{"id":"b97defc6-3427-417d-832d-e30e26d1dc1f","arxiv_id":"2411.18225","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"PATHS, a top-down hierarchical transformer that repeatedly zooms into the most important patches of a whole slide image, reports a mean survival c-index of 0.665 over five TCGA datasets while reading only a small fraction of each slide.","lead":"This paper presents PATHS, a model that analyzes cancer tissue slides by first viewing the whole slide at low magnification and then automatically zooming into the most important regions, the way a pathologist works. It reports survival predictions as good as or better than full-slide models while reading far fewer image patches and running over ten times faster at inference.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's accuracy comparison is confounded by encoder and magnification: five of six baselines use pre-calculated scores from Chen et al. with a different feature extractor at 20x, while PATHS uses UNI at <=10x; Table 5 shows the encoder alone shifts PATHS by more than the reported gain over HIPT.","rationale":"Good-faith reading: PATHS is a clear, well-motivated contribution; the strongest and best-supported claims are efficiency (Appendix C, Figure 3) and interpretability (Figures 4, 6, 7, including counterexamples). The central accuracy claim is conditional. The paper is transparent about the provenance of baseline scores, which is a credit, but transparency does not remove the confound. The reader's weakest assumption is exactly the right one. I agree with the CONDITIONAL verdict. The concern is not that the method is bad: the controlled ZoomMIL comparison and the ablations suggest a real but modest architecture benefit. The issue is that Table 1 cannot support the comparative superiority claim as stated. The concrete test would settle it: same-encoder, same-magnification baselines. If those comparisons are run and PATHS remains ahead, the headline is justified. If not, the paper should be accepted for its efficiency and interpretability contributions with the accuracy claims scaled back. No ad hominem is intended; the issue is experimental design, not integrity.","tokens_in":15455,"tokens_out":5121,"duration_ms":44859,"concrete_test":"Re-run the Table 1 baselines using the same UNI encoder and the same 10x bottom-level patch extraction as PATHS, on the same five folds. A minimal version: run ABMIL and HIPT on UNI features at 10x for IDC, CRC, CCRCC, PRCC and LUAD (and ideally all five baselines on the PATHS 0.625x-10x patch grid), then compare mean c-indices and run a paired permutation test across patients or folds. If PATHS still beats a same-encoder HIPT by more than the fold-level noise, the architecture claim holds; if ABMIL or HIPT with UNI matches or exceeds PATHS, the 'superior performance' statement should be revised to claim efficiency and interpretability advantages only.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline 'superior performance on slide-level prediction tasks' rests on Table 1, but Table 1 is not a controlled architecture comparison. Five of the six baselines (ABMIL, DeepAttnMISL, GCN-MIL, DS-MIL, HIPT) are evaluated with pre-calculated risk scores from Chen et al. [6], which used a different patch encoder (the HIPT SSL-ViT) and patches extracted at 20x magnification. PATHS instead uses UNI features at magnifications from 0.625x to 10x. The paper's own Table 5 shows that the choice of encoder moves PATHS mean c-index from 0.553 (SSL-ViT) to 0.665 (UNI), a +0.112 shift. That is larger than the reported PATHS-vs-HIPT gain of +0.047 and the PATHS-vs-ZoomMIL gain of +0.049. Consequently, the accuracy advantage in Table 1 is plausibly due to the feature extractor, not to the PATHS hierarchy or patch selection. The one same-encoder comparison, versus ZoomMIL (both UNI), shows only a +0.049 mean gain, with no significance test; PATHS is worse on LUAD (0.545 vs 0.551), tied on IDC (0.636 vs 0.634), and fold standard deviations are 0.036-0.097. The ablation (Table 2) also shows that replacing the learned importance with random selection costs only 0.009 mean c-index, so the distinctive selection mechanism contributes little to accuracy on these tasks. Efficiency and interpretability claims are independently supported, but the accuracy claim is under-supported as published.