WorksComputer Science
ExiST: Existence-Aware Self-Training for Semi-Supervised Multi-Organ Segmentation
Abstract
Semi-supervised multi-organ segmentation should learn more than a dozen organs from a handful of labeled scans, and the smallest organs occupy less than one percent of the foreground. We identify a failure mode of self-training that is specific to this regime. When a teacher misses a small organ in a scan, every voxel of that organ receives the confident pseudo-label background, and the student is trained toward the organ absence. Repeated over many scans, this loop extinguishes tail organs. In a pilot study, removing the gallbladder from the pseudo-labels reduced its Dice on AMOS from 36.5 to 1.0. We argue that a missing organ should be treated as unresolved rather than absent and propose Existence-aware Self-Training (ExiST). It builds pseudo-labels by leave-one-out consensus of independently trained networks, screens them with a quality gate on validation scans, and infers, per scan and per organ, whether the organ was found. For every unresolved organ, the background voxels of that scan receive the set-valued target, trained with a marginal cross-entropy that accepts either answer and a Dice term that skips uncertain voxels. We show that this loss never pushes a candidate organ toward absence and never favors one member of the set, which leaves the decision to labeled evidence. ExiST outperforms the current state-of-the-art methods and reaches 77.1 mean Dice on Synapse with 20% labeled scans and 75.9 on AMOS with 5% labeled scans.
Introduction
Multi-organ segmentation of abdominal CT and MRI underlies radiotherapy planning and surgical navigation, and voxel-level annotation of a dozen organs per scan is among the most expensive labeling tasks in medical imaging [21, 28]. Semi-supervised segmentation reduces this cost by training on a few labeled scans together with many unlabeled
ones [2, 9, 45, 57], most often by self-training: a teacher, typically an exponential moving average (EMA) of the student or a co-trained sibling, predicts the unlabeled scans and the student reproduces its confident predictions [44, 56]. Multi-organ data make this regime unusually hard. The liver holds roughly half of the foreground voxels of the labeled scans of Synapse and AMOS, whereas the adrenal glands and the esophagus each hold less than one percent Figure 1. Pseudo-labels inherit and amplify this imbalance, and general methods such as UA-MT [57] and CPS [9] leave several tail organs at zero Dice (Tab. 1). Recent work debiases pseudo-label confidence, re-weights classes, balances heads, or routes organs to experts [32, 41, 46, 60].

Figure 1. Per-organ Dice (lines) on (a) Synapse (20% labeled) and (b) AMOS (5% labeled). Organs are ordered by their share of foreground voxels in the labeled scans (bars).
All of these methods share an assumption that we argue is a real bottleneck for tail organs: background is treated as certain. When the teacher misses a small organ entirely, there is no organ pseudo-label to re-weight. The organ’s voxels are confidently labeled background, the student is pushed toward the organ’s absence, the EMA teacher inherits the weaker student, and the organ is missed in more scans. We call this loop pseudo-label extinction. It is invisible to confidence filtering because the background labels that cause it are confident. A pilot study under the DHC protocol [46] shows the effect directly. When the gallbladder was removed from the pseudo-labels while its voxels stayed confident background, the mean test Dice of the student on AMOS fell from 37.4 to 29.7 and its gallbladder Dice from 36.5 to 1.0. Our central insight is that the right response to a missed organ is not a better background label but a weaker one. If the pseudo-labels of a scan contain far less of organ than that anatomy normally contains, organ was not resolved, and the student should not be told where it is absent. We therefore replace the single background label by a set-valued target. Every accepted background voxel of such a scan may be either background or organ , and the student is trained with the marginal likelihood of the set. A truly background voxel is still fitted, while a voxel of the missed organ no longer generates a gradient toward absence. Set-valued targets are established when the set of annotated organs is known per dataset [43, 61]. Here the set is inferred per scan and per organ from the pseudo-labeler itself. ExiST also decouples the pseudo-labeler from the student. Three networks are trained on the labeled scans with different seeds, each student is paired with the two networks that do not share its seed, a voxel is accepted only when both peers agree with high confidence, and a quality gate on the validation scans decides whether self-training may start.
Contributions:
We identify pseudo-label extinction as a failure mode of self-training for multi-organ segmentation that confidence filtering and class re-weighting cannot fix, and formalize it through the gradient of background pseudo-labels on a missed organ.
We propose existence-aware targets. A per-scan, per-organ test decides whether an organ was found, and unverified background is converted into set-valued targets trained with a marginal cross-entropy and an uncertainty-skipping Dice loss, which provably never decides within the set and never pushes a candidate organ toward absence. The targets are built from leave-one-out consensus pseudo-labels screened by a validation quality gate.
ExiST sets a new state-of-the-art on Synapse and AMOS datasets, with the significant improvements on small organs and single-network inference.
Related Work
Class-imbalanced semi-supervised medical image segmentation. General semi-supervised segmentation relies on EMA teachers [45, 57], consistency constraints [33, 54], mutual supervision [9, 40], data mixing [2, 5, 11, 17], and contrastive or generative terms [1, 18]. These methods treat confident background as correct and degrade sharply on 13 to 15 organs. Class-imbalanced classification corrects the class prior of pseudo-labels [7, 15, 25, 38, 49, 52, 53, 58], and for segmentation, CLD [29] and Basak et al. [3] re-weight classes, DHC [46] co-trains heterogeneous networks with debiased weights, and A&D [47] aggregates and decouples
heads. GA [41] modifies the gradient of the Dice and cross-entropy losses so that classes absent from a label still receive a useful signal, SKCDF [60] decouples a balanced head, and DCA [32] routes head, medium, and tail organs to asymmetric experts. All of these methods act on the organ side of the pseudo-labels, and none tests whether an organ was found in a scan. GA is closest in motivation, but it changes the loss on every voxel, whereas ExiST changes the target only where an organ is in doubt.
Set-valued targets and agreement-based pseudo-labels. Partially supervised segmentation faces a related problem. When datasets annotate different organs, the unannotated organs of a scan hide in its background. Prior-aware networks [61], marginal and exclusion losses [43], and later methods [10, 13, 22, 30, 31, 59] therefore train with a background class that marginalizes over the organs known to be unannotated. ExiST uses the same marginal likelihood, but the set of hidden organs is not dataset metadata. It is inferred per scan and per organ from the consensus of independent pseudo-labelers and applied to pseudo-labels rather than annotations. U2PL [50] uses unreliable pixels as negatives for the classes they are confidently not, whereas ExiST withdraws the negative for the class the voxel may be. Learning from candidate label sets [8, 12, 14, 23, 34, 48] and negative learning [20, 26] provide the theory. Agreement-based pseudo-labels go back to co-training [4, 24, 39, 42, 55] and tri-training [62], which label an example only when other learners agree. The leave-one-out consensus of ExiST follows tri-training, but it is applied once, offline, with frozen networks, and it yields set-valued targets. Foundation models [16, 35, 36, 51] could replace these peers.
