betaFeedback

WorksEngineering

ST-Dist: Distilling Spatial Transcriptomics Predictions into Interpretable Gene Axes for Distant Metastasis Prediction

Abstract

Distant metastasis at diagnosis determines the treatment of prostate cancer, but predicting it from routine biopsy slides is challenging as the signs of metastasis are not clearly localized and are more subtle than in standard tumor detection and subtyping tasks. Multiple instance learning (MIL) on pathology foundation-model features has become the standard approach when labeled data are abundant, but such models remain hard to interpret. Distant metastasis is further challenging, because confirming the label requires clinical follow-up beyond the routine H&E slides. To this end, we propose ST-Dist, which distills the gene expression predicted by a spatial transcriptomics (ST) model into a linear head mapping each tile's foundation-model features onto named gene axes. The slide's projections along these interpretable gene axes are then used for distant metastasis prediction with a simple logistic regression. On 497 prostate needle biopsies (205 patients), ST-Dist reaches an AUROC of 91.37%, above nine MIL baselines. It remains competitive on the public TCGA-COAD and TCGA-KIRC cohorts, and it has the highest mean AUROC over the three cohorts. The ablations on the prostate biopsies show that the gene axes outperform a linear probe on encoder features, dimension-matched principal components, random gene panels and the ST model's own predicted expression, and the gain over the probe holds on all 13 pathology encoder settings that we tested. ST-Dist thus offers a simple, accurate and interpretable route from H&E to distant metastasis prediction.

Introduction

A patient who already has distant metastasis when prostate cancer is diagnosed (stage M1 rather than M0) needs systemic therapy, not local treatment alone. Knowing whether metastasis is present is therefore one of the most important questions at diagnosis, and one that the biopsy from the primary tumor alone does not yet answer. Unlike tumor detection or subtyping, which have abundant labels and a clearly localized signal, metastasis prediction relies on signs that are not confined to one region of the slide [1, 2]. Additionally, confirming metastatic status requires further imaging and clinical follow-up, which makes labels expensive and cohorts small.

Clinical risk groups and tissue-based genomic classifiers such as Decipher [3], the cell cycle progression score [4] and the Genomic Prostate Score [5] help to assess the aggressiveness of the cancer but cannot establish a clear metastatic status. These classifiers additionally need dedicated molecular assays, which are not always available. In contrast, the routine hematoxylin and eosin (H&E) slide exists for every patient, but whether this slide can predict M0/M1 status at diagnosis is largely unexplored.

Recently, deep learning has shown the prognostic value of biopsy histology by combining slide images with clinical variables to predict long-term distant metastasis in phase III trial cohorts [6]. Standard approaches like multiple instance learning (MIL) on pathology foundation-model features [7, 8, 9] learn attention weights that pick out the most informative tiles, which suits localized signals. With few labels, however, simple pooling without attention is often as good as complex MIL [10, 11]. Further, the explanations that MIL offers, mostly attention heatmaps, are not always accepted as interpretable [12].

Gene expression offers a different language for the same slide. Clinicians already reason about metastatic risk in terms of genes, and with models now capable of predicting expression from H&E, predicted expression can be leveraged for downstream tasks. Paired spatial transcriptomics (ST) resources such as HEST-1k [13] have enabled the development of foundation models such as STPath [14], which predicts tens of thousands of genes for every tile of a WSI. Similarly, virtual proteins predicted from H&E have been shown to improve prognostic prediction in lung cancer [15], further showing that predicted expression is useful for prognosis.

In this work, we propose genes as a concept bottleneck [16]. Unlike a recent concept-based MIL method [12], which predicts its concepts with a vision-language model, we distill the concepts (genes) from an ST foundation model, STPath. A linear head on top of a frozen pathology encoder learns to reproduce the expression that the teacher, STPath, predicts for a small, pre-specified panel of metastasis related genes. A logistic regression trained on the outcome labels then predicts metastasis from the slide-averaged gene values. The head can sit on any pathology encoder, independent of the teacher’s backbone. We call this model ST-Dist. Every input of downstream classifiers using ST-Dist is thus a named gene rather than an unnamed image feature.

