{"id":"98635598-2c6e-473a-83f3-765e3df62351","arxiv_id":"2509.11924","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"VMD, a student-teacher-expert distillation network, improves plaque vulnerability classification in unannotated 3D carotid MRI by transferring knowledge from limited annotations and radiology reports.","lead":"A machine learning system called VMD reads 3D carotid artery MRI scans and classifies plaques as vulnerable or stable by combining the scan with limited vessel annotations and radiologist text reports during training. The method works without annotations at test time and compares favorably with junior radiologists on a small in-house dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Radiology reports may contain the ground-truth label; if so, VMD's expert branch performs label leakage rather than domain-knowledge transfer, undermining the central claim.","rationale":"I agree with the reader's weakest assumption: the radiology-report modality is under-specified and may leak the label. This is the single most load-bearing concern because the paper's central contribution is the text-guided distillation of domain knowledge. If the reports simply contain the ground-truth label, the reported improvement is a form of privileged-information distillation, which is a much weaker contribution and not what the authors claim. The concrete test would settle it directly. I also noted the mathematical derivation in Eq. (3)–(5) appears flawed (e.g., the KL in (3) is written between q(zE|xE) and p(zS|xS,y), but (5) replaces the posterior with p(zS|xS); the 'upper bound' language is confused). However, the variational derivation could potentially be repaired, and even a heuristic loss may produce useful regularization; the label-leakage issue, if real, invalidates the contribution's interpretation entirely. Therefore the reader's CONDITIONAL verdict is appropriate: acceptance should require clarification of the report content and/or a leakage-check experiment. My stress-test does not change that verdict; it sharpens the condition.","tokens_in":14637,"tokens_out":4422,"duration_ms":52257,"concrete_test":"Obtain (or ask authors to provide, even redacted) a sample of the radiology reports used in training. Check whether the report text explicitly states the plaque is 'vulnerable'/'stable' or names components (LRC, IPH, ulceration, thrombus) that directly determine the label. Then run a controlled experiment on the same split: re-train VMD after removing the report's diagnostic conclusion (e.g., truncate the final 'Diagnosis' section or mask any sentence containing the label-defining terms), keeping all other settings identical. If the student's test ROC drops from ~0.7136 toward the w/o-expert ablation (~0.6815 in Table IV), label leakage is confirmed. If performance is unchanged, the expert branch is likely contributing independent knowledge.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that VMD leverages radiologists' domain knowledge encoded in radiology reports to improve plaque-vulnerability diagnosis from unannotated 3D MRI. This requires the report modality to contain diagnostic expertise beyond the binary label. The paper never describes the report content: Section IV-A covers patient inclusion and imaging but not the report text; Section III-A defines x_E only as 'radiology report' without specifying what the report contains. In routine carotid MRI reporting, the conclusion typically states whether the plaque is vulnerable or stable (or lists components such as LRC, IPH, ulceration from which the label is directly derived). If so, the expert network is trained on the ground-truth label, and the 'knowledge transfer' from E to S/T is a label-distillation/privileged-information mechanism, not a transfer of domain expertise. This would fully account for the large gains in Tables III and IV: the teacher and student are being guided by a network that has read the answer. The paper provides no evidence against this, and no code/data are released to check. The empirical improvement may still be real, but the method's novelty and interpretation would be substantially weakened. This is the most load-bearing concern because the entire rationale for the text-guided expert branch hinges on the reports being a source of expertise rather than a direct label channel.