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"PATHS is a top-down, hierarchical transformer for whole-slide image analysis that recursively selects a small number of patches (at most K per magnification level) from a WSI pyramid, encodes them with a frozen pre-trained encoder (UNI), and aggregates them with per-level transformers. The paper evaluates the method on survival prediction across five TCGA datasets (IDC, CRC, CCRCC, PRCC, LUAD), reporting a mean c-index of 0.665, which it claims is superior to previous methods while processing far fewer patches, and also reports large inference-speed gains. Additional contributions are an interpretable importance-based patch selection, an ablation study of the architectural components, and zero-shot visualizations on CAMELYON17 showing that selected regions align with tumor annotations.","tokens_in":15696,"tokens_out":3984,"duration_ms":35421,"significance":"If the accuracy claim held under controlled comparison, PATHS would be a meaningful advance: it shows that a top-down hierarchical selector can match or beat full-slide bag-of-patches methods with a fraction of the compute, and its interpretable importance values are clinically attractive. The efficiency evidence is strong and well documented (Figures 3 and 5, Appendix C), and the zero-shot interpretability results on CAMELYON17 (Figures 4 and 7) are convincing and reproducible because the code is provided. The main weakness is that the central accuracy claim, as stated in the abstract and Section 5, is not supported by a controlled architecture comparison: five of six baselines use pre-calculated scores from a different encoder and magnification, and the paper's own Table 5 shows that the encoder choice alone explains more than the reported advantage.","major_comments":[{"comment":"The headline accuracy claim is confounded by the feature encoder and magnification. Five of the six baselines (ABMIL, DeepAttnMISL, GCN-MIL, DS-MIL, HIPT) are evaluated using 'pre-calculated risk scores' from Chen et al. [6], which used the SSL-ViT encoder and 20x patches, whereas PATHS uses UNI at magnifications from 0.625x to 10x. The paper's own Table 5 shows that changing the encoder from SSL-ViT to UNI shifts PATHS mean c-index from 0.553 to 0.665 (a +0.112 improvement), which is larger than the reported PATHS-vs-HIPT gain of +0.047 and the PATHS-vs-ZoomMIL gain of +0.049. Consequently, the 'superior performance' claimed in the abstract and Section 5 is not a controlled architecture comparison. The authors must either rerun the baselines with the same UNI encoder (or otherwise equalize the feature extraction), or substantially qualify the accuracy claim to be about the full pipeline rather than the PATHS architecture itself. The same issue applies to the statement in Section 5 that PATHS 'significantly improves' accuracy over prior methods.","section":"Table 1 and Table 5; Section 4 (Baselines and Patch Embedding)"},{"comment":"The only same-encoder comparison, against ZoomMIL (both use UNI), does not firmly establish an architecture-level advantage. PATHS has a higher mean c-index (0.665 vs 0.616), but it is worse on LUAD (0.545 vs 0.551) and essentially tied on IDC (0.636 vs 0.634), with per-fold standard deviations between 0.036 and 0.097. No significance tests or confidence intervals are reported, so the difference could easily arise from fold variance. Please add paired significance tests (or at least per-fold paired differences) and discuss the LUAD/IDC results explicitly.","section":"Table 1, ZoomMIL row; Section 4 (Baselines)"},{"comment":"The random-selection ablation is not a clean test of the learned selection mechanism. The ablation replaces the predicted α_i with U[0,1] at inference time only, which changes both which patches are selected and the scaling of patch features in Eq. (7) (line 4 of Algorithm 1). Because the aggregator was trained with MLP-produced α values, random α at test time introduces a distribution shift that is not equivalent to random patch selection during training. The observed small drop of 0.009 mean c-index therefore does not cleanly show that learned selection is unimportant. The authors should retrain with a random top-K selection inside the training loop to isolate the effect of the learned importance model.","section":"Table 2 and Algorithm 1, Eq. (7)"}],"minor_comments":[{"comment":"The phrase 'superior performance' overstates the evidence given the encoder confound; recommend rewording to 'comparable or better under matched encoder' or similar until the controlled comparison is provided.","section":"Abstract and Section 5"},{"comment":"Please state whether the 'pre-calculated risk scores' from [6] use the same five folds and the same preprocessing (patch size, Otsu threshold, tissue segmentation) as the PATHS experiments; otherwise the claim that 'all models are evaluated on the same folds' is not verifiable.","section":"Section 4 (Baselines)"},{"comment":"The