acceptodds
Under review as a conference paper at ICLR 2027

Density-Based Online Autocuration for Pathology Image Model Pre-Training

Abstract

Tile streams for pathology image model pre-training are curated at the slide level, or offline at the tile level with an oracle encoder. Morphologies are unevenly distributed within a slide, so random tile sampling carries each slide's imbalance into the training stream and common morphologies overshadow rare ones. This work curates the stream online, during DINOv2 pretraining, from a density estimate the model produces itself. A density cache keeps a fixed-size memory of stored tile signatures and counts how often new tiles land on each; the counts give the typicality of an incoming tile, and tiles enter the batch with probability inversely proportional to it. One hyperparameter, the tilt, sets how far the batch leans toward low-typicality tiles. ViT-B and ViT-L encoders trained with DINOv2 on 287M tiles were evaluated on twelve biomarker mutation-prediction tasks from H&E. Fold-paired against the uncurated baseline at the same checkpoint, every curated run improved mutation prediction on average, by to percentage points of AUROC over the twelve tasks and by to on THCA/NRAS at ViT-B, and the two thyroid tasks improved in every run at both scales. At ViT-L the improvement is on the tasks with the most consistent folds, and two of the three runs are significant when each task is weighted by the precision of its measurement, with a tilt of the best of the three. At ViT-B it is on the tasks with few positive slides per fold and is resolved in one run of seven. Across the eight k runs whose caches did not starve, the gain rises with the spread of hit-rates over the cache, its bandwidth (Spearman ). A k-step pair at ViT-B, four times the schedule, is at parity on aggregate ( percentage points, 95% interval ), and the thyroid and colorectal tasks improve again under this third independent baseline. Mapped back onto a slide, the stored signatures separate morphologically coherent regions at tile resolution, so the cache also yields an unsupervised tissue segmentation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
autocuration, self supervised learning, dinov2, histopathology, cancer

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_43bUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Density-Based Online Autocuration, a method for curating the training stream during pathology image model pre-training. Within DINOv2 self-supervised pre-training, the method constructs tile signatures from representations produced by the evolving model and estimates local density using a fixed-capacity density cache. The estimated density is used to adjust the admission probability of candidate tiles, reducing the sampling frequency of tiles associated with common morphologies without changing the original training loss.

The authors conduct experiments with ViT-B and ViT-L encoders on a pathology image corpus containing approximately 287 million tiles, evaluating the frozen encoders on twelve biomarker mutation prediction tasks using H&E images. Under the 125k-step training schedule, the authors report aggregate mean AUROC improvements of approximately 0.3–1.0 percentage points across the twelve tasks for the curated configurations relative to their corresponding baselines. For ViT-L, the improvements are concentrated in a subset of tasks with more consistent evaluation results, whereas for ViT-B, the gains are more pronounced on tasks with fewer positive slides per evaluation fold. In the extended 500k-step ViT-B experiment, the curated model achieves near parity with its corresponding baseline in aggregate, with a reported mean difference of +0.3 percentage points and a 95% interval of [−0.4, +0.9].

The authors also demonstrate the potential of using the same density cache for unsupervised tissue segmentation. However, the segmentation results are not quantitatively evaluated against manually annotated reference masks. Overall, the work introduces a method for integrating online density estimation into the data stream of self-supervised pre-training and investigates its effects through downstream mutation prediction and qualitative tissue segmentation experiments.

Critical strengths
  1. A simple and interesting idea for online data curation.

I like the idea of using the model's own evolving representations to guide data curation during pre-training, rather than relying on a separate encoder to curate the data offline. The density cache provides a practical way to estimate how common a tile's representation is and adjust its sampling probability accordingly, without changing the original DINOv2 training objective. This is a natural and interesting direction for addressing morphological imbalance in pathology data.