Preliminaries
Setting. A scan is defined on the voxel grid after preprocessing, with label map over the classes , where is background and indexes the organs ( for Synapse, 15 for AMOS). We are given a small labeled set , a large unlabeled set , and a labeled validation set used for model selection and for the quality gate of § 4.2. A network maps a scan to per-voxel logits and posteriors , with prediction . On labeled patches we minimize , the sum of a class-weighted cross-entropy and a soft Dice loss that is standard for imbalanced 3D segmentation [19, 46]. The weight of class in decreases with the voxel share of the class in .
Self-training and pseudo-label extinction. Self-training adds a loss on unlabeled scans whose targets come from a pseudo-labeler , which can be an EMA teacher [45], a co-trained sibling [9], or, as here, a set of frozen networks. The usual target assigns one class per voxel and keeps only confident voxels,
and the unlabeled loss is with a cross-entropy. Class-imbalance methods modify , , or the sampling of organ voxels, but they keep the structure of Eq. (1): a voxel is either assigned one class or discarded. Consider an organ that fails to detect in scan , so that is small on the true voxels of the organ, , while . Every then receives the target 0 with , and the cross-entropy gradient on the organ logit is
Gradient descent lowers on the organ’s own voxels until . When is tied to the student, the weakened student can yield a weaker teacher, and the organ is missed in further scans. This pseudo-label extinction cannot be prevented by raising , because the offending targets are confident, nor by re-weighting organ classes, because no organ target exists on these voxels.
Method
Figure 2 shows ExiST, which has five stages. All scans are first reoriented to RAS, resampled to mm, and intensity-normalized to . (1) Three foundation networks , , are trained on only, with seeds 0, 1, 2, and then frozen. (2) For each student index , the two peers label every unlabeled scan, and a voxel is accepted only where both agree with high confidence (§ 4.1). (3) A quality gate measured on decides whether these pseudo-labels may be used and which organs are unreliable (§ 4.2). (4) Accepted background is relaxed into set-valued targets wherever an organ may have been missed (§ 4.3). (5) A student with the architecture of the foundation networks is trained from scratch on (§ 4.4). Inference uses alone.

Figure 2. Overview of ExiST. Frozen foundation networks produce leave-one-out consensus pseudo-labels, a quality gate screens them, and the local doubt (Rule A) and missing-organ (Rule B) relaxations turn accepted background into set-valued targets for a trained student.
Leave-one-out consensus pseudo-labels
For scan and peer , we average the prediction over the identity, the anterior-posterior flip, the superior-inferior flip, and their combination, and denote the resulting posterior by . The left-right flip is not used because it would exchange the left and right kidneys and adrenal glands. Let be the label predicted by peer and its confidence. The consensus label and the acceptance mask are
with , where is the one-voxel band on each side of every class boundary of the consensus map. The accepted set is , with organ parts and background part . All other voxels are ignored and receive no loss. We also keep the peer-maximum posterior , which records whether at least one peer sees some evidence for class and is used in § 4.3. Requiring the agreement of two independently trained networks is a form of disagreement-based uncertainty [27].
Quality gate
Pseudo-labels are useful only if their accepted organ voxels are mostly correct and if they add more correct information than they remove. We measure both properties by applying the pseudo-labeling of § 4.1 to . The precision of the accepted voxels of organ is
and the pooled precision is the same ratio computed over the accepted voxels of all organs. We further count an accepted voxel as useful if its pseudo-label is correct while the held-out network is wrong, and as harmful in the opposite case. The gate passes if the pooled precision is at least 0.85 on CT (0.80 on MRI) and the useful-to-harmful ratio is at least 1.5. If the gate fails on CT, self-training is not started, and if it fails only on MRI, the MRI scans are removed from . Independently of this decision, any organ with is gate-ignored: its accepted voxels are moved to the ignored set, and the organ is treated as unresolved in every scan.
Existence-aware targets
We replace the single label of an accepted voxel by a candidate set of classes that the student may predict without penalty. Accepted organ voxels keep a singleton, for and . Accepted background voxels start from and are enlarged by two relaxations. The first is local (Rule A in Figure 2), and the second acts on a whole scan (Rule B).
Local Doubt Relaxation (LDR). Let be the most likely organ at under the peers. If its probability is small but not negligible, the voxel may belong to a missed structure, and we admit the organ as a candidate:
with , far below . LDR does not label the voxel as organ. It only withdraws the claim that the organ is absent.
Missing-Organ Relaxation (MOR). LDR cannot help when an organ is missed entirely, because the peers then assign it a negligible probability everywhere. We therefore test, per scan and per organ, whether the organ was found at all. Let be the accepted volume of organ in scan , with the voxel volume, and let be the median volume of organ over . Organ is unresolved in scan if
where is the set of gate-ignored organs. Since counts only voxels that both peers confidently assign to the organ, an organ that is under-segmented by more than a factor of four is also unresolved. With , MOR enlarges every accepted background voxel of the scan:
After both relaxations, a background voxel carries where no organ is in doubt, where one organ is in doubt, and a larger set where several organs are unresolved. The target map is computed once per student. MOR is global within a scan, because an organ that was not found could be anywhere.
Learning objective
A candidate set is trained with the negative log-likelihood of the event that the label lies in the set,
which is the marginal loss of partial-label learning [12, 23] and of partially supervised segmentation [43]. For a singleton it is the ordinary cross-entropy, and for it accepts either answer. The Dice loss, in contrast, needs a binary reference per organ. For organ we define the voxels at which its membership is certain, , and compute the Dice term of only over with reference :
A voxel at which is allowed but not certain is excluded from the Dice term of organ . Otherwise the denominator would penalize any mass placed on a candidate organ and partially undo Eq. (8). Each iteration draws labeled and pseudo-labeled patches and minimizes
with and deep supervision on both terms at every decoder scale. uses the class weights on singleton voxels.
Analysis. Fix a voxel with logits , posterior , and candidate set , and write and .
Proposition 1. For every class , if , and otherwise. Moreover, depends on only through and , so the conditional distribution is unconstrained by .
Proof. Since , the second claim is immediate, and differentiating the two terms gives .
The first part states that gradient descent never decreases the logit of a candidate organ, in contrast to Eq. (2). The second part states that the loss carries no information about which member of the set is correct. The student is relieved of the claim that the organ is absent, but it is not told that the organ is present, and the division of the mass within the set is left to the labeled patches and the accepted organ voxels [12, 14]. Equation (9) preserves both properties, because the Dice term of a candidate organ does not depend on the voxels at which it is only allowed.