assignment arrows '← −' are nonstandard and potentially confusing; please use standard notation such as '←' throughout the pseudo-code.","section":"Algorithm 1"},{"comment":"The column headers 'Neither', 'Hierarchical only', and 'Slide-level only' are ambiguous; please define in the caption exactly which context components are included in each variant (e.g., 'Neither' = no hierarchical context and no slide-level context).","section":"Table 2"},{"comment":"The phrase 'reading in fewer uninformative slides' should be 'reading in fewer uninformative patches'; the current wording is a typo that could confuse readers.","section":"Discussion, second paragraph"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid efficiency and interpretability contribution, and the code availability is a plus. The main risk is that the accuracy claim is not yet supported by a controlled architecture comparison; the authors need to either run the baselines with the same UNI encoder or retitle the claim to be about the full PATHS pipeline. If the controlled comparison still shows a meaningful advantage, the paper would be acceptable; otherwise the accuracy section should be rewritten to reflect the encoder-driven gains. The paper is within scope for a computer vision venue, but the authors should also clarify the fold-matching of pre-calculated scores from [6]."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this paper for the efficiency and interpretability results; read the accuracy headline with caution. What's actually new is the combination of top-down traversal with hard top-K patch selection, RNN-propagated hierarchical context (Eq. 6), and slide-level context across magnifications (Eq. 4). That specific combination is not in ZoomMIL, HIPT, or random-sampling baselines. The inference-speed evidence is concrete: Appendix C documents the latency protocol and patch counts, and the hardware-independent patch-count plot is useful. The interpretability section is honest—zero-shot CAMELYON17 heatmaps, first eight slides alphabetically, with examples that miss micrometastases. That is not cherry-picking. Code is public. The soft spot is the accuracy comparison, and it is real. In Table 1, five of six baselines (ABMIL, DeepAttnMISL, GCN-MIL, DS-MIL, HIPT) are pre-calculated risk scores from Chen et al. [6], which used the SSL-ViT encoder on 20x patches. PATHS uses UNI at 0.625x–10x. Table 5 shows the encoder alone moves PATHS mean c-index from 0.553 to 0.665—a +0.112 shift, larger than the reported +0.047 gain over HIPT and +0.049 over ZoomMIL. So the \"superior performance\" claim is plausibly explained by the feature extractor, not the architecture. The one controlled comparison, versus ZoomMIL (both UNI), shows a modest +0.049 mean gain, PATHS is worse on LUAD and tied on IDC, and there are no significance tests. The ablation also shows random selection costs only 0.009 mean c-index, so the distinctive selection mechanism contributes little on these tasks. To the paper's credit, these limitations are visible in the text—they discuss the random-vs-attentional margin and the LUAD failure—but the abstract does not hedge. The efficiency and interpretability claims stand. The accuracy claim needs either same-encoder baselines or a re-scoped statement. Who is this for? Anyone working on efficient WSI processing or interpretable patch selection. It deserves a serious referee: it is a legitimate new method with reproducible efficiency evidence. If I were reviewing, I would push for major revision that either adds UNI-based baselines or softens the claim to \"comparable performance at a fraction of the cost.\"","headline":"A genuinely useful efficiency + interpretability contribution whose accuracy headline is over-reached by an encoder-confounded baseline comparison.","tokens_in":16415,"tokens_out":2338,"would_cite":true,"duration_ms":20753,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"PATHS, a top-down hierarchical transformer that reads only a few hundred patches per slide, reports superior survival-prediction performance on four of five TCGA datasets and over ten times faster inference than full-slide methods.","keywords":["whole slide image","computational pathology","hierarchical transformer","survival prediction","patch selection","multiple instance learning","gigapixel image","attention"],"falsifier":"Recompute the five baselines with the same UNI encoder and 10x magnification on the same folds, or run PATHS with the SSL-ViT encoder used for the [6] baselines. If the mean c-index gap over HIPT (reported as 0.047) collapses to within noise, then the top-down hierarchical selection is not the