  1. A reasonably broad experimental study with useful observations.

The authors evaluate the method with both ViT-B and ViT-L encoders and explore several curation configurations. I also appreciate that the paper reports less favorable results, including the near-parity between the curated model and its baseline under the extended 500k-step schedule, rather than focusing only on positive outcomes. These experiments provide useful insights into when the method may help and where its benefits remain limited. That said, more independent pre-training runs would help establish how consistently these observations hold.

Critical weaknesses
  1. Downstream evaluation. The authors report the best AUROC from six evaluations on the held-out set without a separate validation set. The five shuffle splits also have overlapping held-out sets, and slides from the same patient can appear in both the downstream head-training and held-out evaluation partitions. The reported patient overlap is 8% on average and 21% for BLCA/FGFR3. Given the relatively small reported gains, patient-disjoint evaluation and a model selection protocol that does not rely on the held-out set would strengthen the results.

  2. Computational cost and baselines. The proposed method requires additional forward passes to score candidate tiles. The authors report that step time increases by 1.65–2.40 times for some ViT-B configurations. However, the main comparisons are made at matched training steps rather than matched computational budgets. A comparison with standard pre-training and a simpler feature-space resampling baseline under matched compute would help clarify whether the proposed method offers sufficient benefits to justify its additional cost.

  3. Reproducibility and training dynamics. Most curated runs resume from their corresponding baseline checkpoints, so the comparisons share the same pre-training history. While this helps isolate the effect of curation, it remains unclear whether the gains persist across independent training runs. The paper mainly reports downstream results at the final checkpoints, while the training dynamics analysis focuses on the density cache rather than downstream performance. Independent pre-training runs with different random seeds and downstream evaluations at intermediate checkpoints would help clarify the consistency of the reported gains.

  4. Comparison with existing pathology foundation models. The supplementary material compares the proposed models with UNI, Virchow, Virchow2, Prov-GigaPath, and H-Optimus-0. However, substantial performance gaps remain on several tasks. For example, on COAD/BRAF, ViT-L (run 11) achieves an AUROC of 69.4, compared with 80.9 for Virchow2 and 80.5 for H-Optimus-0. Although the primary goal is to study online data curation rather than to develop a new state-of-the-art model, the authors should clarify the practical significance of these results and whether the proposed strategy can also benefit stronger pre-trained encoders.

Questions for authors

I’m still wondering how useful this sampling strategy is in practice. Since it requires extra forward passes to select tiles, could the authors clarify whether it actually provides a meaningful benefit over simpler sampling strategies or standard pre-training when the compute budget is the same?

4 comments
AuthorsOpenReview ↗

Response to 43bU Weakness 1

Weakness 1. Downstream evaluation. The authors report the best AUROC from six evaluations on the held-out set without a separate validation set. The five shuffle splits also have overlapping held-out sets, and slides from the same patient can appear in both the downstream head-training and held-out evaluation partitions. The reported patient overlap is 8% on average and 21% for BLCA/FGFR3. Given the relatively small reported gains, patient-disjoint evaluation and a model selection protocol that does not rely on the held-out set would strengthen the results.

We thank the reviewer for this careful reading of our evaluation. We ran both of the analyses that the reviewer proposes, and report that the improvements fortunately remain under each (Table 1).

For the patient-disjoint evaluation, we removed from each held-out fold every slide whose patient also has a slide in the training part of that split (the 8% and 21% that the reviewer cites), and we scored the same trained heads on the slides that were left. Run 11, our main ViT-L run, then improves on its baseline by +1.00 pp, with a 95% interval of +0.40 to +1.57 pp, compared with +0.91 pp on all slides. The improvement becomes larger in ten of the eleven comparisons.

For a model selection that does not rely on the held-out set, we fixed the last epoch (120) for every encoder, task and split, so that nothing is selected at all. Run 11 then improves by +1.23 pp, with a 95% interval of +0.12 to +2.51 pp, and by +1.21 pp on the patient-disjoint folds.

