{"id":"1d561ac5-15b7-4dad-84f9-a42727ba7c5a","arxiv_id":"2502.00695","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A triple-modal network fusing CT, radiomics, and clinical text reaches 83.12% accuracy for chronic liver disease prognosis on a private 184-patient dataset, outperforming the compared baselines.","lead":"This paper introduces a neural network that combines CT scans, radiomic features, and clinical lab values to predict good or bad outcomes for chronic liver disease patients. The authors report that their three-way fusion model beats single-modality and simpler multi-modal baselines on a private 184-patient dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains may be inflated by tuning the TMFF weight lambda on the same folds used for evaluation; no nested validation or significance testing is provided.","rationale":"I agree with the reader's identification of the evaluation protocol as the load-bearing weakness. The paper's own ablations (Table 4) are the clearest evidence: a single hyperparameter lambda is swept over four values and the best CV result (0.6) is reported as the method's performance. Because the same 5-fold CV metrics are used both to select lambda and to report the final numbers, the reported accuracy and AUC are optimistically biased. The effect is amplified by the small cohort (184 patients) and the absence of error bars or significance tests; the largest reported gain over the best competing method (AUC +0.0208) is plausibly within fold-to-fold noise. A secondary inconsistency (baseline without TMFF is 76.61% in Table 4 vs 75.48% in Table 3) further muddies the ablation story. The architectural components (IMA self-attention, cyclic TCAF, pairwise SDM loss) are reasonable and the public code link is a plus, but the central 'significantly outperforms' claim cannot be accepted on the evidence as presented. The proposed nested-CV check directly addresses whether the advantage survives honest model selection.","tokens_in":6816,"tokens_out":4301,"duration_ms":40628,"concrete_test":"Reproduce the full pipeline with a nested 5-fold CV: for each outer test fold, select lambda (and alpha, if also tuned) using only the training folds, e.g., via an inner 4-fold CV, then evaluate on the held-out outer fold; report mean±std over outer folds for ACC/F1/AUC. Also run a paired permutation test (or DeLong test) comparing TMI-CLNet vs HFBSurv on the same folds. If the nested-CV AUC advantage is ≤0.02 or the 95% CI includes 0, the claimed superiority is not supported. Additionally, verify the public GitHub repository reproduces Table 1 exactly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim ('significantly outperforms...') rests on Table 1's metrics (83.12% ACC, 0.7805 F1, 0.8223 AUC) computed via 5-fold cross-validation on a private 184-patient cohort. The weakest point is that the TMFF loss weight lambda is tuned on these same folds: Section 3.2 (Table 4) sweeps lambda over {0.2, 0.4, 0.6, 0.8} and selects 0.6 because it yields the best CV accuracy/AUC, and the paper then reports those selected-fold numbers as the final result. This is a classic 'selection on the test set' protocol. With only five folds, the variance is high (e.g., binomial standard error for AUC at n=184 is about 0.03), so the reported advantage over the best baseline (HFBSurv) of 0.0208 AUC and 3.67% accuracy could easily be within noise once selection bias is accounted for. The paper provides no per-fold results, no error bars, and no significance tests. Also, the baseline in Table 3 (no IMA/TCAF: 75.48% ACC) differs from Table 4 (no TMFF: 76.61% ACC), so the ablation baselines are not consistent, making the component-wise gains hard to interpret. Unless a nested or held-out evaluation is provided, the headline claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TMI-CLNet, a triple-modal network that fuses CT images, radiomic features, and clinical information for chronic liver disease prognosis. The architecture consists of an Intra-Modality Aggregation (IMA) module based on multi-head self-attention, a Triple-modal Cross-Attention Fusion (TCAF) module with cyclic one-direction cross-attention among the three modalities, and a Triple-Modal Feature Fusion (TMFF) loss adapted from the Similarity Distribution Matching loss. On a private 184-patient liver prognosis dataset, the method reports 83.12% accuracy, 0.7805 F1, and 0.8223 AUC, outperforming several unimodal and multimodal baselines. The paper includes ablations on modality configuration, the IMA/TCAF modules, and the TMFF loss weight lambda, plus Grad-CAM visualizations.","tokens_in":7224,"tokens_out":2374,"duration_ms":26754,"significance":"If the reported gains are reliable, the combination of within-modality aggregation, cyclic cross-attention, and pairwise distribution-matching alignment would be a useful template for triple-modal medical prognosis. The paper provides code, runs all comparisons on the same protocol, and includes multiple ablation studies, which are strengths. The architecture itself is internally consistent and the loss derivation in Eq. (5) is a straightforward extension of the cited SDM loss, so there is no circular reasoning in the method. However, the central empirical claim of significant superiority is not currently established because the evaluation protocol selects the TMFF weight on the same folds that produce the final metrics, reports no error bars or statistical tests, and uses a single small private cohort. These issues are load-bearing for the paper's main contribution and need to be addressed before the claim can be accepted.","major_comments":[{"comment":"The TMFF loss weight lambda is tuned on the same 5-fold cross-validation used to report the final results. Table 4 sweeps lambda over {0.2, 0.4, 0.6, 0.8}, selects 0.6 because it gives the best accuracy and AUC, and then those selected-fold numbers are presented as the final performance. This is a selection-on-the-test-set protocol that can inflate the reported gains. Please provide a nested or held-out validation procedure, or clearly separate model selection from evaluation, and report per-fold results with confidence intervals. This is essential for the headline claim of 83.12% accuracy and 0.8223 AUC.","section":"Section 3.2, Table 4 and Eq. (5)"},{"comment":"The statement that TMI-CLNet 'significantly outperforms existing state-of-the-art' methods is not supported