Training and inference
Alg. 1 summarizes the procedure. Foundation networks and students share the architecture, the optimizer, and an organ-balanced sampler, which draws an organ uniformly from and centers a patch on a random voxel of that organ (on unlabeled scans, a random accepted voxel). Students are trained from scratch and selected by validation Dice. At inference the student is applied in a sliding window of with 50% overlap and Gaussian blending, the predictions of the four flips are averaged, and the argmax is resampled to the native grid.
Experiments
Datasets and protocol. We follow the class-imbalanced protocol of DHC [46], which is also used by A&D, GA, SKCDF, and DCA, so that all numbers in Tabs. 1 and Table 2 share splits, label budgets, and evaluator. Synapse [28] (BTCV) has 30 abdominal CT scans with 13 organs: spleen (Sp), right and left kidney (RK, LK), gallbladder (Ga), esophagus (Es), liver (Li), stomach (St), aorta (Ao), inferior vena cava (IVC), portal and splenic veins (PSV), pancreas (Pa), and right and left adrenal gland (RAG, LAG). The split has 20 training, 4 validation, and 6 test scans, and 20% of the training scans (4) are labeled. We use the split released with SKCDF [60]. AMOS [21] has 360 CT and MRI scans with 15 organs: those above without PSV, plus duodenum (Du), bladder (Bl), and prostate/uterus (P/U). The DHC split has 216 training, 24 validation, and 120 test scans, with 5% of the training scans labeled. Dice (%) and average surface distance (ASD, in voxels of the evaluation grid) are computed per organ with the DHC evaluator and averaged over organs and test scans. ExiST is trained and applied at mm, and its predictions are resampled to the evaluation grid (in-plane halved for Synapse, for AMOS) before scoring. Organs that are not annotated in a scan (bladder and prostate/uterus in the MRI scans of AMOS) are scored against an empty reference, as in the evaluator.
Table 2. Comparison with state-of-the-art methods on AMOS (5% labeled). Best in bold, second best underlined. Mean over three seeds (std 0.07 Dice, 0.35 ASD), trained at mm and scored on the DHC grid. Other rows as reported by DCA [32].
| Method | Avg. Dice | Avg. ASD | Sp | RK | LK | Ga | Es | Li | St | Ao | IVC | PSV | RAG | LAG | Du | Bl | P/U |
| V-Net (fully supervised) [37] | 76.50 | 2.01 | 92.2 | 92.2 | 93.3 | 65.5 | 70.3 | 95.3 | 82.4 | 91.4 | 85.0 | 74.9 | 58.6 | 58.1 | 65.6 | 64.4 | 58.3 |
| General | |||||||||||||||||
| UA-MT [57] | 42.16 | 15.48 | 59.8 | 64.9 | 64.0 | 35.3 | 34.1 | 77.7 | 37.8 | 61.0 | 46.0 | 33.3 | 26.9 | 12.3 | 18.1 | 29.7 | 31.6 |
| URPC [33] | 44.93 | 27.44 | 67.0 | 64.2 | 67.2 | 36.1 | 0.0 | 83.1 | 45.5 | 65.4 | 54.4 | 46.7 | 0.0 | 29.4 | 35.2 | 44.5 | 33.2 |
| CPS [9] | 41.08 | 20.37 | 56.1 | 60.3 | 59.4 | 33.3 | 25.4 | 73.8 | 32.4 | 65.7 | 52.1 | 31.1 | 25.5 | 6.2 | 18.4 | 40.7 | 35.8 |
| SS-Net [54] | 33.88 | 54.72 | 65.4 | 68.3 | 69.9 | 37.8 | 0.0 | 75.1 | 32.3 | 68.0 | 56.6 | 33.5 | 0.0 | 0.0 | 0.0 | 0.2 | 0.8 |
| Co-BioNet [40] | 48.32 | 26.04 | 76.6 | 82.1 | 75.1 | 41.5 | 38.2 | 87.9 | 40.4 | 75.2 | 53.7 | 40.8 | 4.8 | 0.0 | 25.1 | 64.2 | 19.2 |
| BCP [2] | 45.67 | 28.12 | 62.3 | 66.1 | 62.5 | 34.8 | 18.7 | 79.4 | 38.2 | 64.3 | 49.1 | 35.6 | 22.4 | 15.2 | 22.8 | 35.1 | 32.9 |
| MagicNet [5] | 54.08 | 29.03 | 80.0 | 84.5 | 86.1 | 47.9 | 0.0 | 85.1 | 50.7 | 81.7 | 69.3 | 57.2 | 46.0 | 0.0 | 40.8 | 62.9 | 19.2 |
| ABD [11] | 50.23 | 22.45 | 68.7 | 70.4 | 68.2 | 39.5 | 25.3 | 84.2 | 42.1 | 69.8 | 54.7 | 41.2 | 28.6 | 21.8 | 29.4 | 42.3 | 38.7 |
| -FFT [17] | 54.80 | 18.76 | 74.2 | 75.8 | 73.6 | 45.2 | 32.1 | 88.9 | 48.7 | 87.3 | 48.7 | 47.8 | 35.2 | 28.9 | 37.0 | 45.9 | 45.1 |
| Imbalance | |||||||||||||||||
| Adsh [15] | 40.33 | 24.53 | 56.0 | 63.6 | 57.3 | 34.7 | 25.7 | 73.9 | 30.7 | 65.7 | 51.9 | 27.1 | 20.2 | 0.0 | 18.6 | 43.5 | 35.9 |
| CReST [52] | 46.55 | 14.62 | 66.5 | 64.2 | 65.4 | 36.0 | 32.2 | 77.8 | 43.6 | 68.5 | 52.9 | 40.3 | 24.7 | 19.5 | 26.5 | 43.9 | 36.4 |
| SimiS [6] | 47.27 | 11.51 | 77.4 | 72.5 | 68.7 | 32.1 | 14.7 | 86.6 | 46.3 | 74.6 | 52.4 | 41.6 | 24.4 | 17.9 | 21.9 | 47.9 | 28.2 |
| Basak et al. [3] | 38.74 | 31.76 | 68.8 | 59.0 | 54.2 | 29.0 | 0.0 | 83.7 | 39.3 | 61.7 | 52.1 | 34.6 | 0.0 | 0.0 | 26.8 | 45.7 | 26.2 |
| CLD [29] | 46.10 | 15.86 | 67.2 | 68.5 | 71.4 | 41.0 | 21.0 | 76.1 | 42.4 | 69.8 | 52.1 | 37.9 | 24.7 | 23.4 | 22.7 | 38.1 | 35.2 |
| DHC [46] | 49.53 | 13.89 | 68.1 | 69.0 | 67.1 | 42.3 | 37.0 | 78.6 | 43.8 | 70.8 | 57.4 | 43.2 | 27.0 | 28.7 | 29.1 | 44.1 | 36.7 |
| A&D [47] | 50.03 | 5.21 | 73.1 | 76.0 | 76.5 | 29.1 | 44.9 | 82.5 | 49.0 | 72.8 | 61.7 | 48.5 | 30.2 | 19.7 | 36.4 | 32.9 | 18.7 |
| GA [41] | 63.51 | 4.58 | 78.9 | 85.5 | 87.2 | 50.0 | 49.1 | 86.9 | 56.2 | 83.4 | 70.3 | 57.4 | 49.1 | 40.8 | 38.3 | 71.6 | 47.9 |
| SKCDF [60] | 53.81 | 5.97 | 77.1 | 77.9 | 71.2 | 34.1 | 50.4 | 88.6 | 51.6 | 80.9 | 58.9 | 48.8 | 33.0 | 30.2 | 32.2 | 45.9 | 26.4 |
| DCA [32] | 69.90 | 2.66 | 83.7 | 87.8 | 90.6 | 59.6 | 63.6 | 91.3 | 66.4 | 87.0 | 76.7 | 64.0 | 55.8 | 48.5 | 48.1 | 72.7 | 51.1 |
| ExiST | 75.91 | 3.40 | 93.1 | 93.0 | 94.6 | 64.3 | 75.6 | 94.9 | 73.7 | 92.2 | 87.4 | 78.0 | 67.9 | 66.0 | 63.3 | 54.3 | 40.4 |
Table 2.
