{"id":"ba0cab72-7ac9-4767-bdcd-5ee6d735f441","arxiv_id":"2411.10894","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Adding BI-RADS lesion descriptors as text input to a dual-view iterative attention model raises benign/malignant classification AUC on CBIS-DDSM from 0.711 to 0.872, outperforming a multi-view transformer baseline.","lead":"This paper shows that adding radiologists' written BI-RADS lesion descriptors to mammogram images improves automated cancer classification, raising AUC from 0.71 to 0.87 on a public dataset. The result matters because it suggests that structured clinical text can be directly fused with images in screening AI, though the gain comes from the text, not better image analysis.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-validation may be at image level rather than patient level; same-patient leakage could fully explain the reported 0.872 vs 0.711 AUC gap.","rationale":"The paper's strongest claim is that adding BI-RADS descriptor text to a dual-view attention model raises AUC from 0.711 to 0.872 on CBIS-DDSM. The architecture is plausible, and the no-descriptor variant is the right control. The most load-bearing threat is experimental validity: the paper never says the five folds are patient-disjoint. Because CBIS-DDSM contains roughly 2.3 abnormalities per patient and two standard views per abnormality, an image-level split creates direct train/test overlap. This would inflate both models, but especially the multimodal one, since the same descriptor vector and the same lesion can appear in both folds. The reader's weakest_assumption identifies exactly this issue, and I agree. I would not move to reject because the authors may have used patient-level grouping and simply omitted the detail; the proposed check settles it. Secondary issues—test-set-driven selection of configurations and augmentations in Tables 3-5, no error bars in Table 2, and the internal 0.872 versus 0.878 AUC discrepancy with the claimed best configuration—all reinforce the need for released splits and code, but they are less decisive than the leakage risk.","tokens_in":8836,"tokens_out":5325,"duration_ms":60847,"concrete_test":"Obtain the exact five-fold split files or splitting code; verify no patient ID occurs in both train and test. If unavailable, re-run with GroupKFold over patient_id, same hyperparameters, both with and without descriptors. Compare patient-level AUC to 0.872/0.711; a substantial drop invalidates the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 states only \"five-fold stratified cross-validation to maintain class balance across folds.\" No patient-level grouping is mentioned. CBIS-DDSM contains 1,566 patients but 3,568 abnormalities, so many patients contribute multiple lesions and both CC and MLO views. If folds are stratified at the image or abnormality level, the same lesion's CC/MLO pair and the same BI-RADS descriptor vector (Section 3.1) can appear in both training and test folds. The model's view-attention layer explicitly fuses the two views (Section 3.3), so under image-level splitting the test fold can contain a lesion whose cross-view appearance and descriptor text were already seen in training. That would make the 0.872-vs-0.711 comparison a test of memorization rather than generalization. No code, fold assignments, or split files are released, so the reported result is not independently verifiable as stated. This issue also compromises the comparison with the multi-view baseline [23], because all models may share the same leaked splits. The central claim therefore depends on an unstated assumption that the five folds are disjoint at the patient level.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Deep BI-RADS, a dual-branch multi-attention architecture that fuses CC/MLO mammogram views with textual BI-RADS lesion descriptors for benign/malignant classification on CBIS-DDSM. The model uses iterative cross-, self-, and view-attention layers, and the authors report that adding BI-RADS descriptors raises AUC from 0.711 to 0.872. The central claim is that expert-documented BI-RADS descriptors, combined with visual features, materially improve cancer detection over image-only models.","tokens_in":8971,"tokens_out":4205,"duration_ms":41460,"significance":"The research question—whether BI-RADS descriptors, which radiologists already record, can complement image-only deep learning—is clinically relevant. The ablation comparing the same architecture with and without descriptors is a sound way to isolate the descriptor contribution, provided the evaluation protocol is unbiased. If the reported gains withstand patient-level cross-validation and statistical testing, the paper would provide a useful empirical result for multi-modal mammography analysis. The manuscript does not ship code, fold assignments, or exact values for several key hyperparameters, which limits verification; these omissions should be addressed.","major_comments":[{"comment":"The paper states only \"five-fold stratified cross-validation to maintain class balance across folds\" and does not specify whether the folds are disjoint at the patient level. CBIS-DDSM contains 1,566 patients but 3,568 abnormalities, so multiple lesions and both CC/MLO views of the same patient can appear in both training and