{"id":"9e955bb8-05b0-4138-9101-6986b9adbbb9","arxiv_id":"2411.14743","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"FOCUS improves few-shot whole slide image classification by progressively removing redundant patches guided by pathology foundation model features and language prompts, then aggregating the surviving tokens with cross-modal attention.","lead":"FOCUS is a method for classifying whole slide cancer images from just a few labeled slides, by first compressing each slide down to its most diagnostic patches using pathology foundation models and language-based descriptions. It could reduce how much expert annotation is needed for AI cancer diagnosis, a step toward practical data-scarce pathology tools.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline 4-shot gains are within roughly one standard error of zero; without paired significance tests the 'consistently outperforms' claim is not established.","rationale":"I focus on statistical reliability rather than the reader's chosen weakest assumption. The central claim is an empirical performance claim, and the numbers in Table 1 are the only support; the paper's own reported variance makes the headline differences indistinguishable from noise under a straightforward unpaired reading. The reader's concern about cosine-similarity-based compression is real and would benefit from a random-token-selection control, but it does not directly decide whether the method beats SOTA. A significance test does. I also note the internal inconsistency in Eq. (6): with increasing θ_i, the condition min(s)<θ_i becomes easier to satisfy, so mask_j=1 is assigned more often and the stage retains more tokens, contradicting the text's claim of increasing filtering intensity. I did not make this the headline because it may be a documentation/code mismatch, but it should be checked in the released code. Overall the verdict stays CONDITIONAL: the method is plausible and the ablations are encouraging, but the headline performance gap requires significance testing before the claim of consistent superiority is accepted.","tokens_in":19409,"tokens_out":8997,"duration_ms":99710,"concrete_test":"Obtain the per-fold Balanced ACC arrays for FOCUS and the best baseline in the 4-shot rows of Table 1 for all three datasets and run a paired permutation test (10,000 permutations) on the difference, reporting p-values and bootstrap 95% confidence intervals for each dataset. If any p≥0.05 or any confidence interval contains zero, the 'consistently outperforms' claim is not supported by the current evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 claims absolute Balanced ACC gains of 1.2%, 4.3%, and 3.3% over the second-best methods in the 4-shot setting. The reported ten-fold standard deviations make these differences statistically fragile: on TCGA-NSCLC FOCUS is 0.819±0.044 vs ViLa-MIL 0.807±0.043 (difference 0.012); on CAMELYON it is 0.701±0.082 vs 0.658±0.036 (difference 0.043); on UBC-OCEAN it is 0.704±0.088 vs 0.671±0.080 (difference 0.033). Each headline difference is only about 0.6–1.5 standard errors under an unpaired comparison, and the paper reports no paired significance test, confidence interval, or effect size. Because the evaluation uses only ten folds with random K-shot sampling, the variance is high enough that the central empirical claim could survive only if the improvements are statistically reliable. The reader's mechanism concern about cosine-similarity-based redundancy is legitimate, but it is secondary: even a correctly functioning compression module would not support 'superior performance' if the headline margins are noise.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FOCUS, a three-stage adaptive visual compression framework for few-shot whole-slide image classification. Using frozen pathology foundation-model features and LLM-generated text prompts, the method first discards visually redundant patches by sliding-window cosine similarity, then selects patches by cross-modal relevance to language prompts, then removes locally similar neighboring tokens with increasing thresholds, and finally aggregates the compressed tokens with a cross-modal attention head. Experiments on TCGA-NSCLC, CAMELYON, and UBC-OCEAN under 4/8/16-shot settings compare FOCUS with MIL and few-shot baselines, reporting gains in Balanced ACC, AUC, and F1, with ablations for the compression modules, foundation-model choice, and prompt source.","tokens_in":19630,"tokens_out":3048,"duration_ms":33243,"significance":"If the empirical claims were statistically reliable, FOCUS would make a modest but useful contribution: it is a simple, largely training-free compression strategy that leverages existing pathology