{"id":"6cd02aad-b6d4-404c-ba0b-fa7a2593d726","arxiv_id":"2501.02922","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A new multiple instance learning model for whole-slide images produces inherently interpretable predictions by combining vision-language concept detection with a linear decision rule.","lead":"Researchers introduce Concept MIL, a method that classifies gigapixel cancer tissue slides while explaining each prediction using human-understandable pathology concepts instead of just heatmaps. It matches the accuracy of leading models on two public datasets and could make medical AI decisions easier for pathologists to verify.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Concept MIL's interpretability hinges on CONCH embedding alignment; the paper's own Fig. 2 shows misalignment, and the filtering step (III-B) is too weak to guarantee reliable concept activations across test data.","rationale":"The reader's weakest assumption points to the same load-bearing concern: the reliability of CONCH's shared embedding space for concept detection. I agree with that assessment. The paper's own Fig. 2 is direct evidence of misalignment, and Sec. III-B's filtering step is described too loosely to establish that the final concept set is trustworthy across the full test distribution. The central claim of inherent interpretability depends entirely on the semantic validity of the concept activations; if these are unreliable, the explanations (Eq. 7, Eq. 8, Fig. 3d) become arbitrary numbers, and the model is not meaningfully interpretable. My check would settle whether the filtering is adequate by measuring concept-detection precision on a broader, more representative sample with multiple expert raters. I do not propose changing the reader's CONDITIONAL verdict, as the concern is real but the paper could address it with additional validation; the current evidence is inconclusive rather than disqualifying.","tokens_in":14616,"tokens_out":3654,"duration_ms":118581,"concrete_test":"Construct a gold-standard evaluation set of 300 patches per concept, sampled from the top 1% of cosine-similarity scores across a held-out split of the test set (including both tumor and normal WSIs). Have three independent pathologists label each patch for presence/absence of the target concept, and compute per-concept precision (fraction of high-activation patches confirmed by majority vote) and Fleiss' kappa. If any concept's precision falls below 0.7, or kappa is below 0.6, the concept filtering in Sec. III-B is insufficient and the central interpretability claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Concept MIL is an inherently interpretable model whose predictions are a linear combination of human-understandable concepts (Eq. 8), so that each concept's influence can be traced. This claim requires that the concept activation vectors f_jc—cosine similarities between CONCH image and text embeddings—reliably indicate the presence of the named pathology concepts. The paper itself demonstrates this assumption fails: Fig. 2 shows fat-cell patches receiving high cosine similarity with the concept 'fibrous tissue'. To address this, Sec. III-B filters misaligned concepts using a pathologist who reviews five high-activation patches per concept from 50K tumor patches in the training set. This filtering is insufficient for three reasons: (1) it verifies only tumor-region patches from the training set, not normal patches or test WSIs, so alignment on the test distribution is unmeasured; (2) five patches per concept is a tiny sample for assessing semantic reliability across heterogeneous gigapixel slides; (3) the pathologist's judgment is a single binary decision with no inter-observer agreement or quantitative precision/recall. Moreover, the concept MIL branch alone achieves only 0.688 AUC on Camelyon16 (Table II), suggesting the concept features carry limited discriminative signal. If the concept activations are unreliable, the explanations are unfaithful, and the model's 'inherent interpretability' claim collapses, regardless of classification accuracy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Concept MIL, a dual-branch multiple-instance-learning model for whole-slide image classification in histopathology. An image MIL branch selects the top K salient patches via attention, and a concept MIL branch classifies the slide using concept activation vectors computed as cosine similarities between CONCH image embeddings and text embeddings of pathologist-refined concepts. The authors claim this is the first concept-based, inherently interpretable WSI model, with predictions obtained as a linear combination of human-understandable concepts, enabling faithful local and global explanations. Experiments on Camelyon16 and PANDA report AUC/accuracy above 0.9, localization scores around 0.85-0.87, and a small user study with three pathologists showing moderate agreement with the model's top concepts.","tokens_in":14941,"tokens_out":4649,"duration_ms":49261,"significance":"If the interpretability claim holds, the paper would be a valuable contribution to computational pathology, offering