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VMD, a student-teacher-expert distillation framework for classifying carotid plaque vulnerability from 3D carotid MRI. The student network diagnoses from unannotated MRI; the teacher network additionally sees limited vessel-wall annotations; the expert network accepts radiology reports. Training maximizes a combination of mutual-information-style terms: variational ELBO/KL terms between latent distributions and supervised contrastive losses between student-teacher and teacher-expert prediction pairs, plus a classification cross-entropy term. On an in-house dataset of 502 images, the student reaches ROC 0.7136±0.013, outperforming a fine-tuned 3DResNet-50 baseline (ROC 0.6567±0.0018) and several prior distillation methods, and it also beats two junior radiologists in speed and accuracy. Ablation studies report that each added term improves the student.","tokens_in":14983,"tokens_out":5752,"duration_ms":70474,"significance":"If the results hold, the paper offers a practically attractive recipe: use easily available radiology reports and cheap limited annotations to improve diagnosis from unannotated 3D MRI without extra inference cost. The clinical comparison with junior radiologists and the five-seed ablation study are useful concrete evidence. However, the significance rests on two pillars that are currently shaky: (1) the variational derivation in Section III-A contains algebraic errors and conflates an ELBO with a mutual information; (2) the expert modality is never described, so the reported gains may reflect label leakage through the radiology reports rather than transfer of radiologists' domain expertise. No code or data is released, which limits reproducibility but is not itself a defect for this venue.","major_comments":[{"comment":"The central derivation is not correct as written. In Eq. (3), the first displayed equality writes the integration variable as dθ (network weights) rather than dz_E. More seriously, the expansion of D_KL(q_θ(z_E|x_E) || p(z_S|x_S,y)) introduces p(y|x_S) and p(y|x_S,z_S) with sign errors: the second line contains a term with log p(z_S|x_S,y) that has the wrong sign, and the intermediate lines do not telescope into the final expression. The final inequality in Eq. (4) is the standard ELBO for log p(y|x_S) only if z_S is drawn from the conditional prior p(z_S|x_S) and q is a variational posterior over z_S; here q is defined over z_E conditioned on x_E, and the paper never shows how z_E enters the likelihood p(y|x_S,z_S). Equation (5) then labels this ELBO as I(S,E), but an ELBO for the label is not a mutual information between the student and expert networks. Because this term is the objecti","section":"Section IV-A / Section III-A"},{"comment":"The content of the radiology reports is never described. Section III-A defines x_E only as 'radiology report,' and Section IV-A describes patient inclusion and imaging but not the report text. The ground-truth labeling criterion is that plaques containing LRC, IPH, thrombus, or ulceration are vulnerable. In routine carotid MRI reporting, the conclusion often states these components or directly states whether the plaque is vulnerable. If the reports contain such statements, then the expert network is effectively trained on the target label through text, and the reported improvements from I(S,E) and I(T,E) are label distillation (privileged information), not transfer of radiologists' domain expertise. This would substantially weaken the central claim. The authors should provide examples or a description of the report text, report the proportion of reports that explicitly mention vulnerabil","section":"Tables III and IV"},{"comment":"The ablation baseline is not consistently reported. Table IV's 'w/o both' row reports ROC 0.6645±0.0193, but Table II's baseline (fine-tuned 3DResNet-50 without teacher/expert) reports 0.6567±0.0018. If these are the same setting, the numbers and standard errors should match; if they are different (e.g., a student network with a different architecture or training schedule), this must be stated. Similarly, full VMD in Table IV reports ROC 0.7109±0.0163, while Table II reports 0.7136±0.013. These discrepancies make it difficult to quantify the incremental contribution of each ablated term, and they should be reconciled or explained.","section":"Section V-A"}],"minor_comments":[{"comment":"The improvement percentages over baseline are internally inconsistent. The text says 'ROC: 4.66%, ACC: 1.01%, PRC: 1.63%', but from Table II the absolute percentage-point increases are 5.69, 6.56, and 3.95, respectively. The later list of relative gains (8.67%, 10.05%, 5.06%, 8.52%, 6.96%) is also not fully consistent (ACC should be about 9.95%). Please correct all reported percentages.","section":"Section V-A"},{"comment":"The algorithm says 'Minimise kld_S_E and kld_T_E', but Eq. (5) and Eq. (8) include both the log-likelihood expectation and the KL term; minimizing only the KL would not implement the stated objective. The wording should be aligned with the equations.","section":"Algorithm 1"},{"comment":"The notation \\hat{x}^T_j and \\hat{x}^S_i is used for 'predictions' in Eq. (6), but the contrastive loss is applied to feature representations. Please clarify whether these are class logits, latent features, or post-softmax probabilities; the gradient behavior and the interpretation as mutual information depend on this choice.","section":"Section III-B"},{"comment":"There are numerous typos and grammar issues: 'Variation inference' in the abstract, 'V ARIATIONAL' in the page header, 'BOLDREPRESNETS' in Table II, 'Minimisekld' in Algorithm 1, 'diagnosis of plaque in 3D carotid artery MRI' in the title (probably should be 'plaque vulnerability'). A careful proofread is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper reports a sensible engineering contribution, but the variational derivation in Section III-A is seriously flawed and the radiology-report content is completely unspecified. The latter is the bigger risk: if the reports state the vulnerability label or its defining components, the entire 'expert knowledge transfer' narrative collapses into label distillation. I would ask the authors to fix the derivation and to provide concrete evidence about the report content before considering the paper further."