On 497 prostate needle biopsies, ST-Dist outperforms nine MIL methods, and we further evaluate it on two public cohorts of The Cancer Genome Atlas (TCGA-COAD and TCGA-KIRC). Our contributions can be summarized as:

  • We introduce ST-Dist, an encoder-agnostic way to read pathology features through named gene axes, distilled from an ST model without outcome labels.

  • We show on three metastasis cohorts that this simple model matches or exceeds nine MIL methods, and that its gain over a linear probe on encoder features holds on 13 different encoder settings on prostate biopsies.

  • We trace the gain to the choice of genes rather than to the reduction in dimension, and validate STPath’s predictions of two of the genes with the largest M1–M0 differences, MSMB and AZGP1, on measured ST.

Method

Problem setting. A slide ss is represented by its NsN_s tissue tiles, with features xi=f(tilei)∈Rdx_i=f(\mathrm{tile}_i)\in\mathbb{R}^d from a frozen pathology encoder ff. The task is to predict the slide label ys∈{0,1}y_s\in\{0,1\} (M0 or M1) from a small labeled cohort. ST-Dist addresses it as shown in Figure 1.

Figure 1. Overview of ST-Dist. Learn Gene Axis Projection: without outcome labels, the gene head (W,b)(W,b) is fit so that its tile values Wx~i+bW\tilde{x}_i+b match STPath’s predicted expression tit_i (ridge loss, Eq. (1)). Downstream Predictor: with the gene head fixed, a logistic regression predicts M0 or M1 from the standardized, slide-averaged gene values zsz_s. Snowflake: frozen. STPath is not used at inference.

Tile encoder. Our primary pathology encoder (ff) is Prov-GigaPath [17], a vision transformer pretrained by self-supervised learning on 1.3 billion tiles (d=1536d=1536). Any other pathology encoder can serve as ff. The teacher, STPath, always receives GigaPath features, so a different encoder changes only the input of the gene head.

Spatial transcriptomics teacher. STPath is a generative foundation model that predicts the expression of 38,984 genes from histology. It is a geometry-aware transformer, pretrained by masked gene-expression prediction on about 1,000 public slides with paired ST profiles from 17 organs. Each tile is one token, which combines the tile’s GigaPath feature, an expression input, an organ token and a sequencing-technology token; tile positions enter through the geometry-aware attention. At inference, all expression inputs are masked, so the prediction rests on histology alone. We use the organ token of each cohort (Prostate, Kidney or Bowel) and no technology token. For tile ii, STPath produces a 512-dimensional latent uiu_i (after layer normalization). Each gene gg is a fixed linear readout of it, tig=agTui+ag0t_{ig}=a_g^\mathsf{T}u_i+a_{g0}. For a panel of GG genes, we collect these values in ti∈RGt_i\in\mathbb{R}^G.

Choosing the gene panel. The gene panel is a predefined set of genes to predict from the teacher. It is fixed before any outcome is seen and is derived from the union of published gene sets for the endpoint: signatures of metastasis and aggressiveness, lineage genes and cancer drivers. Genes outside the teacher’s vocabulary are removed. Our selection of panels is described in § 3.

Distilling the gene head. The student is a linear gene head, W∈RG×dW\in\mathbb{R}^{G\times d} and b∈RGb\in\mathbb{R}^G, that maps a tile feature to the teacher’s values. Let x~i\tilde{x}_i denote the standardized feature of tile ii. Row wgw_g of WW is the axis of gene gg in feature space, and the tile’s value for gene gg is its coordinate along this axis, wgTx~i+bgw_g^\mathsf{T}\tilde{x}_i+b_g. Over the set S\mathcal{S} of training slides, the head minimizes the slide-weighted ridge objective

1∣S∣∑s∈S1Ns∑i∈s∥ti−Wx~i−b∥22+λ∥W∥F2,(1)\frac{1}{|\mathcal{S}|}\sum_{s\in\mathcal{S}}\frac{1}{N_s}\sum_{i\in s}\lVert t_i-W\tilde{x}_i-b\rVert_2^2+\lambda\lVert W\rVert_F^2, \tag*{(1)}

so that every WSI has the same total weight, whatever its size. The regularization parameter λ\lambda is chosen from ten values between 10−410^{-4} and 10210^2, by the mean per-gene R2R^2 between head and teacher on the validation slides. No outcome label enters this step, and the head is refit in every fold.

