{"id":"457136c7-d01f-4914-b4d5-f03e8ce54e4a","arxiv_id":"2608.13148","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A prior-guided hybrid concept bottleneck model improves concept detection at low annotation rates without losing diagnostic accuracy, as shown on mammographic masses, calcifications, and lung nodules.","lead":"The authors propose a concept bottleneck model that combines sparse concept annotations with class-level prevalence priors to keep cancer image predictions interpretable. In low-annotation settings, it improves concept detection over a standard CBM on mammography and CT benchmarks while preserving diagnostic accuracy.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Data priors are estimated from full-training concept labels, so the hybrid receives aggregate label information at every annotation fraction; the 'less annotation' claim is not yet supported by the comparison.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the method's headline benefit rests on aggregate concept statistics computed from the full training cohort. My read sharpens this into a concrete comparison flaw: the hybrid model at low annotation fractions uses all training concept labels in aggregate form, while the 'matched' standard CBM uses only the annotated subset. Because data priors outperform manual priors and prior initialization is the main ablation driver, the experimental design cannot separate the method's mechanism from this information advantage. The authors honestly flag the need for historical aggregate statistics in Section 4.6, but the abstract and conclusion do not carry that caveat. This does not invalidate the experiments as measurements of a conditional claim; it makes the central 'less annotation' contribution weaker than stated. A hold-out-prior recomputation would settle whether the gains persist when priors are estimated from a separate cohort or from only the annotated subset. Since the reader already assigned CONDITIONAL and the concern supports that verdict without changing it, no adjustment is needed.","tokens_in":8380,"tokens_out":4169,"duration_ms":40885,"concrete_test":"Hold out the concept labels of a random half of training patients entirely, never using them for either instance-level loss or prior estimation; estimate data priors only from the remaining half, and rerun the 0/5/10/20% annotation comparisons with identical hyperparameters. If the hybrid-versus-standard concept AUC gaps at 10% shrink substantially, the reported gains are attributable to full-cohort aggregate label access rather than to the prior-guided training procedure. Alternatively, force the method to use only manual priors, which require no concept labels; the gap between manual-prior and data-prior performance would then quantify the residual annotation burden.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 states that sampled annotation fractions apply only to labels used in the instance-level concept loss, while 'data-derived aggregate priors were computed separately from the training patients of each fold.' This means pi_y in Eq. (3) and P(y|c_k) in Eq. (4) are estimated from concept labels of the entire training cohort, including patients whose instance labels are withheld from the standard CBM baseline. At every annotation fraction, including 0%, the hybrid receives aggregate label information that the matched standard CBM does not. The headline gains at 10% annotation (0.619 to 0.741, 0.650 to 0.787, 0.597 to 0.642) therefore conflate the proposed mechanism with an information advantage: the baseline is denied the same aggregate supervision. Supplementary Fig. S1 and Table 1 also show that data-derived priors generally outperform manual priors, and that prior initialization, which uses P(y|c_k), is the main driver of gains. Since manual priors do not require labels but are weaker, the practical 'reduced annotation burden' claim depends on label-requiring aggregate statistics. Section 4.6 acknowledges this limitation, but the abstract and conclusion present the annotation reduction as established rather than conditional on full-label prior estimates.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a prior-guided hybrid concept bottleneck model (PriorHybrid-CBM) that combines limited instance-level concept annotations with class-conditional distribution matching on unannotated patients and prior initialization of the concept-to-diagnosis head. The method is evaluated on CBIS-DDSM mammographic masses and calcifications and LIDC-IDRI pulmonary nodules across concept annotation fractions from 0% to 100%. The central empirical claim is that in the clinically relevant 0-20% annotation regime, the hybrid CBM consistently improves mean concept ROC-AUC over a matched standard CBM while maintaining diagnostic AUC close to black-box models, with reported 10%-annotation improvements from 0.619 to 0.741 (masses), 0.650 to 0.787 (calcifications), and 0.597 to 0.642 (nodules). The paper also reports that zero-shot vision-language models are insufficient for fine-grained tumor-level concept prediction.","tokens_in":8539,"tokens_out":3865,"duration_ms":34530,"significance":"If the reported gains hold under a properly matched information comparison, the method would be a meaningful step toward reducing concept-annotation costs for interpretable medical imaging. The empirical work has notable strengths: percentile bootstrap confidence intervals, paired significance tests, patient-level splitting, a component ablation, and publicly released code. The paper