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: this is a modest but real empirical contribution—student-teacher-expert distillation for 3D carotid MRI plaque vulnerability, with ablations showing each term helps. The AUC gain over a 3D ResNet baseline (0.714 vs 0.657) is not huge but consistent across seeds, and the comparison to four KD baselines is fair. The idea of using limited vessel-wall annotations as a teacher modality and radiology reports as an expert modality is sensible and clinically motivated.\n\nMain soft spot: the report modality. The paper never says what is in the radiology reports. In routine carotid MRI reporting, the conclusion usually states whether the plaque is vulnerable or stable, or lists LRC/IPH/ulceration from which that label follows directly. If that's the case, the expert branch is effectively reading the ground-truth label, and VMD becomes label distillation rather than transferring radiologists' domain expertise. That doesn't necessarily kill the method—privileged information can still help—but it changes the central claim substantially. The authors need to describe the reports and ideally show that the expert still helps when the report text is stripped of the conclusion.\n\nThe math in III-A is also shaky. Equation (3) has algebraic sign errors and an unjustified expansion; the jump to the ELBO in (4) doesn't follow; and equating I(S,E) to that ELBO in (5) is not justified. The final objective might still be a reasonable loss, but the paper's theoretical framing needs a corrected derivation or an explicit statement that the objective is heuristic.\n\nOther soft spots are smaller: single-center data, no code or data release, and the junior-doctor comparison uses only 27 images per doctor, so that statistical claim is weak. The authors do acknowledge the single-center and single-sequence limitations.\n\nAll that said, the setup is honest: ablations are consistent, baselines are reasonable, and the limitations section is candid. The paper deserves referee time, but it should not be accepted without clarification of the report content and a fix to the derivation. For now, I'd treat the empirical results as evidence for a label-distillation effect rather than demonstrated domain-knowledge transfer.","headline":"A clinically sensible distillation setup with consistent empirical gains, but the report modality's unspecified content may leak the label, and the variational derivation has real mathematical problems.","tokens_in":15427,"tokens_out":2329,"would_cite":false,"duration_ms":26935,"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":"VMD transfers radiology text and coarse annotation knowledge into a 3D CNN, improving unannotated carotid plaque MRI diagnosis from ROC 0.657 to 0.714.","keywords":["carotid plaque vulnerability","3D carotid MRI","multimodal knowledge distillation","variational inference","contrastive learning","radiology reports","student-teacher-expert network","medical image classification"],"falsifier":"Open the radiology reports and check whether they contain explicit terms like 'vulnerable plaque', 'stable plaque', or the component labels (IPH, LRC, thrombus, ulceration). Retrain VMD with every sentence that states or implies the diagnosis removed from the reports; if the ROC improvement over baseline largely disappears, the expert modality was providing label leakage, not independent domain knowledge.","tokens_in":14555,"feed_emoji":"🩺","tokens_out":9630,"duration_ms":90463,"temperature":0.7,"pith_summary":"This paper proposes VMD, a student-teacher-expert knowledge distillation strategy that trains a network to diagnose plaque vulnerability directly from unannotated 3D carotid MRI. The teacher branch consumes the same MRI multiplied by a cheap vessel-wall annotation; the expert branch consumes the radiology report text. VMD transfers knowledge from both by maximizing mutual information through contrastive learning and variational inference, and at inference only the student needs the raw MRI. The authors report ROC of 0.7136 versus 0.6567 for the fine-tuned baseline, with ablations showing each component helps. If the result holds, it would reduce annotation effort and give radiologists a fast, reasonably accurate second reader.","feed_headline":"Radiologist