{"id":"5dcfded7-3dd0-48ff-90ce-8e2eb102fb1e","arxiv_id":"2508.15904","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PathPT improves few-shot rare cancer subtyping by using zero-shot vision-language models to create tile-level pseudo-labels and learning prompt tokens with spatial context, outperforming standard MIL baselines when the backbone has strong zero-shot grounding.","lead":"This paper presents PathPT, a method that fine-tunes pathology vision-language AI models for rare cancer subtyping using only a few labeled whole slide images. It generates tile-level training signals from slide-level labels and showed higher accuracy and tumor localization than standard multiple-instance learning on several rare and pediatric cancer datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Few-shot protocol ambiguity: manual prompt selection (Sec 4.4) may run on the full 15-WSI training set, not the k-shot subset, giving PathPT access to more labels than MIL baselines.","rationale":"The reader's weakest assumption is that tile-level pseudo-labels are accurate enough to supervise training. That is a real concern, but it does not directly threaten the headline KEEP result: even if pseudo-labels are noisy, PathPT-KEEP could still outperform MIL on EBRAINS. The prompt-selection ambiguity is more load-bearing because it questions whether the KEEP result is a genuine few-shot result at all. If the full 15-WSI training set is used to choose prompts, PathPT sees more labels than the MIL baselines, and the comparison is confounded. A single code inspection and one re-run with prompt selection restricted to the k-shot subset would settle this. The paper's own results already show that PathPT is not consistently superior across weaker backbones, so the abstract's 'consistently delivers superior performance' needs qualification regardless; my concern is about the strongest positive result, not the overclaim. I therefore maintain the reader's CONDITIONAL verdict rather than moving it.","tokens_in":31698,"tokens_out":8428,"duration_ms":97286,"concrete_test":"Check the released code (github.com/MAGIC-AI4Med/PathPT) for the prompt-selection step. Specifically, verify whether the ranking in Sec 4.4 iterates over all 15 training WSIs per subtype or only the WSIs sampled for the current k-shot run. Then rerun the EBRAINS 10-shot PathPT-KEEP experiment with prompt selection restricted to the 10 sampled WSIs per class (same folds) and compare balanced accuracy to DGRMIL's 0.650. If PathPT no longer exceeds the best MIL baseline, the central claim is not supported; if performance is unchanged, the concern is dismissed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's few-shot evaluation may give PathPT access to more labeled data than the MIL baselines. Section 2.2 fixes 15 WSIs per subtype as the training set, then samples 1/5/10 WSIs for few-shot training. Section 4.4 (Manual Prompt Selection) says 200 prompt groups are ranked by WSI classification performance 'on the training dataset' and the top 100 are selected. If 'training dataset' means the full 15-WSI set rather than the current k-shot subset, then prompt selection uses all 15 labeled WSIs per class, while MIL baselines are trained only on the k-shot subset. This would inflate PathPT's balanced accuracy and could explain the KEEP margin (0.679 vs 0.650) and the collapse on weak backbones (the selected prompts may be tuned to the full training set). The paper never states that prompt selection is restricted to the k-shot sampled WSIs, and the reported gain is therefore not yet established as a true few-shot result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"PathPT is a prompt-tuning framework for pathology foundation models that replaces slide-level MIL aggregation with tile-level classification, using frozen vision/text encoders, a spatial-aware aggregator, learned prompts, and pseudo-labels derived from zero-shot VL predictions filtered by WSI labels. The paper benchmarks PathPT against ABMIL, CLAM, TransMIL, and DGRMIL on eight rare-cancer and three common-cancer subtyping datasets with 1/5/10-shot settings, and on three segmentation datasets. The headline result is PathPT-KEEP reaching 0.679 balanced accuracy on EBRAINS 10-shot; the paper claims consistent superiority and improved tumor-region grounding. The manuscript includes code and detailed supplementary tables.","tokens_in":31996,"tokens_out":4578,"duration_ms":44554,"significance":"If the few-shot comparison is valid, the method is a useful adaptation recipe for VL pathology models in low-data settings: it preserves tile-level interpretability, is parameter-efficient, and ships code. The curated pediatric benchmarks and comparisons across four backbones/four MIL baselines are valuable. However, the paper's own tables show the method only consistently outperforms MIL with the KEEP backbone; with PLIP, MUSK, and CONCH it often underperforms. The grounding DICE numbers are computed on training-set annotations, which weakens the interpretability claim. These issues are fixable by re-analysis and tempering claims.","major_comments":[{"comment":"The 200 prompt groups are ranked by WSI classification performance 'on the training dataset.' The protocol in Section 2.2 defines a 15-WSI-per-subtype training split