also provides a clear negative result for zero-shot VLM transfer. However, the central 'less annotation' claim currently rests on a comparison in which the hybrid model receives aggregate concept-label statistics from the full training cohort, while the standard CBM baseline does not. This limits the validity of the headline conclusion as stated.","major_comments":[{"comment":"The principal concern raised in the review process is confirmed by the manuscript text. Section 3.2 states that 'data-derived aggregate priors were computed separately from the training patients of each fold' and that annotation fractions 'applied only to labels used in the instance-level concept loss.' As a result, at every annotation fraction, including 0%, the hybrid model receives class-conditional concept prevalence vectors pi_y (Eq. 3) and P(y|c_k) (Eq. 4) estimated from the concept labels of the entire training cohort, whereas the matched standard CBM receives none of this aggregate label information. The reported concept-AUC gains in the 0-20% regime, including the 10% figures in the abstract, therefore conflate the proposed mechanism with an information advantage. The authors should either re-run the experiments with priors estimated only from the annotated subset (or from a separate external source) to provide a truly matched comparison, or re-frame the central claim as conditional on the availability of full-training-set aggregate priors. As it stands, the 'less annotation' claim is not supported by the current experimental design.","section":"Section 4.6, Table 1"},{"comment":"The ablation titled 'Random init.' removes prior initialization but retains the distribution-matching loss (lambda_d) and the anchoring loss (lambda_r). Since the distribution-matching targets are data-derived priors computed from the full training cohort, this ablation does not isolate the effect of prior initialization from the use of full-cohort aggregate label information. The observation in Section 4.4 that random-init diagnosis heads drift far from the data-prior direction is made at 0% concept annotation, where the standard CBM has no concept supervision at all; the comparison is therefore not information-matched. The conclusion that 'prior initialization is the main component' needs to be re-stated with the caveat that this component is evaluated in the presence of full-training-set prior statistics.","section":"Section 4.4, Section 4.6"},{"comment":"The paper's own results show that data-derived priors generally outperform manual priors, which are label-free. For example, in the 0-20% regime, mean concept AUC improves from 0.681 to 0.716 for masses and from 0.726 to 0.757 for calcifications when moving from manual to data-derived priors. This implies that the practical annotation-reduction benefit depends on having access to cohort-level label statistics, which may themselves require substantial annotation effort. The limitation section (4.6) acknowledges this, but the abstract and conclusion state the annotation reduction as established rather than conditional on such statistics. The framing should be corrected to match the evidence.","section":"Supplementary Fig. S1, Section 4.6"}],"minor_comments":[{"comment":"Typo: 'Several approached have been proposed' should be 'Several approaches have been proposed.'","section":"Introduction"},{"comment":"Typo: 'specific dataset' is written as 'specifc dataset.'","section":"Section 2.2"},{"comment":"The nodule crop size is written as '483' without a superscript; it should be 4^3 (64) or the intended notation should be clarified.","section":"Section 3.1"},{"comment":"The table captions contain 'T able' rather than 'Table'; please fix the formatting.","section":"Table captions"},{"comment":"The selection of lambda_d is performed using validation concept AUC over the 0-20% regime. This is a second, smaller use of concept labels beyond the training objective and should be mentioned when summarizing the label budget of the method.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the core experiments are sound, but the headline 'less annotation' is not actually supported by the comparison as run. The priors that drive the gains are estimated from the full training cohort's concept labels, so the hybrid model receives aggregate label information at every annotation fraction, including 0%. The matched standard CBM baseline is denied that. That is a real information mismatch, and the stress-test note is right. The authors do acknowledge this in Section 4.6, but the abstract and conclusions present it as established rather than conditional.\n\nWhat's new: the specific combination of instance-level concept loss, class-conditional distribution matching, and prior-based head initialization for low-annotation cancer imaging isn't in the prior work. The experiments are carefully done: bootstrap confidence intervals, paired significance tests, ablations, and released code. The finding that prior initialization of the diagnosis head is the main driver is plausible and backed by the drift analysis. The zero-shot VLM comparison is a useful negative result; those models are not ready for fine-grained tumor-level concepts.