{"id":"5a8efdc6-53b8-4007-9bbd-53ea5a7e927c","arxiv_id":"2412.10853","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SEW combines a thumbnail-level graph classifier with a focus predictor and local detail branch to classify whole-slide pathology images quickly, and uses clustered features to propose new prognostic tumor markers.","lead":"This paper introduces SEW, a system that reads cancer biopsy slides in two steps: a fast global look at the whole slide, then a zoomed-in check of suspicious regions. It reports top accuracy and much faster inference than comparable methods, and uses focused features to suggest possible new tumor markers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The local branch's supervision on public datasets is underspecified: §3.3.3 requires lesion masks, but §4.1 only confirms masks for in-house HCC/GC/CRC; for PANDA/BRCA/LUAD no mask source is given, so SEW's SOTA claim may rest on unavailable supervision.","rationale":"The reader's weakest assumption is exactly the gap I would stress. The SEW pipeline has three interacting supervision signals: L_focus needs a heatmap, L_local_CLS needs group-level labels, and the consistency constraint L_cst connects global nodes to local class tokens. Of these, L_local_CLS is the only one that injects lesion-area knowledge into the local branch; without it, the local branch cannot learn to distinguish lesion from non-lesion, and the 'self-calibration' loop in §3.2 collapses. The paper states in §4.1 that only the three in-house datasets have pathologist lesion-area annotations. For CAMELYON16, pixel-level tumor masks are available in the challenge training set, but PANDA, BRCA, and LUAD are not supplied with such masks. The paper does not describe any automatic mask generation, nor does it state that slide-level labels are broadcast. This is not a matter of tuning: the loss definitions in §3.3.3 require y_t_gt. A reader cannot reproduce Table 1 without making an unstated assumption. The claimed speed advantage is less threatened: thumbnail-graph inference is plausibly fast. The marker-mining claim has a separate circularity, but it is secondary; if the classification claim is built on an underspecified supervision protocol, the marker claim inherits that fragility. I therefore keep the reader's CONDITIONAL verdict and ask for the concrete data-provenance check.","tokens_in":12282,"tokens_out":4112,"duration_ms":39409,"concrete_test":"Ask the authors for the exact training-time computation of y_t_gt in L_local_CLS for CAMELYON16, PANDA, BRCA, and LUAD, and inspect the released preprocessing/training code. Concretely, instrument the code to log the provenance of every y_t_gt: if any public-dataset y_t_gt is derived from slide-level labels or from an undocumented mask-generation step, rerun the Table 1 experiments with that step ablated and with the documented supervision; if SEW no longer exceeds all baselines on PANDA/BRCA/LUAD, the SOTA claim fails. A minimal version: train SEW on PANDA with only slide-level labels by setting all y_t_gt to the slide label, and compare against CLAM/TransMIL under identical splits.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim 'highest accuracy and fastest speed on all types' depends on the full SEW pipeline being trainable on CAMELYON16, PANDA, BRCA, and LUAD as well as the three in-house sets. Section 3.3.3 defines L_local_CLS with y_t_gt that 'indicates whether the corresponding area belongs to the lesion area' when a lesion mask is supplied, or a multi-dimensional one-hot region label. Section 4.1 states that only HCC, GC, and CRC are 'annotated by professional pathologists with the lesion area and grade'. For CAMELYON16, some pixel-level tumor annotations exist in the challenge data, but PANDA, BRCA, and LUAD do not have standard lesion-area masks; the paper never specifies how y_t_gt is obtained for these datasets. If slide-level labels are broadcast to every focused group, the local branch degenerates to a MIL-like patch classifier and the self-calibration mechanism described in §3.2–3.3 is not actually trained as claimed. If masks were produced by an undocumented heuristic, the results depend on that heuristic and are not reproducible. This gap directly affects the headline comparison: a method that needs per-slide lesion annotations is not fairly compared with weakly supervised MIL baselines that use only slide labels.