For the overlapping held-out sets (any two of them share 27% of their slides), we applied the correction of Nadeau and Bengio (2003), which widens the standard error of the precision-weighted mean by a factor of 1.87 (1.81 on the patient-disjoint folds). The improvement of run 11 remains significant in both cases (p=0.008p = 0.008 and p=0.004p = 0.004). A bootstrap over patients, which keeps all slides of a patient together across the five splits, gives +0.91 pp with a 95% interval of +0.06 to +1.79 pp.

Table 1. Curated run minus its baseline, in pp of AUROC, as the mean of the fold-paired differences over the 60 task-folds. Runs 2 to 8 are ViT-B, runs 10 to 12 are ViT-L, and run 14 is the ViT-B pair at 500k steps. "Best epoch" is the protocol of the paper, and "epoch 120" fixes the last epoch for every encoder, task and split. The first column of runs 10 and 12 is recomputed from the stored per-slide scores (+0.26 and +0.42 in the paper).

RunAll slides, best epochPatient-disjoint, best epochAll slides, epoch 120Patient-disjoint, epoch 120
2+0.80+1.08+1.29+1.59
3+0.75+0.84+0.71+0.72
4+0.74+0.82+1.06+1.03
5+0.62+0.66+0.66+0.63
6+0.87+1.11+1.16+1.42
7+0.82+1.00+0.43+0.42
8+1.04+1.21+1.30+1.26
10+0.31+0.41−0.02−0.13
11+0.91+1.00+1.23+1.21
12+0.56+0.59+0.82+0.79
14+0.28−0.03−0.18−0.56

In conclusion, we ran both of the analyses that you proposed, and the improvements remain under each (+1.00 pp for run 11 on patient-disjoint folds and +1.23 pp at a fixed epoch). The improvement of run 11 also holds with the overlap of the held-out sets taken into account (p=0.008p = 0.008). We should add Table 1 to the appendix for our final revision, which we aim to do on the final day of discussions. We hope you will take this requested evidence into consideration.

References. Nadeau C., Bengio Y. Inference for the generalization error. Machine Learning 52:239–281, 2003.

AuthorsOpenReview ↗

Response to 43bU Weakness 2 and Question 1

Weakness 2. Computational cost and baselines. The proposed method requires additional forward passes to score candidate tiles. The authors report that step time increases by 1.65–2.40 times for some ViT-B configurations. However, the main comparisons are made at matched training steps rather than matched computational budgets. A comparison with standard pre-training and a simpler feature-space resampling baseline under matched compute would help clarify whether the proposed method offers sufficient benefits to justify its additional cost.

Question 1. I’m still wondering how useful this sampling strategy is in practice. Since it requires extra forward passes to select tiles, could the authors clarify whether it actually provides a meaningful benefit over simpler sampling strategies or standard pre-training when the compute budget is the same?

We thank the reviewer for this question, which goes to the practical value of the method. We took the suggestion and trained a new pair of ViT-L runs for it. At matched compute, the curated run is level with standard pre-training. Its macro-mean AUROC over the twelve tasks is 75.5 at 200k steps, against 75.5 for the baseline at 240k steps and 75.2 for the baseline at 260k steps, where the curated run is +0.36 pp above and wins 36 of the 60 task-folds (Table 1).

The paper matches training steps because our claim is about which tiles the model learns from. A scout pass, the forward pass that scores a candidate tile, carries no gradient, so at matched steps the two runs differ only in the tiles of each batch. Selection methods and curated datasets are commonly tested this way (Mindermann et al., 2022, and Oquab et al., 2024). As for the cost, the step-time ratios of 1.65 to 2.40 that the reviewer cites include CPU time in the data loader. Counted as GPU work alone, the scout passes add 22% to 29% to a ViT-B run. At ViT-L, where the GPU sets the step time, they added 16% to run 11, our main run.