Baselines. We compare against the 19 semi-supervised methods reported by DCA [32] under this protocol, nine general and ten class-imbalance methods, and against a V-Net trained on all labels with the recipe of the protocol. These numbers are taken from DCA, where each method uses its own recipe and input resolution. Since ExiST uses a residual V-Net at mm, row (a) of Table 3 reports our network trained on the labeled scans alone, which isolates the contribution of backbone and resolution. For Figure 3 and Figure 4 we ran DHC, SKCDF, GA, and MagicNet with released code and BCP through our multi-organ 3D port.
| Variant | Dice | ASD | Tail |
| (a) Foundation only, no self-training | 63.1 | ||
| (b) (a) + consensus, no relaxation | 63.9 | ||
| (c) (b) + LDR only† | 64.8 | ||
| (d) (b) + MOR only† | 66.2 | ||
| (e) (b) + LDR + MOR, standard Dice† | 63.8 | ||
| (f) Full† | 66.3 | ||
| (g) Full, peers include † | 66.9 | ||
| (h) Full, single peer† | 64.4 | ||
| Full ExiST | 65.9 |
Table 3. Component ablation on Synapse (20% labeled). Tail is the mean Dice over Es, Pa, RAG, and LAG. †Earlier gate configuration, under which the gallbladder is gate-ignored.

Figure 3. Qualitative comparison on the Synapse test scan with the median Dice of ExiST (20% labeled). The slices pass through the reference pancreas, and the white boxes mark the enlarged regions.

Figure 4. t-SNE of per-voxel features before the classifier on the Synapse test set (40 voxels per organ and scan).
Implementation. All networks are residual V-Nets [37] with five levels of 32 to 512 channels, instance normalization, and deep supervision at four decoder scales. CT intensities are clipped to HU and MRI intensities to the 0.5 to 99.5 percentile range before scaling to . Foundation networks and students are trained for 40k iterations with AdamW (learning rate $3\times10^{-4}$, weight decay , warm-up followed by cosine decay) on batches of two labeled and, for students, two pseudo-labeled patches. Augmentation consists of the two flips of § 4.1, intensity perturbations, and mild rotation and scaling. The class weights are $w_c\propto for organs, with the voxel count of class in , and 0.1 for background. We use , , , , and the gate thresholds of § 4.2 on both benchmarks. Results are the mean and standard deviation over the three students (seeds 0, 1, 2). The network has 31.0M parameters. On one NVIDIA A100, the pipeline for one student, including the three foundation networks, takes 11.2 GPU hours on Synapse and 31.1 on AMOS, against 2.3 and 5.4 for the labeled-only network. Inference takes 15.1 and 18.7 seconds per scan in both cases.
Table 1. Comparison with state-of-the-art methods on Synapse (20% labeled), meanstd over three runs. Best in bold, second best underlined. Trained at mm and scored on the DHC grid. Other rows as reported by DCA [32].
| Method | Avg. Dice | Avg. ASD | Sp | RK | LK | Ga | Es | Li | St | Ao | IVC | PSV | Pa | RAG | LAG |
| V-Net (fully supervised) [37] | 62.091.2 | 10.283.9 | 84.6 | 77.2 | 73.8 | 73.3 | 38.2 | 94.6 | 68.4 | 72.1 | 71.2 | 58.2 | 48.5 | 17.9 | 29.0 |
| General | |||||||||||||||
| UA-MT [57] | 20.262.2 | 71.677.4 | 48.2 | 31.7 | 22.2 | 0.0 | 0.0 | 81.2 | 29.1 | 23.3 | 27.5 | 0.0 | 0.0 | 0.0 | 0.0 |
| URPC [33] | 25.685.1 | 72.7415.5 | 66.7 | 38.2 | 56.8 | 0.0 | 0.0 | 85.3 | 33.9 | 33.1 | 14.8 | 0.0 | 5.1 | 0.0 | 0.0 |
| CPS [9] | 33.553.7 | 41.219.1 | 62.8 | 55.2 | 45.4 | 35.9 | 0.0 | 91.1 | 31.3 | 41.9 | 49.2 | 8.8 | 14.5 | 0.0 | 0.0 |
| SS-Net [54] | 35.082.8 | 50.816.5 | 62.7 | 67.9 | 60.9 | 34.3 | 0.0 | 89.9 | 20.9 | 61.7 | 44.8 | 0.0 | 8.7 | 4.2 | 0.0 |
| Co-BioNet [40] | 58.832.7 | 7.505.8 | 82.8 | 90.0 | 86.5 | 11.6 | 19.5 | 92.3 | 47.7 | 77.5 | 77.7 | 51.3 | 30.3 | 47.5 | 50.2 |
| BCP [2] | 50.231.5 | 18.452.3 | 72.4 | 58.1 | 52.7 | 38.2 | 12.5 | 92.3 | 35.6 | 48.9 | 45.2 | 18.7 | 22.1 | 8.4 | 5.2 |
| MagicNet [5] | 60.572.5 | 22.486.3 | 82.5 | 91.0 | 89.5 | 11.2 | 0.0 | 89.4 | 62.7 | 77.6 | 79.0 | 66.1 | 47.3 | 36.8 | 54.3 |
| ABD [11] | 55.671.2 | 15.321.8 | 76.8 | 64.3 | 58.9 | 42.5 | 18.9 | 93.7 | 41.2 | 55.4 | 51.6 | 24.3 | 28.7 | 14.6 | 11.8 |
| -FFT [17] | 59.120.9 | 12.181.1 | 81.2 | 70.5 | 65.4 | 48.7 | 25.3 | 95.1 | 47.8 | 62.1 | 58.3 | 30.9 | 35.2 | 21.5 | 18.4 |
| Imbalance | |||||||||||||||
| Adsh [15] | 35.290.5 | 39.614.6 | 55.1 | 59.6 | 45.8 | 52.2 | 0.0 | 89.4 | 32.8 | 47.6 | 53.0 | 8.9 | 14.4 | 0.0 | 0.0 |
| CReST [52] | 38.333.4 | 22.859.0 | 62.1 | 64.7 | 53.8 | 43.8 | 8.1 | 85.9 | 27.2 | 54.4 | 47.7 | 14.4 | 13.0 | 18.7 | 4.6 |
| SimiS [6] | 40.070.6 | 32.980.5 | 62.3 | 69.4 | 50.7 | 61.4 | 0.0 | 87.0 | 33.0 | 59.0 | 57.2 | 29.2 | 11.8 | 0.0 | 0.0 |