text distilled into MRI boosts plaque diagnosis","feed_subtitle":"Student-teacher-expert network raises ROC from 0.657 to 0.714 on unannotated carotid 3D MRI.","key_machinery":"The load-bearing machinery is the three-branch student-teacher-expert architecture combined with a dual mutual-information objective. The expert is a pretrained biomedical language model that encodes radiology reports; the teacher takes the element-wise product of the MRI volume with a limited vessel-wall annotation mask; the student sees only the raw MRI. The optimization objective, equation (1), maximizes I(S,T), I(S,E), and I(T,E) while minimizing classification cross-entropy. I(S,T) uses a label-aware infoNCE-style contrastive loss; I(S,E) uses a variational ELBO with reparameterization; I(T,E) combines both. This joint objective lets the student inherit text-derived diagnostic knowledge","core_discovery":"The central claim is that radiologists' domain knowledge, encoded in free-text reports, and anatomical priors, encoded in coarse vessel-wall annotations, can be transferred into a 3D convolutional student network so that plaque vulnerability is classified accurately without any annotation at inference time. The transfer is formulated as a mutual-information maximization: a contrastive term pulls same-class feature representations together across the teacher-student and teacher-expert pairs, and a variational term minimizes the KL divergence between the expert's latent posterior and the student's/teacher's latent priors. The paper presents experiments on an in-house dataset of 502 carotid 3D-","pith_inferences":["If the radiology reports contain explicit vulnerability statements, the reported gain may come from label leakage through the text encoder rather than from genuine domain expertise; redacting all diagnostic conclusion sentences before training would test this.","The same-class contrastive alignment assumes that vulnerable plaques look similar across patients; this assumption may weaken in multi-center data with different scanners, so external validation is a natural next test.","The dual mutual-information formulation is not specific to carotid MRI; it could transfer to other 3D medical classification tasks where coarse annotations and text reports coexist, such as prostate or liver MRI.","Because the student and teacher share a classification head, the architecture is unusually compact; probing how much of the gain comes from weight sharing versus the MI losses would clarify the mechanism."],"forward_implications":["Plaque vulnerability can be diagnosed directly from unannotated 3D carotid MRI, eliminating the need for pixel-level component segmentation at inference.","Rough vessel-wall annotations, which are far cheaper than full plaque-component labeling, can act as an effective teacher modality.","Radiology report text, routinely generated in clinical practice and usually discarded by automated MRI pipelines, can be recycled to improve visual classification.","The teacher network itself becomes a better annotator when constrained by the expert, suggesting cascaded distillation improves intermediate representations."],"fun_headline_variants":["Distilling radiologist notes into MRI sharpens plaque reads","Radiologist text enhances unannotated carotid MRI diagnosis","Text-distilled VMD lifts plaque vulnerability from MRI","Radiologist knowledge distilled into MRI improves plaque diagnosis"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The expert channel must carry diagnostic knowledge that is not already contained in the label; if the report explicitly says the plaque is vulnerable or stable, the student is effectively reading the label through the text encoder, and the claimed distillation of domain expertise collapses.","fun_headline_variants_meta":{"raw":{"variants":["Distilling radiologist notes into MRI sharpens plaque reads","Radiologist text enhances unannotated carotid MRI diagnosis","Text-distilled VMD lifts plaque vulnerability from MRI","Radiologist knowledge distilled into MRI improves plaque diagnosis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001014,"raw_usage":{"total_tokens":4096,"prompt_tokens":696,"completion_tokens":3400,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":440,"completion_tokens_details":{"reasoning_tokens":3347}},"tokens_in":440,"tokens_out":3400,"duration_ms":29633,"temperature":1.0,"reasoning_tokens":3347,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T16:40:11.268896+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Open the radiology reports and check whether they contain explicit terms like 'vulnerable plaque', 'stable plaque', or the component labels (IPH, LRC, thrombus, ulceration). Retrain VMD with every sentence that states or implies the diagnosis removed from the reports; if the ROC improvement over baseline largely disappears, the expert modality was providing label leakage, not independent domain knowledge.","supporting_citations":[],"review_version":1}