For the match itself, our cluster mixes SXM and PCIe variants of the H100 node, and the same run is slower on the PCIe nodes, so a match by measured GPU hours is unreliable in our work. We matched on the step times of the SXM nodes, where a curated step takes 1.27 times as long as a baseline step, and this sets the curated run at 200k steps equal to the baseline at 240k steps. A hardware-independent count of the scout passes gives 254k steps, so we also report the baseline at 260k steps.

A simpler feature-space resampling baseline would not avoid this cost, since it also needs a forward pass over every tile that it chooses from. Done offline, that is an oracle encoder and one pass over all 286.6M tiles of our corpus before the first training step, and again whenever slides are added. Our curator scores each tile in the step that draws it, with no oracle encoder and no stored embeddings, which is where it is most useful in practice. Training this baseline would take a further ten-day ViT-L run, so we give its cost here.

Table 1. The new ViT-L pair at matched compute. Both runs use a constant schedule, as in DINOv3 (Siméoni et al., 2025), so that every checkpoint is a usable end point. Each difference is the curated run at 200k steps minus its baseline, in pp of AUROC, as the mean of the fold-paired differences over the 60 task-folds, with a 95% interval from a bootstrap over tasks.

Baseline atMacro-mean AUROC, curated against baselineDifference, all slidesFolds won of 60Difference, patient-disjoint foldsDifference, fixed epoch (120)
240k steps75.5 against 75.5+0.01 (−0.83 to +0.87)31−0.19 (−1.02 to +0.77)+0.24 (−0.53 to +1.10)
260k steps75.5 against 75.2+0.36 (−0.41 to +1.21)36+0.25 (−0.72 to +1.30)+0.82 (−0.29 to +1.93)

In conclusion, we trained the comparison at matched compute that you asked for, and the curated run is level with standard pre-training (75.5 against 75.5). A feature-space sampler would need the same forward passes as our curator, or an oracle encoder. We will add Table 1, with its per-task results, to the appendix. If this answers your question, we would be grateful if you would revisit your score.

References. Mindermann S. et al. Prioritized training on points that are learnable, worth learning, and not yet learnt. ICML 2022, PMLR 162:15630–15649. Oquab M. et al. DINOv2: Learning robust visual features without supervision. TMLR 2024. Siméoni O. et al. DINOv3. arXiv:2508.10104, 2025.

AuthorsOpenReview ↗

Response to 43bU Weakness 4

Weakness 4. Comparison with existing pathology foundation models. The supplementary material compares the proposed models with UNI, Virchow, Virchow2, Prov-GigaPath, and H-Optimus-0. However, substantial performance gaps remain on several tasks. For example, on COAD/BRAF, ViT-L (run 11) achieves an AUROC of 69.4, compared with 80.9 for Virchow2 and 80.5 for H-Optimus-0. Although the primary goal is to study online data curation rather than to develop a new state-of-the-art model, the authors should clarify the practical significance of these results and whether the proposed strategy can also benefit stronger pre-trained encoders.

We thank the reviewer for this question, and we are happy to clarify both points.

The gap to the published encoders reflects the scale of their training, i.e., the scale of their data, and the epochs covered during training. Our 125k-step ViT-L runs see each tile 1.34 times, whereas UNI saw each of its tiles 3.8 times (Section 4 of the paper), and the other four encoders were trained on 171,189 to 3.1M slides against our 69,790 (Table 8 of the paper). When we trained longer, the gap narrowed. A new ViT-L baseline that we trained for the rebuttal period uses a constant schedule (like DINOv3) and reaches a macro-mean AUROC of 75.45 at 200k steps, compared with 74.21 for run 9 at 124k steps, which brings the gap to UNI from 5.73 to 4.49 pp. We plan to extend this to 400k iterations to further close this gap, hopefully within the discussion period (we will provide the results table to you as an addendum to this response if possible). However, we cannot extend the curated counterpart to 400k steps within the discussion period.