from which 1/5/10 WSIs are sampled for few-shot training. If prompt selection uses the full 15-WSI set, PathPT's tile pseudo-labels are generated with prompt groups chosen using labels that the MIL baselines never see. The paper never states that selection is restricted to the k-shot subset. Please clarify and re-run if necessary; this is essential for the EBRAINS KEEP margin (Table S1: 0.679 vs 0.650) to be a valid few-shot comparison.","section":"Section 4.4, Manual Prompt Selection"},{"comment":"The claim of 'consistently superior performance' is contradicted by the paper's own results. On EBRAINS with PLIP, PathPT 10-shot is 0.251 vs TransMIL 0.488 and DGRMIL 0.491 (Table S1); with CONCH 10-shot PathPT 0.491 vs TransMIL/DGRMIL 0.621. On TCGA-THYM PLIP 10-shot PathPT 0.430 vs TransMIL 0.683 (Table S3). PathPT is best mainly with KEEP, and is sometimes competitive with CONCH. Please restrict the 'superior' claim to the configurations where it holds and report win/loss counts with uncertainty.","section":"Section 2.2 and Supplementary Tables S1-S11"},{"comment":"The grounding DICE scores are computed on pathologist annotations of 'each WSI in the few-shot training set.' This evaluates the model's fit to training data, not its ability to localize unseen tumors. Since interpretability and grounding are central contributions, report DICE on held-out WSIs (e.g., test split of classification benchmarks) or explicitly state these are training-set localization results; the current wording implies a generalization claim.","section":"Sections 2.2/2.3, Figures 2c/3b"},{"comment":"'PathPT consistently improved segmentation performance as the number of training samples increased' is not monotonic in the supplementary data: e.g., KEEP/CAMELYON16 DICE is 0.591 (5-shot) then 0.560 (10-shot); CONCH/CAMELYON16 is 0.522 then 0.505; MUSK/CAMELYON16 is 0.504 then 0.459. Please either correct the claim or analyze the non-monotonicity.","section":"Section 2.5 vs Table S12"}],"minor_comments":[{"comment":"Abstract says 2,910 WSIs and Introduction says 2,920; also pediatric total text says 1,232 but Table S16 sums to 1,229. Please harmonize.","section":"Abstract/Introduction/Section 2.3"},{"comment":"The selection of 'the top-performing 100 groups' is a hyperparameter; clarify whether it is per-dataset or per-split and how it interacts with the 10 repeats.","section":"Section 4.4"},{"comment":"Typos: 'detialed' (Section 2.2), 'lanaguage' (Section 2.3), 'vision-lanaguage' (Section 2.3). Supplementary Figure S2 caption says 'Rare adult cancer subtyping results' but text refers to pediatric; likely caption error.","section":"Throughout"},{"comment":"No statistical comparison is reported for the main subtyping tables; the segmentation section uses paired t-tests (Figure 5c). Add variance-aware comparisons or caveats, especially since some differences between methods are small relative to the reported interquartile ranges.","section":"Main subtyping tables"}],"recommendation":"major_revision","confidential_remarks":"The abstract and Section 3 claims are substantially stronger than the supplementary tables support. I do not see evidence of intentional misreporting, but the protocol ambiguity in prompt selection must be resolved; this is a legitimate correctness-risk issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper is worth reading for the benchmark and for the KEEP results. Second, the central claim of \"consistently superior performance\" doesn't survive contact with the paper's own tables, and there is an ambiguity in the few-shot protocol that could give PathPT more labeled data than the baselines.\n\nWhat's actually new: a new benchmark suite for rare cancer subtyping, including an internal pediatric cohort, eight rare datasets across 56 subtypes; systematic comparison of four VL backbones and four MIL baselines; 10 repeats with quartiles; an ablation against CoOp and linear probing; and code on GitHub. The method is a reasonable assembly of existing pieces — CoOp-style prompts, spatial aggregation, zero-shot pseudo-labels — but the assembly is sensible and the segmentation results show consistent gains. The pediatric cohort alone is a contribution.\n\nSoft spots. The abstract says \"consistently superior,\" but Supplementary Tables S1–S11 show PathPT underperforming TransMIL/DGRMIL with PLIP, MUSK, and CONCH on many datasets and shot settings. For example, EBRAINS 10-shot with CONCH: PathPT 0.491 vs TransMIL 0.621. With KEEP it wins most, but not all. The \"consistently\" is inaccurate.\n\nMore serious: the manual prompt selection in Section 4.4 ranks 200 prompt groups by WSI classification performance \"on the training dataset.\" The paper earlier defines a 15-WSI training set per subtype and then samples k-shot subsets from it. If prompt selection runs on the full 15, PathPT's chosen prompts are tuned on all 15 labeled slides while MIL baselines see only k. That would break the few-shot comparison. The paper never states that prompt selection is restricted to the k-shot subset. This needs to be fixed before the few-shot claim is established. It could be a wording oversight, but it is load-bearing.