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SEW, a whole-slide pathology image analysis framework with three components: a global branch that classifies a thumbnail via a superpixel graph and transformer, a focus predictor that selects top-K regions from the global branch's features, and a local branch that classifies these magnified regions using lesion-area supervision. A consistency constraint (L_cst) aligns global and local features, and a 'pathological prototype vocabulary' is formed by k-means clustering for final classification and tumor-marker mining. Experiments on seven datasets (PANDA, CAMELYON16, BRCA, LUAD, and three in-house sets HCC, GC, CRC) report classification accuracy and inference time, claiming state-of-the-art accuracy and the fastest speed (5.44-10.97 s/slide). The paper also claims discovery of two novel colorectal cancer prognostic markers (mucinous lakes and necrosis within glands) plus a tissue-infiltration spatial marker (§4.3).","tokens_in":12561,"tokens_out":9120,"duration_ms":77109,"significance":"If the claims were fully supported, the framework would be a valuable contribution to computational pathology, combining interpretability, high inference speed, and competitive accuracy, and the marker-mining pipeline could aid biomarker discovery. The paper's strengths include a clearly described architecture with ablations showing that each component helps (Table 4), a generalization study (Table 2), a superpixel-vs-patch comparison (Table 3), and an explicit focus on inference speed. However, the evidence for the headline accuracy claim is statistically weak (margins are often within one standard deviation), the supervision requirements for public datasets are underspecified, a highly relevant speed-focused baseline (TransMIL) is omitted from comparisons, and the marker-mining section relies on in-sample clustering without external validation. These issues currently limit the significance of the results to a promising but not fully established framework.","major_comments":[{"comment":"The local branch supervision is not specified for the public datasets. The loss L_local_CLS uses y_t_gt, which 'indicates whether the corresponding area belongs to the lesion area' when a lesion mask is supplied, or a multi-dimensional one-hot region label. Section 4.1 confirms such annotations only for the three in-house datasets (HCC, GC, CRC). For PANDA, CAMELYON16, BRCA, and LUAD, the paper does not state how y_t_gt is obtained; CAMELYON16 has pixel-level tumor annotations, but standard lesion-area masks are not defined for PANDA, BRCA, and LUAD. If slide-level labels are broadcast to every focused group, the local branch degenerates to a patch-level MIL classifier and the self-calibration mechanism described in §3.2-3.3 is not trained as claimed. If masks are generated by an undocumented heuristic, the results depend on that heuristic and are not reproducible. This gap directly affects the fairness of the SOTA comparison in Table 1, because SEW may be using per-slide lesion annotations that weakly supervised baselines (e.g., CLAM, TransMIL) do not receive. Please specify the mask source for each dataset or adapt the local branch to use only slide-level labels.","section":"§3.3.3, §4.1"},{"comment":"The tumor marker mining is circular with respect to the training data. The CRC model is trained on prognosis labels, and the same cohort (100 cases with follow-up) is then used to extract focused-region features and cluster them. Because the features were optimized to separate good from poor prognosis, the appearance of poor-prognosis-only clusters (red points in Fig. 2a) is expected and does not independently validate 'novel tumor markers.' No external or held-out cohort is used to confirm that mucinous lakes, necrosis within glands, and degree of infiltration carry independent prognostic information. The pathologist's verification is qualitative and does not substitute for a statistical association with prognosis in unseen data. Please either validate the markers on an independent cohort (e.g., show that the identified markers are correlated with prognosis in a separate dataset) or substantially temper the claims in the abstract and conclusion.","section":"§4.3, §3.6"},{"comment":"The accuracy comparison lacks statistical significance testing. The paper reports mean±std but no p-values, confidence intervals, or number of runs/folds. On most datasets the margin over the best baseline is within one standard deviation (e.g., CAMELYON16: SEW 85.69±0.85 vs Tea-Graph 85.62±1.14; HCC: SEW 87.93±0.63 vs CLAM 87.83±1.53; BRCA: SEW 87.44±0.94 vs HIPT 87.26±2.25). Without paired significance tests (e.g., Wilcoxon signed-rank or paired t-test over the same folds/runs), the statement that SEW 'achieves the highest accuracy' (§4.2) is