foundation models and external language priors without requiring paired multi-resolution inputs or reference samples. The authors also release code, which supports reproducibility. However, the paper's central claim—that FOCUS consistently outperforms state-of-the-art few-shot MIL methods—is currently supported only by point estimates with high variance and no significance testing. The value of the method therefore depends on additional evidence that the reported margins are not noise, and on a protocol that prevents large labeled validation sets from driving few-shot model selection.","major_comments":[{"comment":"The headline 4-shot Balanced ACC improvements are statistically fragile and the claim of consistent superiority is not established. In TCGA-NSCLC, FOCUS is 0.819±0.044 versus ViLa-MIL 0.807±0.043 (difference 0.012); in CAMELYON, 0.701±0.082 versus 0.658±0.036 (difference 0.043); in UBC-OCEAN, 0.704±0.088 versus 0.671±0.080 (difference 0.033). Under an unpaired comparison with ten folds, these differences correspond to roughly 0.6, 1.5, and 0.9 standard errors, respectively. The paper reports no paired significance tests, confidence intervals, or effect sizes, although the folds are presumably paired across methods. Either add paired statistical tests with multiple-comparison correction, or soften the 'consistently outperforms' claim to reflect that the observed differences are not statistically significant at the reported level of variance.","section":"Section 4.2, Table 1"},{"comment":"The evaluation protocol uses a validation set with many labels to select the best checkpoint (early stopping based on validation performance) before reporting test results. With a 6:2:2 split, the validation set contains far more labeled slides than the K-shot training set (e.g., for TCGA-NSCLC, about 200 validation slides versus 4 per class in 4-shot). In a few-shot evaluation, this large labeled validation set can act as an additional source of supervision for model selection, potentially inflating the reported few-shot performance. The authors should clarify whether all baselines are selected under the same rule, and should report results obtained without validation-based checkpoint selection, or with a validation set matched to the shot count.","section":"Section 4.1, Implementation Details"},{"comment":"The paper does not validate the key assumption that cosine similarity between patch features reliably indicates diagnostic redundancy. The compression stages discard patches whose mean or neighbor similarity exceeds thresholds (Eq. 2 and Eq. 6), but no experiment compares this selection against random token dropping, feature-only selection without language guidance, or frequency-based selection. Table 2 shows that adding each module improves Balanced ACC, but the increments are small relative to the reported standard deviations and no significance tests are provided. Consequently, the attribution of the gains to language-guided prioritization rather than to mere compression is not demonstrated. Please add ablations that isolate the effect of the cosine-similarity redundancy proxy and the language guidance component, along with sensitivity analysis for the hand-set hyperparameters w=32, gamma=0.8, and theta_base=0.7.","section":"Sections 3.2.1 and 3.3"}],"minor_comments":[{"comment":"The first contribution bullet contains a grammatical error: 'We explores the untapped potential' should read 'We explore'. Also, 'utlize' in Section 4.1 should be 'utilize'.","section":"Introduction, Contributions"},{"comment":"Algorithm 1 is titled 'LA ViC-MIL', which does not match the method name FOCUS used throughout the paper; please correct the title and any internal references for consistency.","section":"Supplementary Material, Algorithm 1"},{"comment":"The explanations for why one LLM prompt outperforms another (e.g., 'comprehensive and clinically precise descriptions' versus 'overly brief descriptions') are post-hoc and unsupported by quantitative analysis; the appendix presents these as established causes rather than as hypotheses.","section":"Section 4.3.3 and Appendix A.4"},{"comment":"The Related Works section discusses PEMP, MSCPT, and FAST as recent few-shot MIL methods, but none of these methods is included in the experimental comparisons; the claim of outperforming 'state-of-the-art' few-shot methods is therefore limited to the subset of baselines evaluated.","section":"Related Works and Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is reasonable and the code release is commendable, but the empirical evidence is currently too weak for acceptance: the main margins are within one standard error when appropriately assessed, and the validation-based early stopping protocol may artificially advantage the method in the few-shot setting. These issues are fixable in revision—adding paired significance tests, reporting results without validation-based model selection, and adding the requested ablations would substantially strengthen the paper. I would not reject at this stage, but the revision needs to demonstrate that the reported gains are not noise."