\n\nGrounding: the tumor-region DICE in Figures 2c/3b is computed on the few-shot training set with pathologist masks. That measures fit, not generalization. The held-out segmentation benchmarks (CAMELYON16, PANDA, AGGC22) are better evidence, but those use mask supervision at tile level, so they don't directly validate the slide-label-based grounding on test WSIs. Also, the pseudo-label loss is enabled only for CONCH and KEEP; that is disclosed but means the method is really a per-backbone selection.\n\nWho this is for: computational pathology researchers and clinicians working on rare cancers, especially pediatric. The benchmark and the KEEP results justify a serious referee. The review should focus on the prompt-selection protocol and the overclaim. I'd send it to review, not desk reject.","headline":"PathPT is a useful few-shot recipe for VL-based WSI subtyping with a genuinely valuable pediatric benchmark, but the consistency claim is overblown and the few-shot protocol has an unresolved label-leakage ambiguity.","tokens_in":32448,"tokens_out":3453,"would_cite":true,"duration_ms":36852,"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":"PathPT converts weak slide-level labels into tile-level supervision for rare cancer, beating four MIL baselines in few-shot subtyping.","keywords":["rare cancer subtyping","few-shot learning","vision-language foundation models","prompt tuning","multiple instance learning","whole slide images","tile-level supervision","tumor region grounding"],"falsifier":"Find or construct a backbone whose zero-shot tile DICE is at chance but whose features still allow MIL gains; if PathPT with that backbone still exceeds MIL, the pseudo-label explanation fails. Conversely, the paper's own PLIP results on TCGA datasets already show the predicted failure mode, offering a controlled case: applying PathPT with perfect ground-truth tile labels should restore the gains if the mechanism is as stated.","tokens_in":31653,"feed_emoji":"🔬","tokens_out":4586,"duration_ms":47176,"temperature":0.7,"pith_summary":"PathPT claims that vision-language pathology foundation models can be adapted to rare cancer subtyping under extreme data scarcity by treating whole-slide labels as a source of tile-level training signals rather than as slide-level classification targets. The framework freezes the vision and text encoders, adds a spatially-aware aggregator and learnable textual prompts, and uses zero-shot tile predictions to generate pseudo-labels that keep only tiles predicted normal or matching the slide's subtype. On eight rare cancer datasets (56 subtypes, roughly 2,900 WSIs) and three common cancer datasets, PathPT with the KEEP backbone reports consistent gains over four MIL baselines, including 0.679 balanced accuracy on the 30-subtype EBRAINS benchmark at 10 shots per subtype, along with improved tumor-region grounding.","feed_headline":"Prompt-tuned model maps rare cancer subtypes from 10 slides","feed_subtitle":"PathPT turns slide-level labels into tile-level tumor maps and beats MIL on 8 rare-cancer datasets.","key_machinery":"Three components carry PathPT. (1) Spatially-aware visual aggregation: parallel 3x3/5x5/7x7 residual convolutions capture local tile interactions and a transformer self-attention layer captures global context, applied to spatially ordered tile features. (2) Task-adaptive prompt tuning: K learnable tokens per class are prepended to the frozen text encoder (a CoOp-style context prompt), so the model learns class-specific text embeddings aligned with histopathological semantics. (3) Tile-level supervision from slide labels: zero-shot VL predictions assign pseudo-labels; tiles whose prediction is normal or matches the WSI label are kept as training signal, and an unlabeled-tile loss pushes remai","core_discovery":"The central claim is that the bottleneck for few-shot rare-cancer subtyping is not feature aggregation but semantic alignment: conventional MIL treats VL models as frozen feature extractors and trains attention over visual features only, while PathPT shows that end-to-end prompt tuning over the frozen text encoder, combined with spatially-aware visual context, converts coarse slide-level labels into reliable tile-level supervision. The paper demonstrates this by benchmarking four VL backbones and four MIL frameworks, finding that gains track zero-shot grounding quality of the backbone and that PathPT outperforms all MIL baselines when paired with KEEP, on both classification and segmentation","pith_inferences":["Because the selective pseudo-labeling acts as a noise filter (keeping only normal or label-consistent tiles), PathPT could serve as a general recipe for weakly-supervised tile classification in any domain where a zero-shot CLIP-like model provides partial grounding—e.g., histology beyond cancer, or medical imaging with coarse labels.","The method's failure mode with PLIP suggests a testable extension: explicitly model pseudo-label confidence or use multiple prompt-sets to detect unreliable backbones, potentially making the framework robust regardless of zero-shot quality.","One could test whether the gains come mostly from the spatial module or the prompts by ablating each on a backbone with strong grounding; the paper's ablation on segmentation suggests the combination matters, but a classification-only ablation across backbones would sharpen this.","If KEEP's advantage is its disease-knowledge