The practical significance of our results lies in the controlled comparison. Each curated run is compared with a baseline that has the same corpus and the same number of steps, so its improvement (+0.91 pp for run 11, our main ViT-L run) is what curation adds at a fixed budget. Curation acts only on which tiles enter the batch and leaves the training objective as it is, so it can be added to the training of a stronger encoder as well, and our response to Weakness 2 gives the curated counterpart of this baseline at 200k steps.

In conclusion, the gap to the published encoders reflects the scale of their training, and it narrowed by 1.24 pp when we trained longer (from 74.21 to 75.45). Our results show what curation adds at a fixed budget (+0.91 pp for run 11). If this clarifies the practical significance of our results, I hope you would take our response into consideration when revisiting the score.

AuthorsOpenReview ↗

Response to 43bU Weakness 3

Weakness 3. Reproducibility and training dynamics. Most curated runs resume from their corresponding baseline checkpoints, so the comparisons share the same pre-training history. While this helps isolate the effect of curation, it remains unclear whether the gains persist across independent training runs. The paper mainly reports downstream results at the final checkpoints, while the training dynamics analysis focuses on the density cache rather than downstream performance. Independent pre-training runs with different random seeds and downstream evaluations at intermediate checkpoints would help clarify the consistency of the reported gains.

We thank the reviewer for these points. They are fair, and we would like to answer them plainly.

On independent runs, most curated runs start from a checkpoint of their baseline, as the reviewer notes. We did this so that a curated run and its baseline differ only in the tiles that they see afterwards. The paper still has three baselines that we trained separately, and Table 1 lists the improvement of every curated run over its baseline. The size of the improvement varies, and its sign does not. We did not repeat a run with only the seed changed. One ViT-L run takes more than a week on eight H100 GPUs, and we chose to spend our compute on more settings and a second model size. We agree that seed repeats would add to this evidence.

On intermediate checkpoints, the paper reports every encoder at the end of its schedule. Our learning rate, weight decay and teacher momentum follow cosine schedules over the full run (Table 3 of the paper). A checkpoint from the middle of a run therefore comes from a model whose training is still under way, and we do not think that it gives a fair comparison. Siméoni et al. (2025) point to a related limit of such schedules in DINOv3, where "employing cosine schedules implies knowing the optimization horizon a priori".

Table 1. Improvement of each curated run over its baseline, in pp of AUROC, as the mean over the 60 task-folds (Table 2 of the paper).

Baseline, trained separatelyCurated runsImprovement of each curated run
Run 1, ViT-B2 to 8+0.80, +0.75, +0.74, +0.62, +0.87, +0.82, +1.04
Run 9, ViT-L10 to 12+0.26, +0.91, +0.42
Run 13, ViT-B with 500k steps14+0.28

We hope we have addressed your concern here as much as was possible in this discussion period.

References. Siméoni O. et al. DINOv3. arXiv:2508.10104, 2025.

Reviewer_Kg3WOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes an online method for selecting training tiles during self-supervised pretraining of pathology image models. A fixed-size cache stores feature signatures from the evolving DINOv2 model and tracks how often similar tiles appear, giving more weight to recent observations. Tiles with less common appearances are more likely to be selected, reducing the dominance of common tissue patterns without relying on a separate encoder. The cache can also be used to generate qualitative tissue segmentation maps.

The authors train ViT-B and ViT-L encoders on 287 million tiles and evaluate their frozen representations on twelve biomarker-prediction tasks. The method yields modest average gains with the shorter training schedule, with more consistent improvements on certain tasks. With longer training, however, the ViT-B model performs about the same overall as the baseline trained without curation.

Critical strengths
  1. The approach works during training without a separate pretrained model. Its fixed-size memory makes it practical to apply to a large image collection.

  2. The experiments are comprehensive. The authors evaluate two model sizes, several sampling settings, twelve prediction tasks, and a longer training schedule. Some task-specific improvements are consistent across configurations.