not supported. Please add appropriate statistical tests and report effect sizes or confidence intervals, or qualify the accuracy claim accordingly.","section":"Table 1, §4.2"},{"comment":"TransMIL, a highly relevant speed-focused baseline, is discussed in §2.2 and §4.2 but is missing from Table 1. Since the paper's headline claim includes 'the fastest speed' (§4.2), the comparison should include TransMIL under the same timing protocol. Moreover, §4.2 states that inference time includes pre-processing and prediction, but then says TransMIL and ZoomMIL have pre-processing that 'takes hundreds of seconds.' The reported ZoomMIL times in Table 1 (e.g., 7.58s for PANDA, 428.19s for CAMELYON16) appear inconsistent with that statement unless pre-processing is sometimes included and sometimes not. Please clarify the timing protocol for every method and include TransMIL in the comparison.","section":"§4.2, Table 1"},{"comment":"The 'pathological prototype vocabulary' is not defined as a shared cross-slide vocabulary. In §3.4.1, k-means is applied to node representations from the local graphs of 'the current WSI,' yielding cluster centers O_c for that WSI. The text claims this vocabulary 'reinforce[s] feature consistency across diverse WSI samples,' but no loss term or update rule is given to align prototypes across slides, and the final prediction in §3.5 averages per-slide cluster centers. This matters for the spatial-distribution marker analysis in §4.3, which assumes that the same semantic tissue type is assigned the same prototype across different WSIs. Please specify how a global prototype vocabulary is constructed, how it is shared across slides, and how it is used during training and inference.","section":"§3.4.1, §3.5"},{"comment":"The consistency loss L_cst = DKL(W_proj_cls CLS_t_local || h'_n) is undefined as written. DKL is a divergence between probability distributions, but both operands are d-dimensional real-valued vectors. Unless a normalization step (e.g., softmax or softmax temperature scaling) is applied to convert both vectors into distributions, the KL divergence cannot be computed. Please define the exact transformation applied to the vectors before computing the divergence, or replace the loss with a bounded similarity measure such as cosine distance or mean squared error.","section":"§3.3.4"}],"minor_comments":[{"comment":"The classification tasks for the public datasets are not specified: the paper does not state whether PANDA is Gleason grading, CAMELYON16 is lymph-node metastasis detection, BRCA is a specific breast cancer task (e.g., ER status or grade), and LUAD is a particular subtype or stage classification. This information is needed for reproducibility and for interpreting the reported accuracy values.","section":"§4.1"},{"comment":"The paper uses 'tumor maker' instead of 'tumor marker' in the Section 4.3 heading, in the conclusion, and in the contribution list ('new tumor mark finding'). These should be corrected to 'tumor marker.'","section":"§4.3, §5, §1"},{"comment":"The notation is confusing in the second equation: CLS_t_local appears on both sides. Please use a distinct symbol (e.g., CLS''_t_local) for the output of the inter-group attention to avoid self-referential notation.","section":"§3.3.3"},{"comment":"In the CRC row, there appears to be an extra '757.79' value before SEW's time of 9.82; please clean up the table formatting.","section":"Table 1"},{"comment":"The phrase 'form the pathological prototype vocabulary for the current WSI' conflicts with the claim that the vocabulary enforces consistency across diverse WSI samples. Please reword to make clear whether the prototypes are computed per-slide or globally.","section":"§3.4.1"},{"comment":"The baseline name is written as 'TeaGraph' in §2.3 and 'Tea-Graph' in §4.2; please use a consistent spelling.","section":"§2.3, §4.2"},{"comment":"The focus predictor loss uses Q_gt, but the paper does not formally define how Q_gt is constructed when the pseudo-label switches from Grad-CAM to the local branch's prediction. Please specify the schedule and the exact form of Q_gt in each phase.