injection, then coupling PathPT with knowledge-augmented backbones may be a general booster for rare-disease tasks—an indirect consequence the paper hints at but does not claim."],"forward_implications":["If PathPT's claims hold, few-shot rare-cancer subtyping can be done with 10 or fewer annotated slides per subtype, which is clinically meaningful for pediatric and rare adult tumors where expert annotations are scarce.","The approach yields tile-level predictions at inference, so the same model outputs both subtype and a spatial tumor map, giving pathologists a directly checkable localization rather than an opaque slide-level score.","Parameter-efficient adaptation (only prompt tokens and a lightweight aggregator are trained) makes it feasible to re-adapt systems when new subtypes are added, since the encoders stay frozen.","The reported dependence on zero-shot grounding quality means the method's ceiling is set by the backbone's ability to separate tumor from normal tissue; improvements in VL alignment should translate directly into PathPT gains."],"supporting_citations":[{"why":"Supplies the KEEP backbone whose strong zero-shot tile grounding powers PathPT's pseudo-labels and the best reported results.","marker":"[39]"},{"why":"CONCH provides a comparison backbone and the manual prompt templates that PathPT's prompt selection averages over.","marker":"[16]"},{"why":"CoOp is the source of the learnable-context-prompt mechanism that PathPT adapts for class-specific prompts.","marker":"[38]"},{"why":"EBRAINS is the 30-subtype brain-tumor benchmark where PathPT–KEEP achieves 0.679 balanced accuracy in 10-shot.","marker":"[24]"},{"why":"ABMIL is one of the four MIL baselines that PathPT is compared against.","marker":"[15]"},{"why":"TransMIL is a MIL baseline that captures long-range dependencies with self-attention, among the strongest MIL competitors.","marker":"[26]"},{"why":"DGRMIL is a MIL baseline using cross-attention and global vectors; it often matches PathPT at higher shot counts, defining the comparison bar.","marker":"[41]"}],"fun_headline_variants":["Few-shot prompt tuning lifts rare cancer subtyping accuracy","PathPT: prompt-tuned VL maps rare cancer from 10 slides","Prompt tuning beats MIL for rare cancer subtyping","Spatially-aware prompt tuning improves rare cancer diagnosis"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"PathPT's training signal is only as good as the zero-shot tile-level predictions of the vision-language backbone; if the backbone cannot distinguish tumor from normal at tile level, the pseudo-labels mislead training and PathPT falls behind MIL baselines.","fun_headline_variants_meta":{"raw":{"variants":["Few-shot prompt tuning lifts rare cancer subtyping accuracy","PathPT: prompt-tuned VL maps rare cancer from 10 slides","Prompt tuning beats MIL for rare cancer subtyping","Spatially-aware prompt tuning improves rare cancer diagnosis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1285,"prompt_tokens":768,"completion_tokens":517,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":449}},"tokens_in":512,"tokens_out":517,"duration_ms":5827,"temperature":1.0,"reasoning_tokens":449,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:39:50.920341+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find or construct a backbone whose zero-shot tile DICE is at chance but whose features still allow MIL gains; if PathPT with that backbone still exceeds MIL, the pseudo-label explanation fails. Conversely, the paper's own PLIP results on TCGA datasets already show the predicted failure mode, offering a controlled case: applying PathPT with perfect ground-truth tile labels should restore the gains if the mechanism is as stated.","supporting_citations":[{"cited_title":"A knowledge-enhanced pathology vision-language foundation model for cancer diagnosis.arXiv preprint arXiv:2412.13126, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies the KEEP backbone whose strong zero-shot tile grounding powers PathPT's pseudo-labels and the best reported results."},{"cited_title":"A visual-language foundation model for computational pathology.Nature Medicine, 30(3):863–874, 2024","cited_arxiv_id":null,"evidence_quote":"CONCH provides a comparison backbone and the manual prompt templates that PathPT's prompt selection averages over."},{"cited_title":"The digital brain tumour atlas, an open histopathology resource.Scientific Data, 9(1):55, 2022","cited_arxiv_id":null,"evidence_quote":"EBRAINS is the 30-subtype brain-tumor benchmark where PathPT–KEEP achieves 0.679 balanced accuracy in 10-shot."},{"cited_title":"Transmil: Transformer based correlated multiple instance learning for whole slide image classification.Advances in Neural Information Processing Systems, 34:2136–2147, 2021","cited_arxiv_id":null,"evidence_quote":"TransMIL is a MIL baseline that captures long-range dependencies with self-attention, among the strongest MIL competitors."},{"cited_title":"DGR-MIL: Exploring Diverse Global Representation in Multiple Instance Learning for Whole Slide Image Classification","cited_arxiv_id":"2407.03575","evidence_quote":"DGRMIL is a MIL baseline using cross-attention and global vectors; it often matches PathPT at higher shot counts, defining the comparison bar."}],"review_version":1}