concept-level explanations without manual concept annotation and with classification performance on par with state-of-the-art MIL models. The work is timely given the rise of pathology foundation models and the need for trustworthy AI in medicine. The authors promise to release code and data, which would facilitate reproducibility. However, the central claim of inherent interpretability is weakened by the non-linear attention gating in the concept branch and by the limited validation of the underlying vision-language embedding alignment.","major_comments":[{"comment":"The prediction is not obtained through a purely linear combination of concepts, because beta_c in Eq. (5) is the output of a non-linear attention network Ac(·) operating on the entire transposed concept feature matrix of the top-K patches (Eqs. 3-4). Consequently, the contribution kappa_c in Eq. (7) is not an independent, counterfactual measure of concept c's influence: beta_c itself depends on all concepts and all selected patches, so changing one concept's activation can alter the apparent contribution of other concepts through the shared attention mechanism. The claim that the model enables 'tracing each concept's influence on the prediction' is therefore not fully justified. The authors should either modify the architecture to obtain a truly linear decomposition (e.g., removing the attention gating) or substantially temper the interpretability claim and discuss the faithfulness limitations introduced by the non-linear gating, in the same way the paper criticizes attention maps in ABMIL.","section":"Section III-D, Eqs. (5)-(8)"},{"comment":"The reliability of concept activations rests on the alignment between CONCH's image and text embedding spaces, and the paper itself demonstrates a misalignment: fat-cell patches receive high cosine similarity with 'fibrous tissue' (Fig. 2). The proposed filtering step—asking a pathologist to verify five high-activation patches per concept from 50K tumor-region patches in the training set—is too weak to establish semantic reliability across the test distribution. It samples only tumor patches, not normal tissue or test WSIs; it uses only five patches per concept, which is a very small sample for heterogeneous gigapixel slides; and it relies on a single pathologist's binary decision with no quantitative precision/recall or inter-observer agreement. Since all downstream explanations are built on these activations, the authors should provide stronger evidence that the refined concept set is reliably detected across the evaluated datasets, or explicitly scope the interpretability claim to the verified distribution.","section":"Section III-B, Fig. 2, Table I"},{"comment":"The concept MIL branch alone achieves only 0.688 AUC on Camelyon16, compared to 0.970 for the image MIL branch, indicating that the concept features carry limited discriminative signal on this dataset. The final model's good performance therefore depends heavily on the image branch's top-K patch selection, which is not concept-based. This raises a conceptual question: is the concept branch truly the decision-maker, or is it largely a post-hoc explainer of the image branch's selections? The paper should clarify the causal role of the two branches, for example by reporting how performance and explanation quality vary with K, or by ablating the image branch's influence at inference time. The current framing as an 'inherently interpretable' model is not fully supported if the core discriminative power comes from a branch whose selection mechanism is not concept-based.","section":"Section IV-E, Table II"}],"minor_comments":[{"comment":"The user study is small (10 samples per dataset, three pathologists) and the agreement scores range from 0.25 to 0.55; the claim that the model's concepts 'align with the concepts used by pathologists' should be reported with confidence intervals or a qualitative disclaimer, given the small sample.","section":"Section IV-F2"},{"comment":"The disease localization score for TransMIL is missing; the text explains why, but a placeholder (e.g., '—') in the table with a footnote would improve readability.","section":"Table III"},{"comment":"The concept filtering step would be more reproducible if the authors reported how many initial GPT4o concepts were proposed and how many were removed by the pathologist, along with the criteria used for removal.","section":"Section III-B"},{"comment":"The standard deviation in Eq. (4) is not clearly defined; please specify whether it is computed over the concept dimension, the patch dimension, or the full matrix, as this affects the scaling behavior.","section":"Eq. (4)"},{"comment":"The final model is taken from the last epoch after 300 epochs without mention of early stopping or validation-based model selection; please justify this choice or describe the validation procedure used for hyperparameter tuning.","section":"Section IV-B"},{"comment":"The term 'label-free' is somewhat misleading because the method still requires a pathologist to refine the concept set and to verify high-activation patches; the authors should clarify this nuance in the