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper makes strong claims of 'novel prognostic tumor markers' and 'highest accuracy and the fastest speed on all types of cancer pathology image datasets,' but the evidence for the accuracy claim is statistically weak, the supervision protocol for public datasets is ambiguous, and the marker mining is performed in-sample without external validation. Given that the marker discovery is a headline contribution, I would encourage the editor to require either a validation on an independent cohort or a careful restatement of the claims as hypothesis-generating rather than validated findings. The manuscript is an extension of the authors' IJCAI 2024 work (ref [29]); the added contribution of the prototype vocabulary and marker analysis should be clearly delineated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"SEW is a coherent two-branch WSI pipeline: thumbnail-level superpixel graph classification, a trainable focus predictor that selects regions for a higher-magnification local branch, and a consistency loss tying the two together. The speed numbers are the most concrete contribution—5-10 seconds per slide with accuracy competitive on seven datasets. The ablations are clean and show each component (local branch, focus predictor, consistency constraint) helps. I’d believe the core method works.\n\nThe problems are concentrated in the comparisons and the supervision protocol. The paper never compares against ref [29], the authors’ own self-reform critical region focusing method, which is the closest prior work. That omission is hard to justify, especially since ref [29] was on the same tasks and by the same group. The stress-test note is on target: §3.3.3 defines the local branch supervision through lesion-area masks or region labels, but §4.1 only confirms such annotations for the three in-house datasets. For CAMELYON16, PANDA, BRCA, and LUAD, the paper does not state where y_t_gt comes from. If masks are broadcast or silently derived, the comparison with weakly supervised MIL baselines is unfair and the self-calibration training is not what is described. This needs to be stated explicitly, dataset by dataset.\n\nThe tumor marker mining is suggestive, not decisive. The model is trained on the CRC prognosis labels and then features from the same cohort are clustered to find prognosis-specific clusters. That’s hypothesis generation, not validation; an external cohort is needed before calling these “novel tumor markers.” Also, the gains over the best baselines are often within one standard deviation, and no significance tests are reported.\n\nStill, the framework is sensible and the core claim is plausible. The paper deserves peer review, not desk rejection. A referee should ask for a comparison with ref [29], a precise mask-supervision table, significance testing, and external marker validation. I’d want to see a revised version, but the idea is worth engaging with.","headline":"A genuinely fast and coherent WSI pipeline whose SOTA claim is currently under-supported by a missing baseline comparison and underspecified mask supervision.","tokens_in":13095,"tokens_out":4133,"would_cite":false,"duration_ms":38972,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SEW reports the highest accuracy and fastest inference on six cancer whole-slide datasets, and surfaces three colorectal prognostic markers from learned features.","keywords":["whole slide image analysis","self-calibration","superpixel graph","focus predictor","pathological prototype vocabulary","tumor marker mining","colorectal cancer","pathology grading"],"falsifier":"Run SEW on CAMELYON16 and PANDA with only slide-level labels, disabling the lesion-mask supervision of the local branch; if accuracy drops to or below the CLAM or ZoomMIL baselines, the reported gains depend on undocumented mask annotations rather than on self-calibration alone.","tokens_in":12079,"feed_emoji":"🔬","tokens_out":8377,"duration_ms":66183,"temperature":0.7,"pith_summary":"SEW is a framework for analyzing gigapixel whole slide pathology images that tries to get both global structure and local detail without paying the computational cost of pyramid features or whole-slide patch MIL. It first classifies a thumbnail through a superpixel graph and transformer, then a focus predictor selects the most suspicious regions, and a detailed branch re-examines those regions at higher magnification; a consistency constraint feeds local detail back into the global branch. The paper reports the highest accuracy and fastest speed across six datasets (CAMELYON16, PANDA, BRCA, LUAD, and in-house HCC, GC, CRC), at 5.44–10.97 seconds per slide. On a colorectal prognosis cohort, k-means clustering of focused features and prototype-based tissue reconstruction lead to two verified novel markers (mucinous lakes, necrosis within glands) and a tissue-infiltration marker. The reason to care: if these results hold, accurate WSI grading and prognosis no longer require hours of compute, and AI can propose candidate tumor markers for pathologists to