Slide gene values. For slide ss with mean standardized tile feature xˉs=1Ns∑i∈sx~i\bar{x}_s=\frac{1}{N_s}\sum_{i\in s}\tilde{x}_i, the slide’s gene values e^s∈RG\hat{e}_s\in\mathbb{R}^G are the average of the tile-level gene values,

e^s=Wxˉs+b=1Ns∑i∈s(Wx~i+b).(2)\hat{e}_s=W\bar{x}_s+b=\frac{1}{N_s}\sum_{i\in s}(W\tilde{x}_i+b). \tag*{(2)}

Each component of e^s\hat{e}_s is standardized with the mean and standard deviation over the training slides, which gives zs∈RGz_s\in\mathbb{R}^G.

Metastasis classifier. The classifier is an ℓ2\ell_2-regularized logistic regression on zsz_s, with inverse regularization strength C=1C=1. This value is fixed in advance and never tuned, so the validation slides serve only to choose the ridge penalty. At inference, a slide needs only the encoder, the gene head and the classifier. The teacher is used only to train the head.

Why concept bottleneck. Because zsz_s is an affine function of xˉs\bar{x}_s, ST-Dist is a linear probe restricted to the subspace spanned by the gene axes, of dimension at most GG. The genes thus act as a prior on which visual features may matter.

Experiments

Cohorts. Our main cohort consists of 497 pretreatment prostate needle biopsies from 205 patients (98 with M1), labeled M0 or M1 for distant metastasis from clinical reports and follow-up. The main cohort data come from a large US hospital network. To test our approach on other cancers, we use two TCGA resection cohorts [18] with the same endpoint: kidney clear-cell carcinoma (TCGA-KIRC, 509 slides from 505 patients, 76 with M1) and colon adenocarcinoma (TCGA-COAD, 387 slides from 380 patients, 60 with M1). The TCGA labels use the pathologic M stage (the clinical one when it is missing). Patients with MX or no M stage are excluded. Each cohort is split into three patient-level folds, each with training, validation and test sets.

Gene panels. Each cohort has its own panel. The prostate panel has 117 genes. Of these, 100 come from published signatures of metastasis and aggressiveness [19, 3, 4, 5]. Ten are prostate lineage genes, such as NKX3-1 and MSMB, and seven are genes often altered in metastases. The kidney (98 genes) and colon (113 genes) panels come from five and seven published sources. The panel files, with the source of every gene, will be released with the code.

A second vocabulary. To test whether the result depends on our hand-built panel, we also use a vocabulary from a single standard source, fixed by name before its result was computed: the 1,236 genes of 10 Molecular Signatures Database (MSigDB) Hallmark pathways linked to metastasis [20]. These are epithelial–mesenchymal transition, TGF-β\beta signaling, apical junction, hypoxia, angiogenesis, inflammatory response, IL6/JAK/STAT3 signaling, TNFα\alpha signaling via NF-κ\kappaB, E2F targets and the G2M checkpoint.

Encoders and baselines. Our main experiments use GigaPath features of 62.5 μ\mum tiles, and robustness is tested on 13 settings of 9 pathology encoders (see § 4). The other settings use 125 μ\mum tiles. For a 125 μ\mum tile, the teacher’s latent is averaged over the four 62.5 μ\mum tiles that it contains, and the gene values follow by the same linear readout. We compare ST-Dist with nine MIL methods. Seven use attention: ABMIL [7], CLAM [21], DSMIL [22], TransMIL [23], DTFD-MIL [24], ILRA [25] and RRT-MIL [8]. Two use simple pooling techniques: SimMPL [10] combines mean-pooling with MLP and Maxsoft [11] takes the highest tile score. All MIL methods share one training recipe: AdamW with a learning rate of 2×10−42\times10^{-4} and a weight decay of 10−410^{-4}, dropout of 0.25 and 20 epochs, keeping the epoch with the highest validation area under the ROC curve (AUROC), with five random seeds per fold. Our main ablation removes the gene head: a mean-pooling linear probe applies the same logistic regression to the slide mean of all encoder features. Random gene panels test whether any genes that the teacher predicts would work as well as the chosen ones. For the curated panel and the Hallmark vocabulary alike, each of 200 random counterparts replaces every gene by one of its 100 nearest genes outside the vocabulary in predicted variance, so that the random genes vary as much across tiles as the original predicted genes.