source of the reported improvement.","tokens_in":15077,"feed_emoji":"🔬","tokens_out":5049,"duration_ms":40707,"temperature":0.7,"pith_summary":"Whole-slide images are gigapixel; standard methods treat each slide as a bag of tens of thousands of patches, which is expensive and noise-prone. This paper proposes PATHS, a top-down hierarchical transformer that mimics a pathologist: it starts at low magnification, learns an importance score for each patch, keeps only the top K (here K=20, so at most 80 per level), zooms into those regions, and repeats across five magnification levels. The paper reports that PATHS outperforms existing methods on survival prediction on four of five TCGA datasets (mean c-index 0.665 vs 0.618 for HIPT) while processing only a few hundred patches per slide and running over ten times faster at inference than full-slide methods. The authors' claim is that reading a small, carefully chosen subset of patches at multiple magnifications is both more accurate and far cheaper than reading the whole slide.","feed_headline":"Reading 80 patches per zoom level beats reading the whole slide","feed_subtitle":"A top-down transformer mimics pathologists, cutting inference time over tenfold while improving survival prediction on four of five cancer…","key_machinery":"The central object is the PATHS hierarchy: n processors, one per magnification level mi forming a geometric sequence with multiplier M=2 from 0.625x to 10x. Each processor computes contextualized patch features by adding an RNN output over the embeddings of ancestor patches at lower magnifications (hierarchical context), predicts a scalar importance α via an MLP with sigmoid, scales features by α, and feeds them to a transformer decoder with 2D positional encoding for global aggregation. A top-K filter keeps the 20 most important patches per level, and MAGNIFY expands those locations at the next magnification, giving a hard upper bound of $M^{2}$K = 80 patches per level. The slide-level features from all levels are summed as slide-level context, making the model a residual hierarchy. This machinery is what enables quadratic self-attention on a small patch set and provides the interpretable importance scores.","core_discovery":"The paper's central claim is that a top-down, cross-magnification hierarchy with learned patch importance can outperform full-slide bag-of-patches and bottom-up hierarchical baselines on slide-level survival prediction, despite processing only a small fraction of the slide. The authors argue that most patches in a whole-slide image are uninformative noise, and that mimicking a pathologist's 'scan, zoom, re-scan' behavior both improves the signal-to-noise ratio and makes quadratic self-attention feasible. They support this with mean c-index 0.665 across IDC, CRC, CCRCC, PRCC and LUAD, best on four of five datasets, and an inference speed-up exceeding a factor of ten at 10x magnification. They also show the learned importance scores localize tumor regions in zero-shot CAMELYON17 visualizations.","pith_inferences":["The reported gain over HIPT may stem largely from switching the patch encoder to UNI: the paper's own Table 5 shows the encoder moves mean c-index by 0.112, more than the 0.047 gain over HIPT, so an encoder-controlled comparison is needed before attributing the improvement to the hierarchy.","Random patch selection with cross-magnification context scores nearly as high as learned selection (mean 0.656 vs 0.665), suggesting the importance module's main added value may be interpretability rather than accuracy.","Since UNI was trained only at 20x, a cross-magnification pre-trained encoder, as the paper suggests, would likely push PATHS further.","The same top-down selection principle could apply to other gigapixel domains, such as satellite imagery, where most pixels are irrelevant to the task."],"forward_implications":["On the five TCGA survival tasks, reading a few hundred carefully chosen patches per slide suffices to match or beat methods that read tens of thousands of patches.","The fixed bound of 80 patches per magnification level makes quadratic self-attention practical for slide-level aggregation.","The learned importance scores localize tumor regions without patch-level labels, as shown in zero-shot CAMELYON17 examples.","Inference latency for a new slide is cut by more than a factor of ten at 10x, because only selected patches are encoded.","Because the method is task-agnostic, it transfers to other weakly supervised gigapixel image tasks."],"supporting_citations":[{"why":"Supplies the HIPT hierarchical baseline and the pre-calculated risk scores used for ABMIL, DeepAttnMISL, GCN-MIL, DS-MIL