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a plausible method paper whose headline claim is not actually supported by the numbers as reported. The progressive three-stage compression — FM-based global similarity filtering, language-guided token prioritization, then neighbor-aware sequential compression — is genuinely new relative to TOP-MIL, ViLa-MIL, PEMP, and MSCPT. The paper is also more clinically practical than some prior work because it runs on single-resolution WSIs rather than paired multi-resolution inputs. The ablations on UBC-OCEAN show each module adds a consistent increment, which is the strongest part of the empirical case. The code link and the prompt-generation ablation across LLMs are also good to see.\n\nThe soft spots are about evaluation, not mechanism. The 4-shot margins over the second-best method are 1.2%, 4.3%, and 3.3% Balanced ACC. With ten folds and the reported standard deviations, each difference is roughly 0.6 to 1.5 unpaired standard errors; no paired test, confidence interval, or effect size is reported. I agree with the stress-test note: the headline margins are statistically fragile, so “consistently outperforms” is not established. Early stopping that selects the best checkpoint on a validation set containing many labels also makes the comparison harder to interpret, since the baselines may not get the same privileged selection. A paired significance test across the same ten folds, or better a validation-free protocol, should be the main revision. The lack of a control that ablates the language guidance (e.g. random or feature-only token selection at the same compression rate) is a genuine gap, but secondary: even a correctly functioning compression module needs a statistically reliable margin to carry the paper’s claim.\n\nThe cosine-similarity redundancy concern is legitimate but minor at this stage; a homogeneous tumor region could in principle be compressed away, and the paper does not analyze what gets discarded. That is a follow-up analysis rather than a fatal flaw. The naming inconsistency in Algorithm 1 (“LA ViC-MIL” vs “FOCUS”) is cosmetic. Citation pattern looks fine, with relevant few-shot MIL and pathology FM work cited.\n\nI would send this to peer review rather than desk-reject. The method is new, the supplement is detailed, and the flaws are fixable with a more careful experimental section. A serious referee can push for paired significance testing, a random-selection control, and a clearer statement about validation-set selection. If the margins survive paired testing, this is a useful contribution to few-shot MIL; if not, the architectural story may still stand, but the superior-performance claim should go.","headline":"A well-built three-stage visual compression pipeline for few-shot WSI classification, but the reported 4-shot gains are statistically fragile and need paired significance testing before the SOTA claim is credible.","tokens_in":20185,"tokens_out":2773,"would_cite":false,"duration_ms":29151,"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":"A three-stage adaptive compression strategy that uses pathology foundation models and LLM-generated text prompts lets few-shot whole-slide classifiers beat prior methods on lung, breast, and ovarian cancer datasets.","keywords":["few-shot learning","whole slide image classification","multiple instance learning","visual token compression","pathology foundation models","language-guided prompting","computational pathology","cancer diagnosis"],"falsifier":"Retain the same total number of tokens but replace the language-guided top-$k$ selection with random selection in the 4-shot UBC-OCEAN setting; if balanced accuracy does not drop, the reported gains come from compression rather than from the LLM prompts. Conversely, on mucinous-carcinoma slides with large uniform mucin pools, progressively lowering $\\tau_g$ should show whether similar-but-diagnostic patches are being pruned.","tokens_in":19190,"feed_emoji":"🔬","tokens_out":8670,"duration_ms":69137,"temperature":0.7,"pith_summary":"FOCUS is a few-shot whole-slide-image classifier that tries to solve the data-scarcity problem by aggressively pruning away uninformative patches before classification. The paper's central claim is that a three-stage, adaptive visual compression