Evaluation. We report the slide-level AUROC as the mean over the three test folds, with its standard deviation (SD) over runs (three folds for logistic models, 15 fold-and-seed runs for MIL). On the main cohort we also report average precision (AP) and balanced accuracy at a 0.5 threshold. All metrics are given in percent. Differences between models are tested with a paired bootstrap over patients (2,000 resamples), in which predictions are averaged over seeds and over each patient’s slides, and pooled over the three test folds. We call a difference significant when its 95% bootstrap confidence interval excludes zero.

RESULTS

Accuracy, ablations and robustness

Performance on the prostate biopsies. Table 1 compares ST-Dist with nine MIL methods on the main cohort. ST-Dist has the highest AUROC (91.37%), average precision (92.29%) and balanced accuracy (82.12%), and its AUROC is 2.2 points above the best MIL method (ABMIL, 89.13%). Among the MIL methods, simple mean pooling (SiMLP, 88.54%) comes within 0.6 points of the best attention-based method, in line with earlier findings for small cohorts [10, 11].

MethodAUROCAPBalanced acc.
ST-Dist (ours)91.37±2.1992.29±1.6082.12±2.15
ILRA87.63±2.8988.56±3.2978.98±4.40
ABMIL<u>89.13±2.28</u><u>90.16±2.90</u>80.28±4.44
RRT-MIL<u>89.08±2.37</u>89.88±2.7876.86±5.82
CLAM88.34±2.4789.34±2.8379.53±4.30
DSMIL86.98±3.3187.99±3.5977.82±4.83
TransMIL83.94±3.8484.55±4.3871.35±7.47
DTFD-MIL72.63±8.5972.39±8.9466.52±8.10
SiMLP88.54±1.5789.78±2.40<u>80.56±2.42</u>
Maxsoft84.23±2.2585.07±2.8474.36±2.95

Table 1. Prostate biopsies (the main cohort), in percent: mean ± SD over runs. Bold: best; <u>underline:</u> second.

Results on the TCGA cohorts. We refit the same recipe on each TCGA cohort, with the cohort’s own gene panel (Table 2). These resection cohorts are harder than the biopsies: M1 patients are rarer (15–16%), and every method scores lower. On TCGA-COAD, ST-Dist (67.71%) outperforms all nine MIL methods, by 2.9 points over the best one (Maxsoft), significantly so against RRT-MIL, TransMIL and DTFD-MIL. On TCGA-KIRC, it ranks second (81.19%), 0.2 points behind ILRA. Without the gene head, the linear probe reaches 78.05% on TCGA-KIRC and 69.16% on TCGA-COAD. Overall, ST-Dist has the highest mean AUROC (80.09%, against 78.68% for the probe and 77.36% for the best MIL method, CLAM).

MethodTCGA-KIRCTCGA-COADMean (3 cohorts)
ST-Dist (ours)<u>81.19±0.52</u>67.71±1.5280.09
ILRA81.38±4.8062.63±5.0477.21
ABMIL79.76±5.0461.35±6.1876.75
RRT-MIL80.62±4.5661.39±4.6277.03
CLAM80.10±4.7563.64±4.99<u>77.36</u>
DSMIL75.79±4.9661.85±7.3074.87
TransMIL72.06±6.6758.81±6.9771.60
DTFD-MIL65.66±7.5051.66±7.0563.32
SiMLP78.71±2.9763.62±6.0076.96
Maxsoft78.67±2.97<u>64.77±5.59</u>75.89

Table 2. TCGA cohorts (resections; 15–16% M1 patients): AUROC in percent, mean ± SD as in Table 1. The mean is over the three cohorts, with the prostate biopsies from Table 1. Bold: best; <u>underline:</u> second.