abstract or introduction.","section":"Abstract and Section III-B"}],"recommendation":"major_revision","confidential_remarks":"The paper is clearly written and the experiments are useful, but the central interpretability claim needs to be reconciled with the non-linear attention gating in the concept branch. The concept-alignment validation is also thin. These are fixable with additional analysis and more cautious claims, so I lean toward major revision rather than rejection. The missing comparison to SI-MIL is understandable given the unavailability of the pipeline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper is a credible first attempt at bringing label-free concept bottleneck models to gigapixel WSI classification. The core idea is simple and well-executed: use CONCH to project patch embeddings into a predefined tissue-concept space, train a standard attention MIL branch to select the top-20 salient patches, and let a concept branch read out a linear classifier on those concept activations. On Camelyon16 and PANDA they match state-of-the-art accuracy/AUC, and the localization scores (87%/85% of top-20 patches inside tumor) are strong. The writing is clear, and the ablation showing the concept branch alone is weak (0.688 AUC on Camelyon16) is honest, because it exposes how dependent the concept branch is on the image branch's patch selection.\n\nThe interpretability claim has two soft spots. First, Eq. 5 is not a linear function of the concepts in the usual CBM sense: the per-concept weights beta are themselves outputs of a gated attention network over the entire selected patch set, so the contribution kappa_c in Eq. 7 is not an independent additive component; changing one patch's concept vector can change all kappa_c. This weakens the 'tracing each concept's influence' claim, though it is not fatal—the same criticism applies to many attention-based explanations. Second, the concept filtering step in Sec. III.B is too thin to guarantee meaningful activations on test data: one pathologist reviews five high-activation patches per concept from 50K tumor-only training patches. That does not measure alignment on normal tissue or on the test distribution, and five patches is a small sample for a gigapixel slide. The paper itself shows a failure case (fat vs. fibrous tissue) in Fig. 2, so this is a real concern, not a nitpick. The user study is also small (3 pathologists, 10 samples per dataset) and agreement numbers are modest (0.25-0.55), but inter-pathologist agreement is similar, which puts the model in context.\n\nNo evidence of circularity: classification and localization are independently grounded. The math is correct as far as it goes. Missing error bars across seeds and code release are standard referee requests, not fatal flaws.\n\nThis paper deserves a serious referee. The novelty is real—first label-free concept MIL for WSI—and the limitations are fixable in revision: temper the 'inherent' language, add a quantitative concept-reliability check, and report variance. I would bring it to a reading group, and I would cite it if I worked on interpretable pathology AI.","headline":"Useful first step: label-free concept bottleneck for WSI, but the 'inherently interpretable' claim is softer than the paper makes it sound.","tokens_in":15497,"tokens_out":2073,"would_cite":true,"duration_ms":20439,"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":"This paper proposes Concept MIL, a whole-slide image classifier that makes each prediction a weighted sum of human-understandable pathology concepts, removing the need for manual concept annotations.","keywords":["multiple instance learning","concept bottleneck models","whole-slide images","histopathology","inherent interpretability","vision-language foundation models","label-free concepts","weakly supervised classification"],"falsifier":"On a held-out set of WSIs with pixel-level annotations for individual concepts (for example, pathologist-outlined regions of 'dense nuclei' or 'cribriform pattern'), compute the correlation between each patch's concept activation score and whether a human labeled that concept as present; if several concepts show near-zero or negative correlation, the concept detection is not grounded in image content and the explanation mechanism fails. A complementary check is to re-run the trained model on a test slide with the largest single concept contribution removed from the linear sum, which should shift the logit by exactly the reported $\\kappa_c$; if it does not, the linear decomposition does not reflect the actual decision process.","tokens_in":14441,"feed_emoji":"🔬","tokens_out":8397,"duration_ms":76651,"temperature":0.7,"pith_summary":"This paper proposes Concept MIL, a whole-slide image classifier whose predictions can be read off as a sum of human-understandable pathology concepts. It claims to be the first concept-based, inherently interpretable model for gigapixel histopathology: instead of only highlighting a heatmap, it tells you that the slide is called tumor because of, say, 'dense nuclei' and 'disorganized cells' in the top 20 patches, and how much each concept contributed. The