| Basak et al. [3] | 33.240.6 | 43.782.5 | 57.4 | 53.8 | 48.5 | 46.9 | 0.0 | 87.8 | 28.7 | 42.3 | 45.4 | 6.3 | 15.0 | 0.0 | 0.0 |
| CLD [29] | 41.071.2 | 32.153.3 | 62.0 | 66.0 | 59.3 | 61.5 | 0.0 | 89.0 | 31.7 | 62.8 | 49.4 | 28.6 | 18.5 | 0.0 | 5.0 |
| DHC [46] | 48.610.9 | 10.712.6 | 62.8 | 69.5 | 59.2 | 66.0 | 13.2 | 85.2 | 36.9 | 67.9 | 61.5 | 37.0 | 30.9 | 31.4 | 10.6 |
| A&D [47] | 60.880.7 | 2.520.4 | 85.2 | 66.9 | 67.0 | 52.7 | 62.9 | 89.6 | 52.1 | 83.0 | 74.9 | 41.8 | 43.4 | 44.8 | 27.2 |
| GA [41] | 68.430.5 | 3.110.2 | 81.4 | 92.4 | 90.8 | 33.5 | 53.3 | 89.1 | 60.9 | 79.1 | 82.1 | 66.7 | 48.7 | 50.3 | 61.4 |
| SKCDF [60] | 64.271.4 | 1.450.1 | 79.5 | 72.1 | 67.6 | 59.8 | 60.7 | 93.3 | 61.7 | 85.4 | 78.5 | 41.8 | 50.9 | 46.4 | 37.8 |
| DCA [32] | 73.200.3 | 1.780.1 | 82.9 | 92.8 | 91.4 | 64.0 | 61.4 | 92.5 | 67.2 | 83.9 | 82.7 | 68.1 | 52.2 | 51.8 | 63.7 |
| ExiST | 77.130.4 | 2.600.2 | 92.7 | 76.7 | 75.7 | 83.6 | 74.1 | 95.3 | 69.9 | 89.1 | 86.6 | 69.4 | 55.6 | 70.4 | 63.7 |
Table 1.
Results
Comparison with the state of the art
Synapse. With 20% labeled scans (Tab. 1), ExiST reaches 77.13 mean Dice, against 73.20 for DCA, 68.43 for GA, and 64.27 for SKCDF. It obtains the best Dice on 11 of the 13 organs, including a tie with DCA on the left adrenal gland. The largest gains over DCA are on small organs: 19.6 Dice on the gallbladder (83.6 against 64.0), 18.6 on the right adrenal gland (70.4 against 51.8), and 12.7 on the esophagus (74.1 against 61.4). The two kidneys are the exception: ExiST reaches 76.7 and 75.7 Dice, close to the fully supervised V-Net (77.2 and 73.8) but well below DCA (92.8 and 91.4).
AMOS. With 5% labeled scans (Table 2), ExiST reaches 75.91 mean Dice, against 69.90 for DCA and 63.51 for GA, which is within 0.6 Dice of the fully supervised V-Net (76.50). ExiST is the best method on 13 of the 15 organs, and the gains over DCA again grow as the organs become smaller: 17.5 Dice on the left adrenal gland, 17.2 on the duodenum, 14.0 on the pancreas, 12.1 on the right adrenal gland, and 12.0 on the esophagus. Its ASD (3.40) is the second lowest after DCA (2.66). DCA has released neither code nor per-case results, which rules out a paired test. Against DHC, re-run with its released code, ExiST is higher on all 120 test scans (mean difference 24.0 Dice, Wilcoxon signed-rank ). The bladder (54.3 against 72.7 for DCA) and the prostate/uterus (40.4 against 51.1) are the two organs on which ExiST is worse. The bladder is the one organ that the quality gate ignores on AMOS, so it never receives an organ pseudo-label and is treated as unresolved in every scan. Both organs are also not annotated in the MRI scans, where MOR still permits the student to predict them.
Organ size. Figure 1 relates these results to organ size, with the organs sorted by their share of the labeled foreground. On the right of the plot, where the share is below one percent, the curve of ExiST lies at or above those of DCA, GA, SKCDF, and DHC, with the prostate/uterus on AMOS as the only exception. The baselines drop sharply there, whereas ExiST stays between 63 and 84 Dice on Synapse and above 65 Dice on the esophagus and adrenal glands of AMOS. Part of the margin over the published numbers comes from the backbone and resolution.
Qualitative comparison
Figure 3 shows the Synapse test scan with the median Dice of ExiST, fixed before inspection. In the axial view, ExiST segments the gallbladder with a shape close to the reference. BCP misses it, GA and MagicNet recover only a fragment, and DHC produces scattered false positives of several organs. In the coronal view, the portal and splenic veins and the two visible parts of the pancreas predicted by ExiST follow the reference, whereas BCP and GA segment the pancreas only partially and SKCDF extends it well beyond the reference. The remaining errors of ExiST are also visible: the stomach is under-segmented in the axial slice, and only a fragment of the pancreatic tail is found.
Feature-space analysis
Figure 4 visualizes the per-voxel features of the penultimate layer on the Synapse test set. For ExiST, most organs form compact and separate clusters, including the gallbladder, the esophagus, and the left adrenal gland. The pancreas, the stomach, and part of the right adrenal gland overlap in the center of the plot, which matches their lower Dice in
Table 1 (PDF p. 7). The features of DHC and BCP mix most organs in one central cloud. The leave-one-out 5-nearest-neighbor organ accuracy in the original feature space summarizes this impression: 91.0% for ExiST, against 89.2% for GA, 88.5% for MagicNet, 80.7% for SKCDF, 71.2% for DHC, and 68.3% for BCP. Part of the difference may come from the input resolution of each method.