Source of the gain. Table 3 tests three explanations of the gain. (i) Compression: if ST-Dist only compressed the 1,536 features

ModelAUROC
ST-Dist91.37±2.19
Nonlinear gene head (MLP)91.68±2.78
Attention-weighted ST-Dist<u>91.53±1.45</u>
Hallmark vocabulary90.81±0.35
STPath expression89.60±2.73
Random gene panels89.49±1.02
ST-Dist on the full latent (512)89.03±0.35
ST-Dist + raw features88.96±1.57
Linear probe (no gene head)88.82±1.80
PCA-11786.78±0.81

Table 3. Ablations on the prostate biopsies (AUROC in percent, mean ± SD as in Table 1; for random gene panels, SD over the 200 panels). Bold: best; <u>underline:</u> second.

into 117, any 117-dimensional projection would work as well. It does not. The 117 principal components fall below the linear probe (86.78% against 88.82%), and random gene panels of the same size reach only 89.49%, with the curated panel above 97% of them. (ii) The teacher’s representation: distilling the full 512-dimensional STPath latent instead of the genes gives almost no gain over the probe (89.03%), so the teacher helps mainly through the chosen genes. (iii) Restriction: adding the raw features back to the gene values (88.96%) returns the AUROC to the level of the probe, because the classifier can again use every visual feature direction. The benefit therefore comes from the choice of genes, not from the reduction in dimension. Two further rows test design choices. Learning tile weights with attention on top of ST-Dist’s classifier gives a similar AUROC (91.53%), and the learned weights stay close to equal (on average, they spread over 91% of the tiles), so the plain slide average suffices. A nonlinear gene head (an MLP with two hidden layers of 1,024 units) reproduces the teacher more closely than a linear layer trained in the same label-free way (validation R2R^2 0.729 against 0.691), but its AUROC on the prostate biopsies is similar, so we keep the simpler linear head.

Robustness to the choice of vocabulary. The gain does not depend on our hand-built panel. Replacing it with the genes of the 10 metastasis-related Hallmark pathways (1,236 genes, of which only 41 are in the 117-panel) yields an AUROC of 90.81 ± 0.35%, above the linear probe (88.82%) and above 97% of its 200 variance-matched random counterparts.

The student outperforms its teacher. ST-Dist retains only the part of STPath’s prediction that a tile’s own features explain linearly. In †^{\dagger} Two in-house runs, self-supervised on about 60% of TCGA [18] slides from all organs. Both use the official MoCo v3 augmentations; SRA (stain reconstruction augmentation [33]) adds one that randomizes the amounts of the H and E stains separately.

all three cohorts, this part predicts metastasis better than STPath’s own prediction of the same genes (STPath expression): 91.37% against 89.60% on the prostate biopsies, 81.19% against 78.90% on TCGA-KIRC and 67.71% against 62.69% on TCGA-COAD, where the difference is significant.

Robustness across encoders. ST-Dist outperforms the linear probe on all 13 encoder settings, by 0.10 to 4.69 points (Table 4). Encoders that reproduce the teacher more closely tend to give better ST-Dist models (Spearman 0.72 between the validation R2R^2 and the ST-Dist AUROC). The curated panel ranks at or above the 90th percentile of variance-matched random gene panels on 9 of the 13 settings.

EncoderR2R^2Linear probeST-Dist
MoCo v3 ResNet-50 (SRA)†^{\dagger}0.50882.20±\pm2.2886.89±\pm4.73
MoCo v3 ResNet-50†^{\dagger}0.53783.73±\pm2.3388.38±\pm3.50
Virchow2 (CLS + mean)0.63686.95±\pm3.2091.48±\pm2.24
Hibou-B0.59487.82±\pm2.1389.85±\pm2.45
Virchow2 (CLS)0.62288.67±\pm2.3391.05±\pm2.32
GigaPath (62.5 μ\mum, main)0.69488.82±\pm1.8091.37±\pm2.19
GigaPath (125 μ\mum)0.63389.66±\pm0.3891.06±\pm2.62
Phikon0.58289.68±\pm1.8090.48±\pm1.72
Virchow (CLS + mean)0.63689.80±\pm0.8991.23±\pm1.99
Virchow (CLS)0.61189.85±\pm2.0390.76±\pm2.35
UNI0.59989.97±\pm1.3391.95±\pm2.03
Phikon-v20.60390.58±\pm1.7890.68±\pm1.43
UNI2-h0.61491.62±\pm0.7192.06±\pm1.69