verify.","feed_headline":"SEW pipeline is fastest and most accurate on six cancer slide datasets","feed_subtitle":"Global thumbnail scan plus focused zoom finds lesions and cuts inference to seconds.","key_machinery":"The load-bearing mechanism is a self-calibration loop built on three learned components. A superpixel graph (SLIC) converts the thumbnail into about 1024 nodes with color histograms plus spatial positions; a 3-layer GCN aggregates neighbor features, then a 12-layer cross-attention transformer with a classification token classifies the slide. The focus predictor reads the transformer's last-layer node tokens and predicts a lesion heatmap, trained first on Grad-CAM pseudo-labels and later on local-branch outputs; top-K non-overlapping subgraphs are zoomed for the local branch. The local branch builds finer superpixel graphs, applies intra-group and inter-group cross-attention with class tokens per group, and is supervised by lesion-area masks. The consistency constraint $L_{cst} = D_{KL}(W_{cls}^{proj} CLS_{local}^t \\| h'_n)$ forces the global node feature to match the local class token, so the global branch learns where to look. Finally, k-means over all local node features forms a pathological prototype vocabulary that ties same-semantic tissues across slides and enables whole-slide reconstruction for spatial marker mining.","core_discovery":"The central claim, stated as the authors would state it, is that SEW—a three-component pipeline of a global superpixel-graph thumbnail classifier, a focus predictor that locates high-probability lesion regions, and a local branch that classifies magnified top-K subgraphs—simultaneously solves the global-vs-local and speed-vs-accuracy problems in WSI analysis. The global branch gives structural context; the focus predictor narrows attention; the local branch supplies cellular-level validation; the KL consistency constraint makes the global branch extract features aligned with the local detail. The paper asserts this pipeline achieves the highest accuracy and the fastest speed on all six evaluated datasets (e.g., CAMELYON16 85.69% at 5.44s versus HIPT 85.57% at 335.74s), and that the focused discriminative features, clustered with k-means and reconstructed through a pathological prototype vocabulary, reveal two novel colorectal cancer markers—mucinous lakes and necrosis within glands—and a third spatial marker, the degree of tumor infiltration.","pith_inferences":["Editorial inference: the same global-focus-local self-calibration pattern could transfer to other gigapixel domains such as satellite or whole-organ imaging, where global context and local detail both matter; the paper's contribution is the mechanism, not just the medical result.","Editorial inference: the novel-marker claims rest on a 100-case CRC subset and pathologist verification; a prospective, blinded study on independent cohorts would be needed to confirm prognostic value, because cluster uniqueness in a single retrospective set can reflect cohort-specific artifacts.","Editorial inference: because mask supervision is documented only for in-house datasets, a clean testable extension is fully weakly-supervised SEW (slide labels only) on CAMELYON16 and PANDA; the gap between that and the reported numbers would quantify how much of the gain is genuinely self-calibration versus lesion-mask supervision.","Editorial inference: the pathological prototype vocabulary could be reused as a tissue-level dictionary for cross-slide registration, stain normalization, or content-based retrieval, beyond its current role in classification and reconstruction."],"forward_implications":["Whole-slide grading and prognosis can run in 5–11 seconds per slide on a single GPU, two orders of magnitude faster than pyramid- or graph-based methods, with equal or better accuracy.","The focus predictor plus local validation makes attention inspectable: heatmaps and top-K regions give a built-in explanation of each decision, useful for clinical review.","The pretrained features transfer across cancer types: fine-tuning from HCC weights converges in about half the epochs and stays within 0.07–0.64% of training from scratch, suggesting reusable pathology representations.","Superpixel-based graph nodes beat fixed 16x16 patches for this task: on HCC and CAMELYON16, superpixel graphs improve accuracy by roughly 2–4.5 points and AUC by 0.05–0.06.","If the CRC marker results hold, SEW gives pathologists a small candidate set from feature clusters rather than an overwhelming patch pool, shortening the loop from data to marker hypothesis."],"supporting_citations":[{"why":"Provides