test folds if the split is image- or lesion-level. Because the view-attention sub-layer (Section 3.3) explicitly fuses CC/MLO features, such leakage could inflate the reported AUC of 0.872 and make the comparison with the image-only variant a test of memorization rather than generalization. The authors must state and justify patient-level disjoint folds, or release the exact fold assignment files.","section":"Section 4.1 and Table 2"},{"comment":"The central result—AUC 0.872 versus 0.711 without descriptors—is reported as a single number with no error bars, confidence intervals, or significance test, whereas Tables 4 and 5 report mean ± std across folds. Given the small number of training iterations (1000) and the absence of per-fold values, the improvement cannot be assessed for statistical reliability. The authors should report per-fold results, the standard deviation or confidence interval for each metric, and an appropriate statistical comparison (e.g., DeLong test for AUC differences).","section":"Table 2 and Section 5"},{"comment":"Several hyperparameters required to reproduce the model are missing: the Fourier positional encoding parameters nbands and mfreq in Eq. (9) are never given numeric values, the latent query length NQ is only described as \"set by the user,\" and the latent vector length L is stated as 256 in Section 3.1 but its role in the attention layers is not fully specified. Without these values, or released code, the reported results cannot be independently reproduced.","section":"Section 3.4, Section 4.2, and Section 3.3"},{"comment":"The multi-view baseline [23] and the single-view baseline [15] are problematic as stated. References [23] and [24] are the same paper (Tulder et al.), so the baseline set is smaller than claimed. Reference [15] (Hover-Trans) is described in the bibliography as an ultrasound-image method, yet Section 5 calls it \"an advanced recent single-view Transformer approach\" for mammograms without explaining how it was adapted, if at all. The comparison should be corrected or the baselines replaced with appropriate mammogram methods.","section":"Section 5 and References"}],"minor_comments":[{"comment":"Reference [2] lists the TI-RADS Atlas with year \"2073,\" which is a typo and should be corrected.","section":"References"},{"comment":"The citation placeholder \"[?,6,28]\" in the first paragraph should be replaced with the actual reference numbers.","section":"Section 1"},{"comment":"The symbol N is used both for the number of descriptor classes in Section 3.1 and for the number of multi-attention layers in the architecture, which is confusing and should be disambiguated.","section":"Section 3.1 and Section 3.3"},{"comment":"The definition of the frequency bands Sb is unclear: b is an integer, but S is a real number, and the notation \"1 ≤ b < nbands\" with b ∈ Z+ is inconsistent with the superscript on Sb. Please clarify the indexing and define mfreq and nbands.","section":"Equation (9)"},{"comment":"The formula for Nk = H'·W' / 2(n−k−1) uses n without defining it; presumably n is the number of layers N. Please make the dependence on N explicit.","section":"Section 3.3"},{"comment":"The ROC curve is described as being for a single fold; specify which fold is shown, or plot all folds with the average curve.","section":"Figure 2"},{"comment":"The rows \"Baseline + 384,\" \"Baseline + 1024,\" and \"Baseline + 2048\" are not clearly labeled: the caption refers to interpolation upsampling, but the text should state that these are resized image dimensions and whether both training and test images are resized.","section":"Table 5"},{"comment":"The sentence \"fromscratch\" in \"trained all models fromscratch following their respective provided training protocols\" has a missing space, and the protocols themselves are not described for the baselines; please provide details or citations.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The central ablation is well designed, but the missing patient-level split description and the lack of error bars on the main result are serious enough to require re-review after revision. I would encourage the editor to ask for the actual fold splits or code, as the reported 0.872 vs 0.711 AUC gap is large and the leakage risk is concrete given the CBIS-DDSM patient-to-abnormality ratio. The paper's scope is a modest empirical contribution; it would be strengthened by a clearer statement about when BI-RADS descriptors are available at test time in practice."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper asks whether BI-RADS descriptor text, which radiologists already record, can improve automated benign/malignant classification when fused with mammogram images. The answer inside the paper is a 0.872 vs 0.711 AUC gain on CBIS-DDSM. That is a legitimate question, and the architecture is a reasonable adaptation of Perceiver-style cross-attention with a dual-view branch. The within-paper ablation (with vs without descriptors) is the right experiment to run, and the reported gains across all metrics make the claim look coherent.