Table 4. ST-Dist on 13 settings of 9 pathology encoders (prostate biopsies, AUROC in percent, mean ±\pm SD over 3 folds), sorted by the linear probe. R2R^2: mean validation R2R^2 of the gene head against the teacher. CLS + mean: the class token concatenated with the mean patch token. Encoders: GigaPath [17], UNI and UNI2-h [26], Virchow [27], Virchow2 [28], Phikon [29], Phikon-v2 [30], Hibou-B [31] and MoCo v3 [32].

Which genes separate M1 from M0

Genes with the largest differences. ST-Dist outperforms its teacher (Table 3). To check that its gene values still carry the teacher’s biology, Figure 2 compares the ten genes with the largest M1–M0 differences in ST-Dist with the same genes in STPath, measured by Hedges’ gg (the difference in patient means, divided by the pooled SD). ST-Dist’s differences closely follow STPath’s (Pearson rr = 0.99 over all 117 genes) and are larger for 8 of the 10 genes. The student thus keeps the teacher’s gene-level signal and separates M1 from M0 slightly more strongly.

Figure 2. The largest M1–M0 differences in ST-Dist and STPath (Hedges’ gg; 98 M1, 107 M0 patients), all significant in both (Mann–Whitney test, p<10−3p < 10^{-3}, corrected for 117 genes). * marks STPath’s prediction validated on measured ST.

Grounding in measured expression. On eight held-out prostate Visium slides from HEST-Bench [13], which STPath did not see during training, STPath predicts 12 of the 115 panel genes that these slides measure with a mean correlation of at least 0.3 (0.32 to 0.52) against measured expression, computed per gene within each slide. These are mostly luminal and androgen-regulated genes (MSMB, AZGP1, SLC45A3, TMPRSS2, KLK2, KLK3, HOXB13, NKX3-1, SPDEF, ANO7, TPM2 and FOXA1). They include two of the genes with the largest M1–M0 differences in both models, MSMB and AZGP1 (Figure 2), whose loss has been associated with aggressive prostate cancer [34, 35]. Both genes are lower in M1 patients, which agrees with these reports.

Recovery of known biology. Used alone in place of the panel, the E2F targets pathway (one of the 10 Hallmark pathways) reaches an AUROC of 91.43 ±\pm 1.07%. This is at the 96th percentile of 200 variance-matched random gene sets of the same size (nominal pp = 0.05). E2F targets form the proliferation program that the cell cycle progression score [4] also measures, a score already used clinically to assess the aggressiveness of prostate cancer.

Discussion and conclusion

ST-Dist shows that distant metastasis, a label-scarce target, is well served by a simple model that reads pathology features through a small set of named genes rather than unnamed image features. The genes do more than make the model interpretable, they act as a prior that selects which visual directions are worth learning from few labels. A task-relevant vocabulary, whether our curated panel or the metastasis-related Hallmark pathways, keeps most of this benefit. In summary, distilling an ST foundation model into named gene axes turns any pathology encoder into a simple, interpretable and competitive predictor of distant metastasis from routine H&E slides.

Compliance with ethical standards

As an analysis of de-identified images, the study is not human subjects research and was exempt from IRB review. The TCGA [18] and HEST-1k [13] data are publicly available, and their use did not require separate ethical approval.

Acknowledgments

This work was supported by U.S. Department of Defense (DoD) HT94252410186. We declare no conflicts of interest. We used Claude (Anthropic) to draft code and edit language. ChatGPT Pro(Open AI) was used to curate gene panels.