\n\nSoft spots: the main one is the annotation-burden claim. Since data priors require aggregate concept statistics from the full training set, the method does not yet show it can reduce annotation effort from scratch. It shows that if you already have cohort-level concept prevalence, you can train with fewer instance-level labels. That is still useful, but it's a different claim. The manual prior sweep in Fig. S1 shows weaker gains, which underscores the point. Also, the closest prior work (Nahiduzzaman et al.) is cited but not used as a baseline, so the marginal contribution over class-level prior supervision alone is not quantified. Minor points: no per-concept error analysis, and the concept-correction experiment is oracle, as the authors note.\n\nOverall: the paper is honest and the engineering is solid. The central mechanism — anchoring the concept-to-diagnosis head to a clinical prior — is reasonable, and the empirical evidence is consistent with it. The 'less annotation' claim needs to be reframed as 'less instance-level annotation given aggregate priors.' I'd send this to peer review; a good reviewer can push them to match the information budget and add the missing baseline. For a reading group it's a useful data point in the CBM literature, though not a must-read.","headline":"Solid empirical study of a prior-guided hybrid CBM, but the 'less annotation' headline is conditional on aggregate priors computed from full training labels, which the standard baseline does not receive.","tokens_in":9122,"tokens_out":1934,"would_cite":true,"duration_ms":18182,"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 prior-guided hybrid concept bottleneck model improves mean concept AUC in the 0–20% annotation regime across three cancer imaging tasks while keeping diagnostic AUC close to black-box baselines.","keywords":["concept bottleneck models","interpretable medical imaging","limited concept supervision","class-conditional priors","annotation efficiency","cancer imaging diagnosis","mammography","pulmonary nodules"],"falsifier":"Compute the data-derived priors from only the annotated 10% of training patients (instead of the full training cohort) and rerun the low-annotation experiments; if the mean concept AUC advantage over the standard CBM shrinks to near zero or reverses, the paper's central benefit is attributable to full-cohort aggregate labels rather than to the prior-guided mechanism itself.","tokens_in":8094,"feed_emoji":"🩻","tokens_out":8243,"duration_ms":59628,"temperature":0.7,"pith_summary":"The paper claims that dense, expensive per-patient concept annotations are unnecessary for interpretable cancer diagnosis. By adding class-conditional concept prevalence priors and a clinically initialized concept-to-diagnosis head to a concept bottleneck model, it achieves substantially higher concept detection AUC than a standard CBM when only 0–20% of patients have concept labels, while keeping diagnostic AUC close to black-box performance. The reported gains at 10% annotation are large for the three studied tasks—mass concept AUC rises from 0.619 to 0.741, calcifications from 0.650 to 0.787, and lung nodules from 0.597 to 0.642—so the practical argument is that a small annotation set plus aggregate clinical statistics suffices for transparent tumor diagnosis.","feed_headline":"Prior guidance lifts concept AUC with few labels in cancer imaging","feed_subtitle":"At 10% concept annotation, mean concept AUC rises to 0.74 for masses, 0.79 for calcifications, and 0.64 for lung nodules.","key_machinery":"The central object is the prior-guided hybrid CBM, a concept bottleneck model whose concept head $g_\\theta(x)$ maps images to concept probabilities $\\hat{c}=\\sigma(z)$, followed by a linear diagnosis head. Three mechanisms carry the argument: (1) a distribution-matching loss $L_{\\text{dist}}$ that compares the mean predicted concept vector of unannotated patients in a minibatch to a class-conditional prevalence prior $\\pi_y$, giving weak supervision from aggregate label proportions; (2) initialization of the diagnosis head weight for concept $k$ as $\\logit\\{P(y=1\\mid c_k=1)\\}-\\logit\\{P(y=1\\mid c_k=0)\\}$, with bias from the malignancy base rate; and (3) an optional L2 anchor $L_{\\text{reg}}$ keeping the head near that initialization. The load-bearing mechanism is the stabilized concept-to-diagnosis mapping: with a plausible start, gradient updates teach image-to-concept mappings while the diagnosis head stays close to the prior, which the drift analysis shows does not happen with random initialization.","core_discovery":"The central discovery is that a hybrid supervision scheme—mixing instance-level concept labels on a few patients with class-conditional concept distribution matching on the rest, plus a prior-initialized diagnosis head—lets concept bottleneck models learn fine-grained radiological concepts annotation-efficiently. The paper reports consistent mean concept ROC-AUC gains in the 0–20% annotation regime relative to a matched standard CBM, with the largest effect coming from prior initialization of the concept-to-diagnosis head rather than from the distribution loss alone. Ablations show that the prior-initialized head drifts almost not at all (relative L2 drift under 0.5%) when no concept labels are available, while randomly initialized heads drift by more than 85%. The paper also shows that zero-shot vision-language models are insufficient for fine-grained tumor-level concept recognition, scoring around 0.50 AUC.","pith_inferences":["If the prior-initialization effect generalizes, prevalence priors drawn from historical radiology reports or clinical databases could cold-start interpretable models for new cancer types where only outcome labels exist, without any new concept annotation.","The distribution-matching loss is learning from label proportions; a natural extension is estimating $\\pi_y$ iteratively from the unannotated pool (e.g., via expectation-maximization), which would remove the need for full-cohort concept statistics and directly test how much aggregate information the method truly requires.","The oracle concept-correction gains suggest a prospective clinician-in-the-loop study: have radiologists correct only high-impact concept predictions, then measure real diagnostic AUC improvement over a no-correction baseline; the paper's oracle numbers bound the possible benefit."],"forward_implications":["A clinically meaningful concept bottleneck for cancer imaging can be trained with as little as 0–20% patient-level concept annotations, with concept AUC close to what fully annotated models give in the tested tasks.","The diagnosis head initialized from clinical priors is robust: at 0% annotation its weights drift by less than 0.5% relative, while randomly initialized heads drift by more than 85%, explaining why prior initialization matters most when labels are absent.","Zero-shot vision-language models do not yet provide reliable fine-grained tumor-level concept predictions, so annotation-free concept bottleneck models are not a viable alternative in this setting.","Correcting concept predictions at test time improves diagnosis AUC by up to 0.139 on lung nodules, indicating that selective clinician review of concept outputs could become a practical workflow.","Data-derived class-conditional priors outperform manual BI-RADS-style priors for concept supervision, showing that cohort-calibrated aggregate statistics are the stronger prior source."],"supporting_citations":[{"why":"Defines concept bottleneck models, the architecture the hybrid builds on.","marker":"[8]"},{"why":"Supplies learning-from-label-proportions theory underlying the class-conditional distribution-matching loss.","marker":"[16]"},{"why":"Introduces class-level prior supervision for weakly supervised concept learning, which this paper extends.","marker":"[13]"},{"why":"Provides the CBIS-DDSM mammography data with mass and calcification cohorts and concept labels.","marker":"[9]"},{"why":"Provides the LIDC-IDRI lung CT data with nodule malignancy and radiological concept labels.","marker":"[2]"},{"why":"Mammo-CLIP serves as the zero-shot vision-language baseline for mammography concept prediction.","marker":"[4]"},{"why":"CT-CLIP serves as the zero-shot vision-language baseline for lung CT concept prediction.","marker":"[5]"}],"fun_headline_variants":["Prior-guided CBM lifts concept AUC with 10% labels","Few-shot concept learning: prior init beats distribution loss","Hybrid CBM: 80% fewer concept labels, near-black-box accuracy","Zero-shot VLMs lag, prior-guided CBMs win on concepts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's benefit depends on having reliable class-conditional concept prevalence estimates (the expected fraction of each radiological concept in benign versus malignant cases), and in the experiments these are computed from the full training set's concept labels even when instance-level supervision is only 0–20%; if those aggregate statistics are not available without near-full annotation, the claimed reduction in annotation burden weakens.","fun_headline_variants_meta":{"raw":{"variants":["Prior-guided CBM lifts concept AUC with 10% labels","Few-shot concept learning: prior init beats distribution loss","Hybrid CBM: 80% fewer concept labels, near-black-box accuracy","Zero-shot VLMs lag, prior-guided CBMs win on concepts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000693,"raw_usage":{"total_tokens":3148,"prompt_tokens":969,"completion_tokens":2179,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":2105}},"tokens_in":585,"tokens_out":2179,"duration_ms":14638,"temperature":1.0,"reasoning_tokens":2105,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:43:45.191488+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the data-derived priors from only the annotated 10% of training patients (instead of the full training cohort) and rerun the low-annotation experiments; if the mean concept AUC advantage over the standard CBM shrinks to near zero or reverses, the paper's central benefit is attributable to full-cohort aggregate labels rather than to the prior-guided mechanism itself.","supporting_citations":[{"cited_title":"In: Proceedings of the 37th International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"Defines concept bottleneck models, the architecture the hybrid builds on."},{"cited_title":"In: Proceedings of the 26th International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"Supplies learning-from-label-proportions theory underlying the class-conditional distribution-matching loss."},{"cited_title":"arXiv preprint arXiv:2511.01131 (2025)","cited_arxiv_id":null,"evidence_quote":"Introduces class-level prior supervision for weakly supervised concept learning, which this paper extends."},{"cited_title":"Scientific Data4, 170177 (2017) Prior-Guided CBMs for Cancer Imaging 11","cited_arxiv_id":null,"evidence_quote":"Provides the CBIS-DDSM mammography data with mass and calcification cohorts and concept labels."},{"cited_title":"Medical Physics38(2), 915–931 (2011)","cited_arxiv_id":null,"evidence_quote":"Provides the LIDC-IDRI lung CT data with nodule malignancy and radiological concept labels."},{"cited_title":"In: Medical Image Computing and Computer Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Mammo-CLIP serves as the zero-shot vision-language baseline for mammography concept prediction."}],"review_version":1}