pipeline—global redundancy removal using pathology foundation-model features, language-guided token prioritization using LLM-generated pathology descriptions, and neighbor-aware sequential filtering—lets a model trained on 4 to 16 labeled slides per cancer class outperform prior few-shot multiple-instance learning methods. On TCGA-NSCLC, CAMELYON, and UBC-OCEAN, it reports the best balanced accuracy and AUC across the three shot settings, with the largest margins at 4 shots: 81.9%, 70.1%, and 70.4% balanced accuracy on the three datasets. If the claim holds, it means that in scarce-data clinical settings, selectively discarding visual content under language guidance is more effective than aligning every patch with a text prompt. The paper also demonstrates through ablations that each compression stage and the choice of foundation model contribute to the gain.","feed_headline":"Adaptive patch pruning lifts few-shot slide diagnosis by 4.3 points","feed_subtitle":"FOCUS prunes redundant slide patches with foundation-model similarity and LLM prompts, improving 4-shot accuracy on three cancers.","key_machinery":"FOCUS's core mechanism is the three-stage progressive visual-token compression pipeline: (1) global redundancy removal via sliding-window cosine similarity with dynamic threshold $\\tau_g=\\mu(S)+\\sigma(S)$; (2) language-guided prioritization using cross-modal attention between learnable-plus-LLM text prompts and patch features, selecting top-$k$ tokens by relevance rank $r_i$; and (3) sequential neighbor-aware compression that drops a token when the minimum cosine similarity to its two neighbors falls under a rising threshold $\\theta_i=\\theta_{base}+i\\Delta\\theta$. A cross-modal multi-head attention aggregator then pools the surviving tokens for prediction. This chain carries the argument because each stage's thresholds control how many patches survive, and the ablation study shows each stage adds accuracy.","core_discovery":"The central discovery is that in few-shot weakly-supervised whole slide classification, the bottleneck is not the aggregator but the enormous number of diagnostically irrelevant patches, and that these can be removed adaptively before aggregation. FOCUS uses a frozen pathology vision-language foundation model (CONCH) to encode patches, computes pairwise cosine similarities inside sliding windows, and discards patches whose mean similarity exceeds the dynamic threshold $\\tau_g = \\mu(S)+\\sigma(S)$. It then ranks the surviving patches by their cross-modal attention relevance to a concatenation of learnable prompts and an LLM-generated pathology knowledge prompt, keeping the top-$k$ where $k=\\min(M_{\\max},\\gamma N')$, and finally removes tokens whose minimum similarity to their spatial neighbors falls below an increasing threshold $\\theta_i=\\theta_{base}+i\\Delta\\theta$. The surviving, spatially coherent tokens are pooled by a text-guided multi-head attention aggregator. The paper reports consistent improvements over prior few-shot MIL methods, notably absolute balanced-accuracy gains of 1.2%, 4.3%, and 3.3% over the second-best method in the 4-shot setting on TCGA-NSCLC, CAMELYON, and UBC-OCEAN, respectively.","pith_inferences":["Nothing in the paper separates the language guidance from the simple act of compression: a comparison against random token pruning with the same retained count would establish whether the LLM descriptions, rather than the compression itself, drive the 4-shot improvements.","The cosine-similarity redundancy proxy suggests an immediate stress test: large homogeneous tumor regions with mutually similar but diagnostically critical patches may be pruned by global or sequential stages, so the method's robustness on mucinous or diffusely growing tumors deserves direct evaluation.","The same recipe—off-the-shelf pretrained encoder similarity pruning, text-prompt reranking, then neighbor-aware thinning—could transfer to other gigapixel or long-document tasks where only a few class labels are available and most input tokens are irrelevant.","A dynamic extension in which the LLM prompt is updated per slide or per compression stage, or in which thresholds are learned rather than set by population statistics, is a natural next step the paper leaves implicit."],"forward_implications":["At 4 shots per class, FOCUS raises balanced accuracy to 81.9% on TCGA-NSCLC, 70.1% on CAMELYON, and 70.4% on UBC-OCEAN, beating the second-best method by 1.2, 4.3, and 3.3 points.","The advantage persists at 8 and 16 shots, with the best AUC scores of 97.2%, 94.3%, and 96.7% at 16 shots across the three datasets.","Each added module helps: on UBC-OCEAN at 4 shots, adding the FM-based global redundancy removal, language-guided