model avoids manual concept annotations by using a vision-language model trained on pathology image-text pairs to score each patch against a concept list, and the final prediction is a linear combination of concept scores, so contributions are exact by construction. On Camelyon16 and PANDA it reports accuracy and AUC above 0.9, on par with standard MIL models, and a pathologist user study indicates the concepts match clinical reasoning. If correct, it means an interpretable gigapixel classifier can be built with slide-level labels only, and every decision can be audited concept by concept.","feed_headline":"Tumor calls explained by pathology concepts, no concept labels needed","feed_subtitle":"Concept MIL reaches over 0.9 AUC on two pathology benchmarks while tracing every prediction to specific human-readable concepts.","key_machinery":"The load-bearing object is the linear concept decomposition. Concept activation vectors are built by cosine similarity between CONCH image patch embeddings and text embeddings of prompts like 'an H&E image of CONCEPT'; the concept MIL branch then applies a sparse, percentile-scaled attention vector $\\beta$ over concepts and a logistic classifier over the sum $\\sum_j \\sum_c w_c f_{jc}\\beta_c$ across the top-$K$ patches chosen by a differentiable Patch Attention-Guided Top-K module. This linear pre-activation is what makes the prediction an exact sum of per-concept contributions $\\kappa_c$, so the explanations are not post-hoc approximations. The architecture also includes a jointly trained attention-based image MIL branch whose only job at inference is to supply the top-$K$ patch indices; the image branch's own prediction is discarded.","core_discovery":"The central claim is that Concept MIL is the first inherently interpretable WSI classifier whose decisions are expressed in pathology concepts rather than spatial heatmaps. The model operates by embedding all WSI patches with the CONCH image encoder, computing concept activation vectors as cosine similarities between patch embeddings and text embeddings of a predefined concept set, selecting the top 20 patches by attention from a jointly trained image MIL branch, and making the final prediction by a logistic function of a weighted sum over concepts and selected patches. Because the pre-activation is a linear combination, the contribution of each concept to the prediction is exactly $\\kappa_c = w_c \\sum_{j=1}^{K} f_{jc}\\beta_c$, giving an inherent local explanation; dataset-level averages and t-SNE projections give global explanations. The authors report that this scheme reaches AUC and accuracy over 0.9 on Camelyon16 and PANDA, places 87.1% and 85.3% of the top-20 patches inside tumor regions, and produces top concepts that overlap with pathologists' choices in a three-pathologist user study.","pith_inferences":["The 'label-free' claim applies to training supervision, not the whole pipeline: the concept list is drafted with a general-purpose language model and then refined and verified by a pathologist, so expert effort moves from labeling concepts to curating them.","The linear logit makes counterfactual explanations easy to build — asking how the prediction would change if a concept were absent — although the paper does not develop this explicitly.","The method's interpretability is only as trustworthy as the embedding alignment of the vision-language model; for other stains, tissues, or concept sets, the manual filtering step the authors used would likely need to be repeated.","The binary design suggests a direct extension to Gleason grading or multi-class tasks by replacing the logistic head with a softmax over the same per-concept contributions, and multi-scale concepts could be handled by pooling concept scores across magnifications."],"forward_implications":["A clinician can trace each tumor or normal call to specific pathology concepts on specific top-K patches rather than to an opaque heatmap.","A WSI classifier with concept-level explanations can be trained using only slide-level labels, with no pixel or concept annotations.","Adapting the model to a new disease reduces to defining a new concept list and text prompts, as long as the vision-language embeddings are reliable for those concepts.","Dataset-level concept contribution plots allow developers to audit whether the model's decisions rest on clinically sensible features before deployment.","Because the prediction is a linear function of concept scores, removing a concept's contribution changes the logit by exactly its stated contribution, making the explanations faithful by construction."],"supporting_citations":[{"why":"Supplies the shared image-text embedding space used both for patch features and for cosine-similarity concept activation; the core enabler of label-free concepts.","marker":"[26]"},{"why":"Provides the dual-branch self-interpretable MIL design and the differentiable PAG Top-K patch selection that Concept MIL adapts.","marker":"[22]"},{"why":"Shows how to build concept bottleneck models without labeled concepts using a vision-language model; the conceptual template for label-free concept