  3. The paper is transparent about its limitations. It reports failed configurations, tasks where performance decreases, additional computational costs, and weaknesses in the evaluation.

Critical weaknesses
  1. The evaluation may overestimate performance. Slides from the same patient can appear in both training and evaluation sets. The authors also select the best training epoch using the evaluation set rather than a separate validation set. These choices make it difficult to trust improvements that are often smaller than one AUROC point.

  2. The gains may not justify the extra computation. The method previews many tiles that it later rejects. Comparisons match training steps, but not total computing time. In the longer experiment, it performs about the same as ordinary training.

  3. Important comparison methods are missing. The study compares against uncurated training but does not establish whether the proposed counted-cache estimator outperforms simpler diversity sampling, online clustering, or curation using frozen warmup features. Offline tile balancing is already explored in Chen et al., making a controlled comparison particularly informative.

  4. Rare patterns are not always useful patterns. The method also increases the proportion of non-diagnostic material, such as blank regions and fat. The paper does not directly show that the gains come from greater exposure to useful rare tissue patterns.

  5. The tissue segmentation results are only qualitative. The colored tissue maps look plausible, but their accuracy is not measured or compared with maps produced using the baseline model.

Questions for authors
  1. Do the improvements remain when all slides from a patient stay in the same split and a separate validation set is used to select the training epoch?

  2. How consistent are the gains across independent training runs? Can the uncertainty estimates account for overlapping evaluation splits and tasks that share patients?

  3. How does the method compare with ordinary training given the same total GPU time?

  4. Would a simpler diversity sampler, or a sampler using features frozen before curation begins, achieve similar results?

  5. Can you show that the method increases exposure to useful rare tissue patterns, and that this explains the downstream improvements?

  6. Are the segmentation maps better than those obtained by clustering features from the baseline model?

6 comments
AuthorsOpenReview ↗

Response to Kg3W Weakness 1 and Question 1

Weakness 1. The evaluation may overestimate performance. Slides from the same patient can appear in both training and evaluation sets. The authors also select the best training epoch using the evaluation set rather than a separate validation set. These choices make it difficult to trust improvements that are often smaller than one AUROC point.

Question 1. Do the improvements remain when all slides from a patient stay in the same split and a separate validation set is used to select the training epoch?

We thank the reviewer for this point. For improvements below one AUROC point it is a fair thing to ask, and we were glad to find that both concerns can be tested directly on the per-slide scores that we stored for every run. We ran both tests, and the answer to the question is yes, the improvements remain (Table 1).

For the patient overlap, we removed from each held-out fold every slide whose patient also has a slide in the training part of that split, so that no patient is on both sides, and we scored the same trained heads on the slides that were left. The removed slides are 8.3% of the held-out slides on average and 21.3% in BLCA/FGFR3. On the patient-disjoint folds, run 11 (our main ViT-L run) improves on its baseline by +1.00 pp, with a 95% interval of +0.40 to +1.57 pp, compared with +0.91 pp on all slides. The improvement becomes larger in ten of the eleven comparisons.

For the epoch selection, we took the selection out altogether. We fixed the last epoch (120) for every encoder, task and split, so that no held-out slide can influence the choice. Run 11 then improves by +1.23 pp, with a 95% interval of +0.12 to +2.51 pp, and by +1.21 pp when the folds are also patient-disjoint. The pooled mean stays positive in nine of the eleven comparisons.

Table 1. Curated run minus its baseline, in pp of AUROC, as the mean of the fold-paired differences over the 60 task-folds. Runs 2 to 8 are ViT-B, runs 10 to 12 are ViT-L, and run 14 is the ViT-B pair at 500k steps. "Best epoch" is the protocol of the paper, and "epoch 120" fixes the last epoch for every encoder, task and split. The first column of runs 10 and 12 is recomputed from the stored per-slide scores (+0.26 and +0.42 in the paper).