Ablation Studies
Effectiveness of key components. Table 3 separates the ingredients of ExiST on Synapse. Row (a) is the foundation network trained on the four labeled scans alone. It reaches 74.23 Dice, which is already above the published methods of Tab. 1, so the following rows must be read against this strong baseline. Self-training with consensus pseudo-labels and no relaxation (b) adds 1.0 Dice and 0.8 tail Dice. The full method adds another 1.9 Dice and 2.0 tail Dice and reaches the lowest ASD of the table (2.60). In total, ExiST gains 2.9 Dice and 2.8 tail Dice over its labeled-only baseline.
Rows (c) to (h) were run with an earlier configuration of the gate, under which the gallbladder was gate-ignored on Synapse, and are compared with row (f), the full method under that configuration. Among them, LDR alone (c) has the highest mean Dice (75.77) and the lowest ASD (2.68). MOR alone (d) reaches a higher tail Dice (66.2 against 64.8) but a lower mean Dice (72.83) and an ASD of 6.71, so under this configuration it recovers small organs at the price of false positives. Replacing the uncertainty-skipping Dice by the standard Dice (e) removes this price (ASD 3.05) but also the gain on tail organs (63.8 against 66.3 in f), which confirms that a standard Dice term penalizes the mass that the marginal loss permits. The agreement of two peers matters, since a single peer (h) loses 1.7 Dice and 1.9 tail Dice relative to (f). Including the own-seed network among the peers (g) changes the result by less than one standard deviation, so the leave-one-out exclusion is a safeguard rather than a source of the gain.
Gate-ignored organs. Row (f) and the full method differ only in the gate configuration, which changes mean Dice from 73.67 to 77.13 and ASD from 5.79 to 2.60. With the gallbladder gate-ignored, its accepted voxels are discarded and MOR is applied to it in every scan, so the organ loses all positive pseudo-labels while every background voxel may become gallbladder. Relaxation thus helps for organs that the peers actually miss, whereas a blanket relaxation of an organ that they mostly find is harmful, as for the bladder on AMOS (Table 2). The gate ratio is also informative across seeds. On Synapse the pooled precision is at least 0.97 for all three peer pairs, and the useful-to-harmful ratios are 0.77, 7.09, and 10.04. The first value is below the threshold of 1.5, but we trained this student nevertheless, and it is the weakest of the three (76.72 Dice against 76.94 and 77.73).
Pseudo-labeler design. Table 4 compares the frozen peers with a single EMA teacher that is trained with our recipe and hard pseudo-labels on AMOS. The EMA teacher is only 0.07 Dice better than its student, so its pseudo-labels carry little information that the student does not already have. Its useful-to-harmful ratio is higher than that of the consensus (3.76 vs. 2.21), but the gallbladder gains only 0.3 Dice during self-training, against 5.5 Dice with ExiST. Under our recipe the EMA teacher does not extinguish the gallbladder, because the labeled-only network already finds it in most scans.
| Pseudo-labeler | Dice | Useful:harm. | Ga Dice |
| EMA teacher, hard labels | +0.07 | 3.76 | 65.8 → 66.1 |
| Consensus + LDR + MOR (ours) | n/a | 2.21 | 65.8 → 71.3 |
Table 4. Pseudo labeler design on AMOS. Dice is teacher minus student, and Ga Dice is given before and after self-training.
Relaxation thresholds. We varied one threshold at a time on Synapse in the configuration of row (f) of Table 3 (73.67 Dice at the defaults). The acceptance confidence and 0.95 gives 72.44 and 73.57, the doubt threshold , 0.05, and 0.10 gives 73.24, 73.83, and 72.83, and the existence ratio , 0.10, and 0.50 gives 73.34, 72.44, and 73.41. All settings lie within 1.4 Dice, about twice the std.
over seeds, so the method is not sensitive to its thresholds.
Conclusion
Standard self-training tells the student that a missed organ is absent, which extinguishes small organs. ExiST replaces this claim by unresolved through gated consensus pseudo-labels, a per-scan and per-organ existence test, and set-valued targets whose marginal loss never pushes a candidate organ toward absence. On Synapse and AMOS, ExiST improves mean Dice by 3.9 and 6.0 points, with the largest gains on small organs and single-network inference. Since the relaxation hurts when applied to an organ in every scan, future existence tests should also use anatomy and modality.
LLM Usage Statement
Generative AI tools were used for language editing of the manuscript and to assist with the literature search. They were not used to generate experimental results, numerical values, or the statements and proofs of the theoretical results, all of which were derived by the authors. The authors take full responsibility for the final content of this work, including all text, claims and artifacts.
References
- [1]Maregu Assefa, Muzammal Naseer, Iyakkutti Iyappan Gana- pathi, Syed Sadaf Ali, Mohamed L. Seghier, and Naoufel Werghi. DyCON: Dynamic uncertainty-aware consistency and contrastive learning for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 30850–30860, 2025. 2arxiv.org/abs/2504.04566
- [2]Yunhao Bai, Duowen Chen, Qingli Li, Wei Shen, and Yan Wang. Bidirectional copy-paste for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11514–11524, 2023. 1, 2, 6arxiv.org/abs/2305.00673
- [3]Hritam Basak, Sagnik Ghosal, and Ram Sarkar. Addressing class imbalance in semi-supervised image segmentation: A study on cardiac MRI. In Medical Image Computing and Computer Assisted Intervention (MICCAI), pages 224–233, 2022. 2, 6arxiv.org/abs/2209.00123
- [4]Avrim Blum and Tom Mitchell. Combining labeled and unlabeled data with co-training. In Proceedings of the Eleventh Annual Conference on Computational Learning Theory (COLT), pages 92–100, 1998. 2
- [5]Duowen Chen, Yunhao Bai, Wei Shen, Qingli Li, Lequan Yu, and Yan Wang. Magicnet: Semi-supervised multi-organ segmentation via magic-cube partition and recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23869–23878, 2023. 2, 6arxiv.org/abs/2212.14310
- [6]Hao Chen, Yue Fan, Yidong Wang, Jindong Wang, Bernt Schiele, Xing Xie, Marios Savvides, and Bhiksha Raj. An embarrassingly simple baseline for imbalanced semi-supervised learning. arXiv preprint arXiv:2211.10806, 2022. 6