prioritization, sequential compression, and cross-modal aggregation moves balanced accuracy from 62.3% to 70.4%.","CONCH, a vision-language foundation model with aligned image-text pretraining, works better as the feature extractor than UNI, GPFM, Virchow, or PLIP in this framework, indicating that cross-modal alignment in the encoder is part of the gain.","The framework operates on standard single-resolution WSIs, avoiding the paired multi-resolution inputs or extra reference slides required by some prior few-shot methods."],"supporting_citations":[{"why":"Supplies CONCH, the pathology vision-language foundation model used to encode all patch features and text prompts, so the whole compression pipeline runs in its embedding space.","marker":"[29]"},{"why":"Introduces the few-shot weakly-supervised learning formulation and the TOP-MIL prompt-learning baseline that FOCUS extends.","marker":"[32]"},{"why":"ViLa-MIL is the strongest prior few-shot vision-language MIL method and the main second-best baseline that FOCUS outperforms.","marker":"[35]"},{"why":"CLAM provides the WSI preprocessing pipeline that segments tissue and crops the 512x512 patches FOCUS consumes.","marker":"[27]"},{"why":"Supplies the CAMELYON16 breast cancer metastasis dataset used in the few-shot evaluation.","marker":"[4]"},{"why":"Supplies the CAMELYON17 dataset used alongside CAMELYON16 in the breast cancer evaluation.","marker":"[3]"},{"why":"Supplies the TCGA-NSCLC lung cancer dataset used for the LUAD/LUSC few-shot experiments.","marker":"[39]"}],"fun_headline_variants":["FOCUS prunes useless patches to sharpen few-shot slide diagnosis","Adaptive compression lifts few-shot cancer diagnosis by 4.3 points","Knowledge-guided patch selection sharpens few-shot pathology","FOCUS: Cut redundant patches, keep the signal in few-shot slides"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that visual redundancy in a whole slide can be read off cosine similarity between patch features, so a large, homogeneous, but diagnostically important tumor region, consisting of many mutually similar patches, could be pruned before it contributes to classification.","fun_headline_variants_meta":{"raw":{"variants":["FOCUS prunes useless patches to sharpen few-shot slide diagnosis","Adaptive compression lifts few-shot cancer diagnosis by 4.3 points","Knowledge-guided patch selection sharpens few-shot pathology","FOCUS: Cut redundant patches, keep the signal in few-shot slides"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1531,"prompt_tokens":1085,"completion_tokens":446,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":701,"completion_tokens_details":{"reasoning_tokens":372}},"tokens_in":701,"tokens_out":446,"duration_ms":5456,"temperature":1.0,"reasoning_tokens":372,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:56:38.015759+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retain the same total number of tokens but replace the language-guided top-$k$ selection with random selection in the 4-shot UBC-OCEAN setting; if balanced accuracy does not drop, the reported gains come from compression rather than from the LLM prompts. Conversely, on mucinous-carcinoma slides with large uniform mucin pools, progressively lowering $\\tau_g$ should show whether similar-but-diagnostic patches are being pruned.","supporting_citations":[{"cited_title":"A visual- language foundation model for computational pathology","cited_arxiv_id":null,"evidence_quote":"Supplies CONCH, the pathology vision-language foundation model used to encode all patch features and text prompts, so the whole compression pipeline runs in its embedding space."},{"cited_title":"The rise of ai language pathologists: Exploring two-level prompt learning for few-shot weakly-supervised whole slide image classification","cited_arxiv_id":null,"evidence_quote":"Introduces the few-shot weakly-supervised learning formulation and the TOP-MIL prompt-learning baseline that FOCUS extends."},{"cited_title":"Vila-mil: Dual-scale vision-language multiple instance learning for whole slide image classification","cited_arxiv_id":null,"evidence_quote":"ViLa-MIL is the strongest prior few-shot vision-language MIL method and the main second-best baseline that FOCUS outperforms."},{"cited_title":"Diagnos- tic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer","cited_arxiv_id":null,"evidence_quote":"Supplies the CAMELYON16 breast cancer metastasis dataset used in the few-shot evaluation."},{"cited_title":"Review the cancer genome atlas (tcga): an im- measurable source of knowledge","cited_arxiv_id":null,"evidence_quote":"Supplies the TCGA-NSCLC lung cancer dataset used for the LUAD/LUSC few-shot experiments."}],"review_version":1}