RunAll slides, best epochPatient-disjoint, best epochAll slides, epoch 120Patient-disjoint, epoch 120
2+0.80+1.08+1.29+1.59
3+0.75+0.84+0.71+0.72
4+0.74+0.82+1.06+1.03
5+0.62+0.66+0.66+0.63
6+0.87+1.11+1.16+1.42
7+0.82+1.00+0.43+0.42
8+1.04+1.21+1.30+1.26
10+0.31+0.41−0.02−0.13
11+0.91+1.00+1.23+1.21
12+0.56+0.59+0.82+0.79
14+0.28−0.03−0.18−0.56

To conclude, we ran both of the tests that you asked for, and the improvements remain with every patient kept on one side (+1.00 pp for run 11) and with no selection on the held-out slides (+1.23 pp). We will add Table 1 to the appendix in the revision aimed for the day before the end of this discussion period. We hope we have satisfied this point of contention enough to revisit your score.

AuthorsOpenReview ↗

Response to Kg3W Question 2

Question 2. How consistent are the gains across independent training runs? Can the uncertainty estimates account for overlapping evaluation splits and tasks that share patients?

We thank the reviewer for these two questions, and we can surely provide some numbers to back our claims and improve comprehension to your satisfaction.

On consistency, the paper has eleven curated runs, compared with three baselines that we trained separately, and Table 1 lists the improvement of each one. while the size of the improvement varies, its sign does not. The seven ViT-B runs lie between +0.62 and +1.04 pp.

On uncertainty, the answer is yes on both counts. For the overlapping splits (any two held-out sets share 27% of their slides), we applied the correction of Nadeau and Bengio (2003), which widens the standard error of the precision-weighted mean by a factor of 1.87, and the improvement of run 11, our main ViT-L run, remains significant (p=0.008p = 0.008). For the tasks that share patients, we ran a bootstrap over patients, in which a patient brings all of its slides and the tasks of one cohort draw the same patients. It gives run 11 an improvement of +0.91 pp, with a 95% interval of +0.06 to +1.79 pp.

Table 1. Improvement of each curated run over its baseline, in pp of AUROC, as the mean over the 60 task-folds (Table 2 of the paper).

Baseline, trained separatelyCurated runsImprovement of each curated run
Run 1, ViT-B2 to 8+0.80, +0.75, +0.74, +0.62, +0.87, +0.82, +1.04
Run 9, ViT-L10 to 12+0.26, +0.91, +0.42
Run 13, ViT-B with 500k steps14+0.28

Therefore, the improvement appears in all eleven curated runs of the paper, on three baselines that we trained separately (Table 1). The improvement of run 11 also holds when the overlapping splits and the shared patients are taken into account (p=0.008p = 0.008, and an interval of +0.06 to +1.79 pp), and we will add both analyses to the appendix. We hope that this additional evidence helps you in reconsidering your score.

References. Nadeau C., Bengio Y. Inference for the generalization error. Machine Learning 52:239–281, 2003.

AuthorsOpenReview ↗

Response to Kg3W Weakness 5 and Question 6

Weakness 5. The tissue segmentation results are only qualitative. The colored tissue maps look plausible, but their accuracy is not measured or compared with maps produced using the baseline model.

Question 6. Are the segmentation maps better than those obtained by clustering features from the baseline model?

We thank the reviewer for this suggestion, which we were happy to follow. We measured our maps against pixel annotations and compared them with maps from the baseline model, and the two are on par in mIoU (Table 1).

We used 10 CAMELYON16 slides (tumor and normal) and 15 TIGER slides (five tissue classes). Our maps come from the cache of run 11, with 18 classes, and the baseline maps from k-means with 18 clusters on the features of run 9 (the uncurated baseline), fitted to each slide separately. No difference in mIoU reaches p<0.05p \lt 0.05, on either dataset and under any of the three matching rules.