\n\nThe problem is that the validation, as reported, cannot be trusted yet. The five-fold stratified cross-validation is not stated to be patient-level. CBIS-DDSM has 1,566 patients and 3,568 abnormalities; many patients contribute multiple lesions and both CC and MLO views. If folds are split at the image or abnormality level, the same lesion's paired views and its descriptor vector can appear in both training and test. The view-attention layer explicitly fuses the two views, so under image-level splitting the model could be memorizing rather than generalizing. That would fully explain the reported gap. No code, split files, or fold assignments are released, so this is not independently verifiable. The stress-test note is on target: this is a load-bearing ambiguity, not a minor omission.\n\nThere are also smaller issues. No error bars or significance tests are given for the main Table 2 numbers, and Table 3 reports AUC 0.878 for the configuration that Table 2 says is 0.872. The baseline list includes an ultrasound model (ref 15) used as a mammogram baseline, and references 23 and 24 are duplicates. Hyperparameters for the Fourier positional encoding bands are incomplete.\n\nOn the positive side, the descriptors are external expert annotations and the labels are biopsy pathology, so there is no circularity in the label derivation. Using BI-RADS descriptors as input rather than output is not common, and the paper is straightforward about the fact that the gain comes from expert text, not from the images alone.\n\nThe fix is straightforward: report patient-level splits, error bars over folds, and release code. If the result survives that, it is a solid contribution for the medical imaging community. As it stands, the central claim rests on an unstated assumption about the splitting. This is a classic leakage risk, and the authors need to address it before the numbers mean anything.\n\nWho gets value from this? Researchers working on multimodal medical imaging, especially those who want to leverage structured radiology reports. It deserves peer review, but only if reviewers push hard for the split details, code, and error bars.","headline":"A plausible and useful idea—BI-RADS descriptors as privileged inputs—but the main result is unverifiable until patient-level cross-validation is confirmed and code is released.","tokens_in":9582,"tokens_out":2083,"would_cite":false,"duration_ms":21406,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Fusing radiologist BI-RADS text with mammograms lifts cancer-detection AUC to 0.872","keywords":["Cancer Detection","BI-RADS","Deep Learning","Mammograms","Breast Cancer","Attention","Transformer","Multi-Modal"],"falsifier":"Re-run the same training with patient-level five-fold splits, assigning all CC/MLO views and all lesions of each patient to the same fold, and compare the AUC with the paper's 0.872; a large drop, or a shrinkage of the gap over the image-only 0.711, would show that part of the reported benefit came from patient leakage.","tokens_in":8550,"feed_emoji":"🩺","tokens_out":8840,"duration_ms":80087,"temperature":0.7,"pith_summary":"This paper asks whether deep learning breast-cancer screening can be improved by feeding the model the same structured lesion descriptions that radiologists already write down. The answer it argues for is yes: on the CBIS-DDSM dataset, a dual-branch model that combines CC and MLO mammogram images with BI-RADS lesion descriptors reaches AUC 0.872 for benign-versus-malignant classification, versus 0.711 for the same model without the descriptors and 0.811 for a multi-view transformer baseline. The gain appears across accuracy, specificity, precision, recall, and F1 as well. The practical point is that an inexpensive, already-recorded textual signal can be fused with images instead of being discarded or predicted from pixels.","feed_headline":"BI-RADS text plus mammogram images lifts cancer-detection AUC to 0.872","feed_subtitle":"On CBIS-DDSM, adding radiologist-written lesion descriptors beats image-only models (AUC 0.872 vs 0.711).","key_machinery":"The mechanism is a dual-branch stack of six multi-attention layers, one branch per mammogram view (CC and MLO). Inside each layer, three attention sub-layers run in sequence: cross-attention fuses the encoded BI-RADS descriptors, a binary vector over descriptor classes such as margin, shape, and distribution, with the current multi-resolution visual features from a Big Transfer backbone; self-attention spreads context within each view; and view-attention swaps query and key information between the two branches, with the value coming from the current view. The best-performing wiring takes the query and keys from the opposite view, which the paper reports as configuration 2.","core_discovery":"The paper's central claim is that radiologist-authored BI-RADS descriptors carry information about lesion malignancy that is not fully recoverable from the mammogram pixels alone, and that an attention-based fusion architecture can extract that information. The experiments on CBIS-DDSM mass lesions show the descriptor-fused model at AUC 0.872, the descriptor-free variant at 0.711, and a multi-view transformer baseline at 0.811, with corresponding improvements in accuracy, specificity, precision, recall, and F1. The authors interpret this as evidence that handcrafted clinical features and end-to-end visual features are complementary, and that a model need not choose between them.","pith_inferences":["A natural next experiment