References

  1. [1]Gustavo Ayala et al., “Reactive stroma as a predictor of biochemical-free recurrence in prostate cancer,” Clin. Cancer Res., vol. 9, no. 13, pp. 4792–4801, 2003.pubmed.ncbi.nlm.nih.gov/14581350
  2. [2]Larisa Nonn, Vijayalakshmi Ananthanarayanan, and Peter H. Gann, “Evidence for field concretization of the prostate,” Prostate, vol. 69, no. 13, pp. 1470–1479, 2009.DOI
  3. [3]Nicholas Erho et al., “Discovery and validation of a prostate cancer genomic classifier that predicts early metastasis following radical prostatectomy,” PLoS One, vol. 8, no. 6, pp. e66855, June 2013.DOI
  4. [4]Jack Cuzick et al., “Prognostic value of an RNA expression signature derived from cell cycle proliferation genes in patients with prostate cancer: a retrospective study,” Lancet Oncol., vol. 12, no. 3, pp. 245–255, Mar. 2011.DOI
  5. [5]Eric A Klein et al., “A 17-gene assay to predict prostate cancer aggressiveness in the context of Gleason grade heterogeneity, tumor multifocality, and biopsy undersampling,” Eur. Urol., vol. 66, no. 3, pp. 550–560, May 2014.DOI
  6. [6]Andre Esteva et al., “Prostate cancer therapy personalization via multi-modal deep learning on randomized phase III clinical trials,” npj Digit. Med., vol. 5, no. 1, pp. 71, June 2022.DOI
  7. [7]Maximilian Ilse, Jakub Tomczak, and Max Welling, “Attention-based deep multiple instance learning,” in Proc. ICML, 2018, pp. 2127–2136.
  8. [8]Wenhao Tang, Fengtao Zhou, Sheng Huang, Xiang Zhu, Yi Zhang, and Bo Liu, “Feature re-embedding: Towards foundation model-level performance in computational pathology,” in Proc. CVPR, 2024, pp. 11343–11352.arxiv.org/abs/2402.17228
  9. [9]Sandesh Pokhrel, Hamid Manoochehri, Beatrice S. Knudsen, and Tolga Tasdizen, “Predicting metastatic risk from primary cancer tissue architecture via distance-aware spatial modeling,” arXiv:2606.28676, 2026.arxiv.org/abs/2606.28676
  10. [10]Jiawen Li et al., “Can we simplify slide-level fine-tuning of pathology foundation models?,” arXiv:2502.20823, 2025.arxiv.org/abs/2502.20823
  11. [11]Hossein Jafarinia et al., “Navigating the MIL trade-off: Flexible pooling for whole slide image classification,” in Proc. NeurIPS, 2025, pp. 96786–96921.DOI
  12. [12]Susu Sun et al., “Label-free concept based multiple instance learning for gigapixel histopathology,” arXiv:2501.02922, 2025.arxiv.org/abs/2501.02922
  13. [13]Guillaume Jaume et al., “HEST-1k: A dataset for spatial transcriptomics and histology image analysis,” in Proc. NeurIPS, 2024, pp. 53798–53833.arxiv.org/abs/2406.16192
  14. [14]Tinglin Huang, Tianyu Liu, Mehrtash Babadi, Rex Ying, and Wengong Jin, “STPath: a generative foundation model for integrating spatial transcriptomics and whole-slide images,” npj Digit. Med., vol. 8, no. 1, pp. 659, Nov. 2025.DOI
  15. [15]Zhe Li et al., “AI-enabled virtual spatial proteomics from histopathology for interpretable biomarker discovery in lung cancer,” Nat. Med., vol. 32, no. 1, pp. 231–244, 2026.DOI
  16. [16]Pang Wei Koh et al., “Concept bottleneck models,” in Proc. ICML, 2020, pp. 5338–5348.arxiv.org/abs/2007.04612
  17. [17]Hanwen Xu et al., “A whole-slide foundation model for digital pathology from real-world data,” Nature, vol. 630, no. 8015, pp. 181–188, 2024.DOI
  18. [18]John N Weinstein et al., “The Cancer Genome Atlas Pan-Cancer analysis project,” Nat. Genet., vol. 45, no. 10, pp. 1113–1120, 2013.