What our maps add is where their classes come from. The model builds them itself during pre-training, as a by-product of the cache, and all slides share them. Per-slide k-means has to be fitted again on every slide, so its clusters carry no identity from one slide to the next. Qualitative maps are also a common way to show what a self-supervised encoder has learned (the attention maps of DINO, Caron et al., 2021, and the PCA maps of DINOv2, Oquab et al., 2024), but thanks to your intervention, Table 1 now adds a measurement.

Table 1. Our maps minus the baseline maps, both scored against the pixel annotation, as the mean over slides ± its standard error with the paired t-test p-value. A positive value favors our maps. The three mIoU rows differ in how clusters are matched to the annotated classes.

Score against the pixel annotationOurs − baseline, CAMELYON16 (10 slides)Ours − baseline, TIGER (15 slides)
mIoU, each cluster named by its majority class−0.053 ± 0.046 (p=0.27p = 0.27)−0.064 ± 0.037 (p=0.10p = 0.10)
mIoU, clusters matched one-to-one to classes (Hungarian)−0.052 ± 0.046 (p=0.29p = 0.29)−0.019 ± 0.040 (p=0.65p = 0.65)
mIoU, each cluster named by its most enriched class−0.099 ± 0.044 (p=0.052p = 0.052)−0.056 ± 0.037 (p=0.16p = 0.16)
Tile accuracy, share of tiles given their annotated class−0.018 ± 0.006 (p=0.019p = 0.019)−0.028 ± 0.027 (p=0.32p = 0.32)
ARI, agreement of clusters and classes, no naming needed+0.009 ± 0.003 (p=0.029p = 0.029)−0.015 ± 0.047 (p=0.75p = 0.75)
NMI, information shared by clusters and classes, no naming needed−0.017 ± 0.007 (p=0.047p = 0.047)−0.028 ± 0.029 (p=0.35p = 0.35)

Thus, we have now measured the maps against pixel annotations and compared them with maps from the baseline model, as you asked. The two are on par in mIoU, and ours come with classes that all slides share. We will add Table 1 to the appendix if this adds to your consideration for a revised score.

Reviewer_ENiyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a method for online autocuration during pathology pretraining by tracking a density cache of M signatures. These signatures can store the morphologies learned during the pretraining and help avoid overlapped morphologies, while preferring the new ones. Empirical results highlight its outstanding performances.

Critical strengths
  • Good visualizations
  • Extensive empirical results
Critical weaknesses
  • Signature estimation: What is the rationale behind the proposed signature estimation? My understanding is that it projects representations from DINO heads, which may be unstable or arbitrary during early training iterations.

  • Connection to nonparametric clustering: The cache mechanism resembles nonparametric clustering, where sparse assignments are encouraged through a nonparametric prior [1]. Such methods can also learn meaningful prototypes corresponding to tissue morphologies. Could the authors provide similar analyses or visualizations?

  • Hyperparameter sensitivity: The method appears to require substantial hyperparameter tuning. Could these parameters be learned directly, or could the design be made less heuristic?

[1] Nguyen et al., Nonparametric Distribution Matching for Self-Supervised Whole-Slide Image Condensation, NeurIPS 2026

Questions for authors

See Weaknesses

Comments on the paper

1 comment
Area_Chair_wZTdOpenReview ↗

A Gentle Nudge: Let's Keep the Discussion Going!

Dear Reviewers,

The authors have now posted their responses, addressing your concerns and incorporating your suggestions. We are already halfway through the discussion period!

One of the distinguishing features of ICLR is its interactive review process, where constructive discussions not only help strengthen individual papers but also contribute valuable insights to the broader scientific community.

So, let's not let the discussion be a one-way street! I encourage you to revisit the authors' responses, acknowledge their efforts, and engage in further discussion where needed. After all, a good scientific discussion is often as valuable as the final verdict!

-AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.