is to parse full free-text radiology reports instead of the curated CBIS-DDSM descriptor lists; the binary-vector encoding used here may discard report nuance, and a text-encoder comparison would test whether the remaining information matters.","The model was evaluated on biopsied lesions, which are enriched for malignancy compared with a general screening population; a screening cohort would be a sterner test of whether the descriptor gain persists at lower cancer prevalence.","The patient-level split question is testable directly from the fold assignment; if folds are not patient-exclusive, the 0.872 figure would need to be re-estimated before the method is compared with prior work.","Because the descriptors are available at interpretation time, the same fusion idea could be transferred to other structured reporting lexicons, for example TI-RADS for thyroid nodules, to see whether the benefit generalizes."],"forward_implications":["If the result holds, the descriptors need no new imaging or workflow to obtain, because radiologists already record them during interpretation.","Because the model accepts a variable number of lesion-descriptor vectors, a single mammogram containing several lesions can be classified in one forward pass without collapsing the lesions into one summary.","The descriptor input excludes the radiologist's final BI-RADS suspicion score and uses only the physical lexicon, so the reported gain cannot be explained merely by the model reading the verdict off the report.","The ablation around view-attention wiring suggests the cross-view direction matters: query and keys from the opposite view outperform same-view wiring, pointing to a concrete design rule for similar multi-view fusion models."],"supporting_citations":[{"why":"Supplies the CBIS-DDSM mammograms and the BI-RADS descriptor metadata that the multi-modal model consumes.","marker":"[12]"},{"why":"Defines the BI-RADS descriptor vocabulary that the binary input vector enumerates.","marker":"[1]"},{"why":"Provides the iterative cross-attention with latent bottleneck that the multi-attention layers are built on.","marker":"[10]"},{"why":"Provides the Big Transfer feature extractor whose multi-resolution outputs feed the attention layers.","marker":"[11]"},{"why":"Supplies the PatchCamelyon pre-training data for the feature extractor.","marker":"[26]"},{"why":"Serves as the multi-view transformer baseline whose 0.811 AUC is compared against.","marker":"[23]"},{"why":"Gives the scaled dot-product attention and Q/K/V formulation used by all attention sub-layers.","marker":"[25]"}],"fun_headline_variants":["Radiologists' BI-RADS text plus mammograms lifts AUC to 0.872","Fusing BI-RADS descriptors with mammograms improves detection to AUC 0.872","Mammograms + radiologist BI-RADS text: AUC 0.872 vs 0.711","Attention-based fusion of BI-RADS and mammograms beats image-only models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gain rests on the unstated assumption that the five-fold cross-validation keeps every patient's images in a single fold, so the model cannot inflate its score by memorizing a patient whose lesions already appeared in training.","fun_headline_variants_meta":{"raw":{"variants":["Radiologists' BI-RADS text plus mammograms lifts AUC to 0.872","Fusing BI-RADS descriptors with mammograms improves detection to AUC 0.872","Mammograms + radiologist BI-RADS text: AUC 0.872 vs 0.711","Attention-based fusion of BI-RADS and mammograms beats image-only models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001116,"raw_usage":{"total_tokens":4576,"prompt_tokens":805,"completion_tokens":3771,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":421,"completion_tokens_details":{"reasoning_tokens":3678}},"tokens_in":421,"tokens_out":3771,"duration_ms":28548,"temperature":1.0,"reasoning_tokens":3678,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:10:28.058405+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same training with patient-level five-fold splits, assigning all CC/MLO views and all lesions of each patient to the same fold, and compare the AUC with the paper's 0.872; a large drop, or a shrinkage of the gap over the image-only 0.711, would show that part of the reported benefit came from patient leakage.","supporting_citations":[{"cited_title":"Amer- ican College of Radiology, Reston, VA (2013)","cited_arxiv_id":null,"evidence_quote":"Defines the BI-RADS descriptor vocabulary that the binary input vector enumerates."},{"cited_title":"In: Computer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Pro- ceedings, Part V 16","cited_arxiv_id":null,"evidence_quote":"Provides the Big Transfer feature extractor whose multi-resolution outputs feed the attention layers."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the PatchCamelyon pre-training data for the feature extractor."},{"cited_title":"In: International Conference on Medical Im- age Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Serves as the multi-view transformer baseline whose 0.811 AUC is compared against."},{"cited_title":"Advances in Neural Information Pro- cessing Systems30, 5998–6008 (2017)","cited_arxiv_id":null,"evidence_quote":"Gives the scaled dot-product attention and Q/K/V formulation used by all attention sub-layers."}],"review_version":1}