- [7]Hao Chen, Ran Tao, Yue Fan, Yidong Wang, Jindong Wang, Bernt Schiele, Xing Xie, Bhiksha Raj, and Marios Savvides. SoftMatch: Addressing the quantity-quality tradeoff in semi-supervised learning. In International Conference on Learning Representations (ICLR), 2023. 2arxiv.org/abs/2301.10921
- [8]Hao Chen, Ankit Shah, Jindong Wang, Ran Tao, Yidong Wang, Xiang Li, Xing Xie, Masashi Sugiyama, Rita Singh, and Bhiksha Raj. Imprecise label learning: A unified framework for learning with various imprecise label configurations. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2arxiv.org/abs/2305.12715
- [9]Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2613–2622, 2021. 1, 2, 3, 6arxiv.org/abs/2106.01226
- [10]Xiaoyang Chen, Hao Zheng, Yuemeng Li, Yuncong Ma, Liang Ma, Hongming Li, and Yong Fan. Versatile medical image segmentation learned from multi-source datasets via model self-disambiguation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11747–11756, 2024. 2arxiv.org/abs/2311.10696
- [11]Hanyang Chi, Jian Pang, Bingfeng Zhang, and Weifeng Liu. Adaptive bidirectional displacement for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4070–4080, 2024. 2, 6arxiv.org/abs/2405.00378
- [12]Timothee Cour, Ben Sapp, and Ben Taskar. Learning from partial labels. Journal of Machine Learning Research, 12: 1501–1536, 2011. 2, 4, 5
- [13]Xi Fang and Pingkun Yan. Multi-organ segmentation over partially labeled datasets with multi-scale feature abstraction. IEEE Transactions on Medical Imaging, 39(11):3619–3629, 2020. 2arxiv.org/abs/2001.00208
- [14]Lei Feng, Jiaqi Lv, Bo Han, Miao Xu, Gang Niu, Xin Geng, Bo An, and Masashi Sugiyama. Provably consistent partial-label learning. In Advances in Neural Information Processing Systems (NeurIPS), pages 10948–10960, 2020. 2, 5arxiv.org/abs/2007.08929
- [15]Lan-Zhe Guo and Yu-Feng Li. Class-imbalanced semi-supervised learning with adaptive thresholding. In International Conference on Machine Learning (ICML), pages 8082–8094, 2022. 2, 6
- [16]Yufan He, Pengfei Guo, Yucheng Tang, Andriy Myronenko, Vishwesh Nath, Ziyue Xu, Dong Yang, Can Zhao, Benjamin Simon, Mason Belue, Stephanie Harmon, Baris Turkbey, Daguang Xu, and Wenqi Li. VISTA3D: A unified segmentation foundation model for 3d medical imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20863–20873, 2025. 2arxiv.org/abs/2406.05285
- [17]Ming Hu, Jianfu Yin, Zhuangzhuang Ma, Jianheng Ma, Feiyu Zhu, Binbing Wu, Ya Wen, Meng Wu, Cong Hu, Bingliang Hu, and Quan Wang. B-FFT: Nonlinear interpolation and differentiated training strategies for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 30839–30849, 2025. 2, 6DOI
- [18]Kaiwen Huang, Yi Zhou, Yizhe Zhang, Jingxiong Li, and Tao Zhou. SemiGDA: Generative dual-distribution alignment for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026. 2arxiv.org/abs/2604.23274
- [19]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. 2DOI
- [20]Takashi Ishida, Gang Niu, Weihua Hu, and Masashi Sugiyama. Learning from complementary labels. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2
- [21]Yuanfeng Ji, Haotian Bai, Chongjian Ge, Jie Yang, Ye Zhu, Ruimao Zhang, Zhen Li, Lingyan Zhang, Wanling Ma, Xiang Wan, and Ping Luo. AMOS: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. In Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, pages 36722–36732, 2022. 1, 5
- [22]Xixi Jiang, Dong Zhang, Xiang Li, Kangyi Liu, Kwang-Ting Cheng, and Xin Yang. Labeled-to-unlabeled distribution alignment for partially-supervised multi-organ medical image segmentation. Medical Image Analysis, 99:103333, 2025. 2arxiv.org/abs/2409.03228
- [23]Rong Jin and Zoubin Ghahramani. Learning with multiple labels. In Advances in Neural Information Processing Systems (NeurIPS), pages 921–928, 2002. 2, 4
- [24]Zhanghan Ke, Di Qiu, Kaican Li, Qiong Yan, and Rynson W. H. Lau. Guided collaborative training for pixel-wise semi-supervised learning. In European Conference on Computer Vision (ECCV), pages 429–445, 2020. 2arxiv.org/abs/2008.05258
- [25]Jaehyung Kim, Youngbum Hur, Sejun Park, Eunho Yang, Sung Ju Hwang, and Jinwoo Shin. Distribution aligning refinery of pseudo-label for imbalanced semi-supervised learning. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2arxiv.org/abs/2007.08844
- [26]Youngdong Kim, Junho Yim, Juseung Yun, and Junmo Kim. NLNL: Negative learning for noisy labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 101–110, 2019. 2arxiv.org/abs/1908.07387
- [27]Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In Advances in Neural Information Processing Systems (NeurIPS), pages 6402–6413, 2017. 3arxiv.org/abs/1612.01474
- [28]Bennett Landman, Zhoubing Xu, Juan Eugenio Iglesias, Martin Styner, Thomas Robin Langerak, and Arno Klein. MICCAI multi-atlas labeling beyond the cranial vault – workshop and challenge. In Proceedings of the MICCAI Multi-Atlas Labeling Beyond the Cranial Vault – Workshop and Challenge, 2015. 1, 5
- [29]Yiqun Lin, Huifeng Yao, Zezhong Li, Guoyan Zheng, and Xiaomeng Li. Calibrating label distribution for class-imbalanced barely-supervised knee segmentation. In Medical Image Computing and Computer Assisted Intervention (MICCAI), pages 109–118, 2022. 2, 6arxiv.org/abs/2205.03644
- [30]Han Liu, Zhoubing Xu, Ruqiang Gao, Hao Li, Jianing Wang, Guillaume Chabin, Ipek Oguz, and Sasa Grbic. COSST: Multi-organ segmentation with partially labeled datasets using comprehensive supervisions and self-training. IEEE Transactions on Medical Imaging, 43(5):1995–2009, 2024. 2DOI
- [31]Jie Liu, Yixiao Zhang, Jie-Neng Chen, Junfei Xiao, Yongyi Lu, Bennett A. Landman, Yixuan Yuan, Alan Yuille, Yucheng Tang, and Zongwei Zhou. CLIP-driven universal model for organ segmentation and tumor detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21152–21164, 2023. 2DOI
- [32]Yajun Liu. Divide, conquer, and aggregate: Asymmetric experts for class-imbalanced semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8503–8513, 2026. 1, 2, 5, 6
- [33]Xiangde Luo, Wenjun Liao, Jieneng Chen, Tao Song, Yinan Chen, Shichuan Zhang, Nianyon Chen, Guotai Wang, and Shaoting Zhang. Efficient semi-supervised gross target volume of nasopharyngeal carcinoma segmentation via uncertainty rectified pyramid consistency. In Medical Image Computing and Computer Assisted Intervention (MICCAI), pages 318–329, 2021. 2, 6arxiv.org/abs/2012.07042