  19. [19]Rohina Rubicz et al., “Gene expression panel predicts metastatic-lethal prostate cancer outcomes in men diagnosed with clinically localized prostate cancer,” Mol. Oncol., vol. 11, no. 2, pp. 140–150, 2017.
  20. [20]Arthur Liberzon, Chet Birger, Helga Thorvaldsdóttir, Mahmoud Ghandi, Jill P. Mesirov, and Pablo Tamayo, “The Molecular Signatures Database Hallmark gene set collection,” Cell Syst., vol. 1, no. 6, pp. 417–425, 2015.pubmed.ncbi.nlm.nih.gov/26771021
  21. [21]Ming Y. Lu, Drew F. K. Williamson, Tiffany Y. Chen, Richard J. Chen, Matteo Barbieri, and Faisal Mahmood, “Data-efficient and weakly supervised computational pathology on whole-slide images,” Nat. Biomed. Eng., vol. 5, no. 6, pp. 555–570, 2021.DOI
  22. [22]Bin Li, Yin Li, and Kevin W. Eliceiri, “Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning,” in Proc. CVPR, 2021, pp. 14313–14323.DOI
  23. [23]Zhuchen Shao et al., “TransMIL: Transformer based correlated multiple instance learning for whole slide image classification,” in Proc. NeurIPS, 2021, pp. 2136–2147.arxiv.org/abs/2106.00908
  24. [24]Hongrun Zhang et al., “DTFD-MIL: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification,” in Proc. CVPR, 2022, pp. 18780–18790.arxiv.org/abs/2203.12081
  25. [25]Jinxi Xiang and Jun Zhang, “Exploring low-rank property in multiple instance learning for whole slide image classification,” in Proc. ICLR, 2023.
  26. [26]Richard J. Chen et al., “Towards a general-purpose foundation model for computational pathology,” Nat. Med., vol. 30, no. 3, pp. 850–862, 2024.DOI
  27. [27]Eugene Vorontsov et al., “A foundation model for clinical-grade computational pathology and rare cancers detection,” Nat. Med., vol. 30, no. 10, pp. 2924–2935, 2024.DOI
  28. [28]Eric Zimmermann et al., “Virchow2: Scaling self-supervised mixed magnification models in pathology,” arXiv:2408.00738, 2024.arxiv.org/abs/2408.00738
  29. [29]Alexandre Filiot et al., “Scaling self-supervised learning for histopathology with masked image modeling,” medRxiv, 2023.DOI
  30. [30]Alexandre Filiot, Paul Jacob, Alice Mac Kain, and Charlie Saillard, “Phikon-v2, a large and public feature extractor for biomarker prediction,” arXiv:2409.09173, 2024.arxiv.org/abs/2409.09173
  31. [31]Dmitry Nechaev, Alexey Pchelnikov, and Ekaterina Ivanova, “Hibou: A family of foundational vision transformers for pathology,” arXiv:2406.05074, 2024.arxiv.org/abs/2406.05074
  32. [32]Xinlei Chen, Saining Xie, and Kaiming He, “An empirical study of training self-supervised vision transformers,” in Proc. ICCV, 2021, pp. 9620–9629.arxiv.org/abs/2104.02057
  33. [33]Hamid Manoochehri, Bodong Zhang, Beatrice S. Knudsen, and Tolga Tasdizen, “SRA: A novel method to improve feature embedding in self-supervised learning for histopathological images,” arXiv:2410.17514, 2024.arxiv.org/abs/2410.17514
  34. [34]James D. Brooks et al., “Loss of expression of AZGP1 is associated with worse clinical outcomes in a multi-institutional radical prostatectomy cohort,” Prostate, vol. 76, no. 15, pp. 1409–1419, 2016.DOI
  35. [35]Anna Dahlman et al., “Evaluation of the prognostic significance of MSMB and CRISP3 in prostate cancer using automated image analysis,” Mod. Pathol., vol. 24, no. 5, pp. 708–719, 2011.DOI

Paper details

vlnc:W_29p7v6qu
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Claude, ChatGPT Pro

View statement
Looking for related papers…
Contents