SLIC superpixel segmentation used to build the global and local graph nodes.","marker":"[1]"},{"why":"Grad-CAM heatmaps initialize the focus predictor and break the cold-start problem.","marker":"[18]"},{"why":"PANDA challenge dataset is one of the public benchmarks where SEW reports 82.62% accuracy at 1.85 s per slide.","marker":"[4]"},{"why":"CAMELYON16 dataset is the public metastasis benchmark where SEW reports 85.69% accuracy at 5.44 s per slide.","marker":"[14]"},{"why":"CLAM is the attention-based weakly supervised MIL baseline whose accuracy and speed SEW compares against on all datasets.","marker":"[15]"},{"why":"HIPT is the hierarchical transformer baseline that achieves near-SEW accuracy but needs hundreds of seconds per slide.","marker":"[6]"},{"why":"Patch-GCN is the graph-based survival and grading baseline that motivates the superpixel graph design and the patch-versus-superpixel ablation.","marker":"[5]"},{"why":"ZoomMIL is the differentiable zooming MIL baseline that targets speed, setting the bar for per-slide inference time.","marker":"[24]"}],"fun_headline_variants":["SEW: fastest, most accurate on six slide datasets","Self-calibration pipeline aces six cancer datasets","SEW: global scan + focused zoom = top accuracy, speed","SEW sets records on six slide datasets, reveals markers","Self-calibration yields rapid, precise slide analysis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The local branch is trained on lesion-area masks or region-level labels, but the paper only says those masks exist for the three in-house datasets; for the four public datasets, mask availability is not described.","fun_headline_variants_meta":{"raw":{"variants":["SEW: fastest, most accurate on six slide datasets","Self-calibration pipeline aces six cancer datasets","SEW: global scan + focused zoom = top accuracy, speed","SEW sets records on six slide datasets, reveals markers","Self-calibration yields rapid, precise slide analysis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001549,"raw_usage":{"total_tokens":6188,"prompt_tokens":934,"completion_tokens":5254,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":5175}},"tokens_in":550,"tokens_out":5254,"duration_ms":36962,"temperature":1.0,"reasoning_tokens":5175,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:32:45.543680+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SEW on CAMELYON16 and PANDA with only slide-level labels, disabling the lesion-mask supervision of the local branch; if accuracy drops to or below the CLAM or ZoomMIL baselines, the reported gains depend on undocumented mask annotations rather than on self-calibration alone.","supporting_citations":[{"cited_title":"Slic superpixels","cited_arxiv_id":null,"evidence_quote":"Provides SLIC superpixel segmentation used to build the global and local graph nodes."},{"cited_title":"Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra","cited_arxiv_id":null,"evidence_quote":"Grad-CAM heatmaps initialize the focus predictor and break the cold-start problem."},{"cited_title":"Artificial intelligence for diagnosis and gleason grading of prostate cancer: the panda challenge","cited_arxiv_id":null,"evidence_quote":"PANDA challenge dataset is one of the public benchmarks where SEW reports 82.62% accuracy at 1.85 s per slide."},{"cited_title":"1399 h&e-stained sentinel lymph node sections of breast cancer patients: the camelyon dataset","cited_arxiv_id":null,"evidence_quote":"CAMELYON16 dataset is the public metastasis benchmark where SEW reports 85.69% accuracy at 5.44 s per slide."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CLAM is the attention-based weakly supervised MIL baseline whose accuracy and speed SEW compares against on all datasets."},{"cited_title":"Scaling vision transformers to gigapixel images via hierarchical self-supervised learning","cited_arxiv_id":null,"evidence_quote":"HIPT is the hierarchical transformer baseline that achieves near-SEW accuracy but needs hundreds of seconds per slide."},{"cited_title":"Whole slide images are 2d point clouds: Context-aware survival prediction using patch-based graph convolutional networks","cited_arxiv_id":null,"evidence_quote":"Patch-GCN is the graph-based survival and grading baseline that motivates the superpixel graph design and the patch-versus-superpixel ablation."},{"cited_title":"Differen- tiable zooming for multiple instance learning on whole-slide images","cited_arxiv_id":null,"evidence_quote":"ZoomMIL is the differentiable zooming MIL baseline that targets speed, setting the bar for per-slide inference time."}],"review_version":1}