and HIPT comparisons.","marker":"[6]"},{"why":"UNI patch encoder that PATHS and ZoomMIL use; encoder choice largely determines performance.","marker":"[7]"},{"why":"ZoomMIL is the matched learnable-patch-selection MIL baseline compared under the same patch budget and encoder.","marker":"[28]"},{"why":"ABMIL is the canonical attention MIL baseline that represents full-slide bag-of-patches processing.","marker":"[15]"},{"why":"Shows that random sampling of few patches can already support survival prediction, backing the paper's efficiency premise.","marker":"[30]"},{"why":"DS-MIL is a multi-magnification MIL baseline and prior evidence that cross-magnification information helps.","marker":"[17]"},{"why":"Multi-resolution MIL work supporting the claimed benefit of incorporating multiple magnification levels.","marker":"[18]"},{"why":"DINO self-supervised pretraining for HIPT's bottom hierarchy stages, relevant to the baseline's setup.","marker":"[3]"}],"fun_headline_variants":["Pathologist-inspired zooming cuts slide analysis time 10x","A few informative patches beat the whole slide for cancer prognosis","PATHS: top-down transformer mimics pathologist, boosts survival prediction","Reading 80 patches per zoom level beats whole slide","Zoom-in hierarchy improves cancer survival prediction, cuts compute"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed performance advantage rests on the comparability of PATHS scores (computed with the UNI encoder at magnifications up to 10x) to pre-calculated baseline risk scores taken from Chen et al. [6], which used a different patch encoder at 20x; if this encoder mismatch explains the gap, the architecture itself has not been shown superior.","fun_headline_variants_meta":{"raw":{"variants":["Pathologist-inspired zooming cuts slide analysis time 10x","A few informative patches beat the whole slide for cancer prognosis","PATHS: top-down transformer mimics pathologist, boosts survival prediction","Reading 80 patches per zoom level beats whole slide","Zoom-in hierarchy improves cancer survival prediction, cuts compute"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000824,"raw_usage":{"total_tokens":3600,"prompt_tokens":938,"completion_tokens":2662,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":2580}},"tokens_in":554,"tokens_out":2662,"duration_ms":18718,"temperature":1.0,"reasoning_tokens":2580,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:25:11.374767+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the five baselines with the same UNI encoder and 10x magnification on the same folds, or run PATHS with the SSL-ViT encoder used for the [6] baselines. If the mean c-index gap over HIPT (reported as 0.047) collapses to within noise, then the top-down hierarchical selection is not the source of the reported improvement.","supporting_citations":[{"cited_title":"Chen, Chengkuan Chen, Yicong Li, Tiffany Y","cited_arxiv_id":null,"evidence_quote":"Supplies the HIPT hierarchical baseline and the pre-calculated risk scores used for ABMIL, DeepAttnMISL, GCN-MIL, DS-MIL and HIPT comparisons."},{"cited_title":"Towards a general-purpose foundation model for com- putational pathology","cited_arxiv_id":null,"evidence_quote":"UNI patch encoder that PATHS and ZoomMIL use; encoder choice largely determines performance."},{"cited_title":"Differentiable zooming for multiple instance learn- ing on whole-slide images","cited_arxiv_id":null,"evidence_quote":"ZoomMIL is the matched learnable-patch-selection MIL baseline compared under the same patch budget and encoder."},{"cited_title":"Attention-based deep multiple instance learning","cited_arxiv_id":null,"evidence_quote":"ABMIL is the canonical attention MIL baseline that represents full-slide bag-of-patches processing."},{"cited_title":"Steiner, Zhaoyang Xu, Apaar Sad- hwani, Hongwu Wang, Isabelle Flament-Auvigne, Craig H","cited_arxiv_id":null,"evidence_quote":"Shows that random sampling of few patches can already support survival prediction, backing the paper's efficiency premise."},{"cited_title":"Eliceiri","cited_arxiv_id":null,"evidence_quote":"DS-MIL is a multi-magnification MIL baseline and prior evidence that cross-magnification information helps."},{"cited_title":"Sisk, Huihui Ye, William D","cited_arxiv_id":null,"evidence_quote":"Multi-resolution MIL work supporting the claimed benefit of incorporating multiple magnification levels."},{"cited_title":"Emerg- ing properties in self-supervised vision transformers","cited_arxiv_id":null,"evidence_quote":"DINO self-supervised pretraining for HIPT's bottom hierarchy stages, relevant to the baseline's setup."}],"review_version":1}