- [34]Jiaqi Lv, Miao Xu, Lei Feng, Gang Niu, Xin Geng, and Masashi Sugiyama. Progressive identification of true labels for partial-label learning. In International Conference on Machine Learning (ICML), pages 6500–6510, 2020. 2arxiv.org/abs/2002.08053
- [35]Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15:654, 2024. 2
- [36]Qinghe Ma, Jian Zhang, Zekun Li, Lei Qi, Qian Yu, and Yinghuan Shi. Steady progress beats stagnation: Mutual aid of foundation and conventional models in mixed domain semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5175–5185, 2025. 2arxiv.org/abs/2503.16997
- [37]Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-Net: Fully convolutional neural networks for volumetric medical image segmentation. In International Conference on 3D Vision (3DV), pages 565–571, 2016. 5, 6
- [38]Youngtaek Oh, Dong-Jin Kim, and In So Kweon. DASO: Distribution-aware semantics-oriented pseudo-label for imbalanced semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9786–9796, 2022. 2arxiv.org/abs/2106.05682
- [39]Yassine Ouali, Céline Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12674–12684, 2020. 2arxiv.org/abs/2003.09005
- [40]Himashi Peiris, Munawar Hayat, Zhaolin Chen, Gary Egan, and Mehrtash Harandi. Uncertainty-guided dual-views for semi-supervised volumetric medical image segmentation. Nature Machine Intelligence, 5(7):724–738, 2023. 2, 6
- [41]Wenbo Qi, Jiafei Wu, and S. C. Chan. Gradient-aware for class-imbalanced semi-supervised medical image segmentation. In European Conference on Computer Vision (ECCV), pages 473–490, 2024. 1, 2, 6DOI
- [42]Siyuan Qiao, Wei Shen, Zhishuai Zhang, Bo Wang, and Alan Yuille. Deep co-training for semi-supervised image recogni- tion. In European Conference on Computer Vision (ECCV), pages 135–152, 2018. 2
- [43]Gonglei Shi, Li Xiao, Yang Chen, and S. Kevin Zhou. Marginal loss and exclusion loss for partially supervised multi-organ segmentation. Medical Image Analysis, 70: 101979, 2021. 2, 4
- [44]Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A. Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. FixMatch: Simplifying semi-supervised learning with consistency and confidence. In Advances in Neural Information Processing Systems (NeurIPS), pages 596–608, 2020. 1
- [45]Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Advances in Neural Information Processing Systems (NeurIPS), pages 1195–1204, 2017. 1, 2, 3arxiv.org/abs/1703.01780
- [46]Haonan Wang and Xiaomeng Li. DHC: Dual-debiased heterogeneous co-training framework for class-imbalanced semi-supervised medical image segmentation. In Medical Image Computing and Computer Assisted Intervention (MICCAI), pages 582–591, 2023. 1, 2, 5, 6arxiv.org/abs/2307.11960
- [47]Haonan Wang and Xiaomeng Li. Towards generic semi-supervised framework for volumetric medical image segmentation. In Advances in Neural Information Processing Systems (NeurIPS), pages 1833–1848, 2023. 2, 6arxiv.org/abs/2310.11320
- [48]Haobo Wang, Ruixuan Xiao, Yixuan Li, Lei Feng, Gang Niu, Gang Chen, and Junbo Zhao. CoCo: Contrastive label disambiguation for partial label learning. In International Conference on Learning Representations (ICLR), 2022. 2arxiv.org/abs/2201.08984
- [49]Xudong Wang, Zhirong Wu, Long Lian, and Stella X. Yu. Debiased learning from naturally imbalanced pseudo-labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14647–14657, 2022. 2arxiv.org/abs/2201.01490
- [50]Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudo-labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4248–4257, 2022. 2arxiv.org/abs/2203.03884
- [51]Jakob Wasserthal, Hanns-Christian Breit, Manfred T. Meyer, Maurice Pradella, Daniel Hinck, Alexander W. Sauter, Tobias Heye, Daniel T. Boll, Joshy Cyriac, Shan Yang, Michael Bach, and Martin Segeroth. TotalSegmentator: Robust segmentation of 104 anatomic structures in CT images. Radiology: Artificial Intelligence, 5(5):e230024, 2023. 2DOI
- [52]Chen Wei, Kihyuk Sohn, Clayton Mellina, Alan Yuille, and Fan Yang. CReST: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10857–10866, 2021. 2, 6arxiv.org/abs/2102.09559
- [53]Tong Wei and Kai Gan. Towards realistic long-tailed semi-supervised learning: Consistency is all you need. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3469–3478, 2023. 2DOI
- [54]Yicheng Wu, Zhonghua Wu, Qianyi Wu, Zongyuan Ge, and Jianfei Cai. Exploring smoothness and class-separation for semi-supervised medical image segmentation. In Medical Image Computing and Computer Assisted Intervention (MICCAI), pages 34–43, 2022. 2, 6arxiv.org/abs/2203.01324
- [55]Yingda Xia, Dong Yang, Zhiding Yu, Fengze Liu, Jinzheng Cai, Lequan Yu, Zhuotun Zhu, Daguang Xu, Alan Yuille, and Holger Roth. Uncertainty-aware multi-view co-training for semi-supervised medical image segmentation and domain adaptation. Medical Image Analysis, 65:101766, 2020. 2DOI
- [56]Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7236–7246, 2023. 1arxiv.org/abs/2208.09910
- [57]Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In Medical Image Computing and Computer Assisted Intervention (MICCAI), pages 605–613, 2019. 1, 2, 6arxiv.org/abs/1907.07034
- [58]Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. FlexMatch: Boosting semi-supervised learning with curriculum pseudo labeling. In Advances in Neural Information Processing Systems (NeurIPS), pages 18408–18419, 2021. 2
- [59]Jianpeng Zhang, Yutong Xie, Yong Xia, and Chunhua Shen. DoDNet: Learning to segment multi-organ and tumors from multiple partially labeled datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1195–1204, 2021. 2arxiv.org/abs/2011.10217
- [60]Zheng Zhang, Guanchun Yin, Bo Zhang, Wu Liu, Xiuzhuang Zhou, and Wendong Wang. A semantic knowledge complementarity based decoupling framework for semi-supervised class-imbalanced medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25940–25949, 2025. 1, 2, 5, 6DOI
- [61]Yujin Zhou, Zhe Li, Song Bai, Chong Wang, Xinlei Chen, Mei Han, Elliot Fishman, and Alan L. Yuille. Prior-aware neural network for partially-supervised multi-organ segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10672–10681, 2019. 2arxiv.org/abs/1904.06346
- [62]Zhi-Hua Zhou and Ming Li. Tri-training: Exploiting unlabeled data using three classifiers. IEEE Transactions on Knowledge and Data Engineering, 17(11):1529–1541, 2005. 2DOI