features.","marker":"[28]"},{"why":"Attention-based MIL forms the image branch whose attention scores select the top-K patches for the concept branch.","marker":"[20]"},{"why":"The key inherently interpretable MIL baseline whose heatmap-only explanations motivate concept-based explanations.","marker":"[21]"},{"why":"Camelyon16 dataset provides the breast cancer metastasis slides and pixel-level tumor masks used for training and localization evaluation.","marker":"[2]"},{"why":"PANDA dataset provides the prostate biopsy slides, Gleason/ISUP labels, and tumor masks used for the second evaluation.","marker":"[4]"},{"why":"GPT4o drafts the initial pathology concept lists that are later refined by a pathologist.","marker":"[29]"},{"why":"Pointing game is extended into the disease localization score that quantifies whether top-K patches fall in tumor regions.","marker":"[45]"}],"fun_headline_variants":["No concept labels needed: AI explains tumor slides using pathology terms","Concept MIL: Gigapixel pathology explained in human terms, no annotations","AI reads histopathology and tells you why, using pathologist concepts","Pathology concepts, zero concept labels: interpretable gigapixel tumor detection","Tumor predictions traced to human-readable pathology concepts, no labeling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that CONCH's shared image-text embedding space is aligned well enough that a patch's cosine similarity to a concept text reliably detects that concept; the paper itself shows this fails for 'fibrous tissue' and requires a manual pathologist filtering step, so if that alignment is unreliable for a chosen concept set, the concept activations and their explanations lose meaning.","fun_headline_variants_meta":{"raw":{"variants":["No concept labels needed: AI explains tumor slides using pathology terms","Concept MIL: Gigapixel pathology explained in human terms, no annotations","AI reads histopathology and tells you why, using pathologist concepts","Pathology concepts, zero concept labels: interpretable gigapixel tumor detection","Tumor predictions traced to human-readable pathology concepts, no labeling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000306,"raw_usage":{"total_tokens":1797,"prompt_tokens":1034,"completion_tokens":763,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":671}},"tokens_in":650,"tokens_out":763,"duration_ms":7011,"temperature":1.0,"reasoning_tokens":671,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:00:13.931504+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out set of WSIs with pixel-level annotations for individual concepts (for example, pathologist-outlined regions of 'dense nuclei' or 'cribriform pattern'), compute the correlation between each patch's concept activation score and whether a human labeled that concept as present; if several concepts show near-zero or negative correlation, the concept detection is not grounded in image content and the explanation mechanism fails. A complementary check is to re-run the trained model on a test slide with the largest single concept contribution removed from the linear sum, which should shift the logit by exactly the reported $\\kappa_c$; if it does not, the linear decomposition does not reflect the actual decision process.","supporting_citations":[{"cited_title":"A visual-language foundation model for computational pathology","cited_arxiv_id":null,"evidence_quote":"Supplies the shared image-text embedding space used both for patch features and for cosine-similarity concept activation; the core enabler of label-free concepts."},{"cited_title":"Si-mil: Taming deep mil for self-interpretability in gigapixel histopathology","cited_arxiv_id":null,"evidence_quote":"Provides the dual-branch self-interpretable MIL design and the differentiable PAG Top-K patch selection that Concept MIL adapts."},{"cited_title":"Additive mil: Intrinsically interpretable multiple instance learning for pathology","cited_arxiv_id":null,"evidence_quote":"The key inherently interpretable MIL baseline whose heatmap-only explanations motivate concept-based explanations."},{"cited_title":"Diagnostic assessment of deep learning algorithms for detec- tion of lymph node metastases in women with breast cancer","cited_arxiv_id":null,"evidence_quote":"Camelyon16 dataset provides the breast cancer metastasis slides and pixel-level tumor masks used for training and localization evaluation."},{"cited_title":"Steiner, Hester van Boven, Robert Vink, Christina Hulsbergen van de Kaa, Jeroen van der Laak, Mahul B","cited_arxiv_id":null,"evidence_quote":"PANDA dataset provides the prostate biopsy slides, Gleason/ISUP labels, and tumor masks used for the second evaluation."},{"cited_title":"Chatgpt: Optimizing language models for dialogue","cited_arxiv_id":null,"evidence_quote":"GPT4o drafts the initial pathology concept lists that are later refined by a pathologist."},{"cited_title":"Top-down neural attention by excitation backprop","cited_arxiv_id":null,"evidence_quote":"Pointing game is extended into the disease localization score that quantifies whether top-K patches fall in tumor regions."}],"review_version":1}