charlotte Cordonnier, Franz Fazekas, Richard Frayne, Richard I Lindley, John T O’Brien, Frederik Barkhof, Oscar R Benavente, et al. Neuroimaging standards for research into small vessel disease and its contribution to ageing and neurodegeneration. The Lancet Neurology, 12(8):822–838, 2013.DOI
- [93]Michael W. Weiner, Dallas P. Veitch, Paul S. Aisen, Laurel A. Beckett, Nigel J. Cairns, Robert C. Green, Danielle Harvey, Clifford R. Jack, William Jagust, John C. Morris, Ronald C. Petersen, Jennifer Salazar, Andrew J. Saykin, Leslie M. Shaw, Arthur W. Toga, and John Q. Trojanowski. The Alzheimer’s Disease Neuroimaging Initiative 3: Continued innovation for clinical trial improvement. Alzheimer’s & Dementia, 13(5):561–571, 2017.DOI
- [94]Andrew Gordon Wilson and Ryan Prescott Adams. Gaussian process kernels for pattern discovery and extrapolation. In International Conference on Machine Learning (ICML), pages 1067–1075, 2013.arxiv.org/abs/1302.4245
- [95]Tao Yang, Xiuying Wang, Hao Liu, Guanzhong Gong, Lian-Ming Wu, Yu-Ping Wang, and Lisheng Wang. Unsupervised anomaly detection in brain MRI via disentangled anatomy learning. Medical Image Analysis, 2026. arXiv:2512.21924.DOI
- [96]Vitjan Zavrtanik et al. DRAEM: A discriminatively trained reconstruction embedding for surface anomaly detection. In ICCV, 2021.arxiv.org/abs/2108.07610
- [97]Xinru Zhang, Ni Ou, Berke Doga Basaran, Marco Visentin, Mengyun Qiao, Renyang Gu, Paul M. Matthews, Yaou Liu, Chuyang Ye, and Wenjia Bai. A foundation model for lesion segmentation on brain MRI with mixture of modality experts. IEEE Transactions on Medical Imaging, 44(6):2594–2604, 2025. 2arxiv.org/abs/2405.10246