by any statistical analysis. The dataset has only 184 patients, the reported improvements over HFBSurv are 3.67 percentage points in accuracy and 0.0208 in AUC, and no standard errors, confidence intervals, or significance tests are provided. With five folds and this sample size, the AUC difference is within the range of plausible sampling noise. Please add statistical testing (e.g., DeLong test for AUC differences, McNemar for accuracy), per-fold results, and error bars, or soften the significance claim accordingly.","section":"Section 3.2, Table 1 and abstract"},{"comment":"The ablation baselines are inconsistent between the two tables. Table 3 reports that removing both IMA and TCAF yields ACC 75.48%, whereas Table 4 reports that removing TMFF loss yields ACC 76.61%. If both rows correspond to the same configuration (full IMA+TCAF without the TMFF loss), the discrepancy of 1.13 percentage points needs an explanation. If the configurations differ in another way, that difference should be stated explicitly. As written, the component-wise gains attributed to the modules versus the loss are not cleanly separable.","section":"Section 3.2, Tables 3 and 4"},{"comment":"External validity is a central concern. All conclusions rest on a single private 184-patient cohort with 5-fold cross-validation and no held-out test set or multi-center validation. The authors themselves list multi-center validation as future work in Section 4, but the abstract and Section 3.2 state that the method 'significantly outperforms' alternatives without acknowledging this limitation. Please either add external validation or substantially temper the generalizability claims in the abstract and conclusion.","section":"Section 3.1 and Section 4"}],"minor_comments":[{"comment":"There are typographical and grammatical issues, including the capitalized 'We' after 'However' in the abstract and 'infomation' in the conclusion. A careful proofread is needed.","section":"Abstract and Section 4"},{"comment":"The clinical information is described as 'clinical text' in the methodology, but the dataset description lists structured numerical data (gender, height, BMI, CBC, and fat analysis). Please clarify how this structured information is converted into the text modality and what the BioBERT input looks like.","section":"Section 3.1, Dataset description"},{"comment":"The classification head is described as a 'pre-trained 1D DenseNet-121,' but no details are given about its input dimension, output dimension, or how the fused global feature vector is reshaped for this head. A short explanation or architecture diagram would improve reproducibility.","section":"Section 3.1, Implementation Details"},{"comment":"The paper states that alpha in Eq. (6) is set to 1, but no ablation or sensitivity analysis for alpha is reported. Since the TMFF loss weight lambda is heavily analyzed, a brief note on alpha would complete the hyperparameter discussion.","section":"Section 2.3, Eq. (5)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable engineering contribution with a well-structured architecture, but the evaluation protocol has a selection-bias issue that directly affects the central 'significantly outperforms' claim. I do not see grounds for rejection, because the problems are fixable within the manuscript's scope: nested validation, statistical significance tests, per-fold results, and a more cautious wording of the claims. I would also note that the code repository may help reviewers verify the exact protocol, but I did not inspect it in this review. The fit for a computer vision venue is acceptable given the multimodal fusion focus."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nTMI-CLNet is a reasonable engineering contribution: it fuses CT, radiomics, and clinical text for liver prognosis using cyclic triple-modal cross-attention and an alignment loss adapted from SDM. The components are familiar, but the specific arrangement is new to this task, and the authors include ablations, code, and a small XGradCAM analysis. None of that is groundbreaking, but it is a solid template for triple-modal medical fusion.\n\nThe problem is the evidence for the headline claim. The reported 83.12% accuracy and 0.8223 AUC come from 5-fold cross-validation on 184 private patients, and the TMFF weight lambda was chosen by sweeping {0.2, 0.4, 0.6, 0.8} and keeping the value that gave the best CV numbers. That is selection on the test set. With five folds, the variance is high enough that a 3.67-point accuracy gain and 0.0208 AUC advantage over HFBSurv could easily be noise. The paper reports no per-fold results, no error bars, no significance tests. It also has an internal inconsistency: the no-IMA/no-TCAF baseline in Table 3 is 75.48% accuracy, while the no-TMFF baseline in Table 4 is 76.61% — same network minus different modules should give the same number. That makes the component-wise ablations hard to interpret.\n\nThe architecture itself is not circular; the TMFF loss is a direct adaptation of an external SDM loss, so the novelty claim is honest. The self-citations to the authors' prior cross-modal work are appropriate context, not padding. The real soft spot is entirely in the evaluation protocol.\n\nThis paper is for readers working on multimodal medical fusion who want a concrete example of cyclic cross-attention and an SDM-style alignment loss. It deserves a serious referee, but only if the referee insists on a nested validation or a held-out test set, plus significance testing. As is, the central claim is not established. I'd send it back for revision rather than accept.\n\nRecommendation: engage with it as a conditional — the method is worth a look, but the numbers need to be re-derived under a clean protocol.","headline":"A sensible triple-modal fusion architecture for liver prognosis whose reported gains are likely inflated by tuning a loss weight on the same cross-validation folds; the method deserves a look, but the evaluation needs to be redone.","tokens_in":7638,"tokens_out":1918,"would_cite":false,"duration_ms":19015,"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":"The paper claims a triple-modal network fusing CT images, radiomic features, and clinical data predicts chronic liver disease prognosis with 83.12% accuracy, outperforming existing unimodal and multimodal baselines.","keywords":["multi-modal learning","cross attention","radiomics","chronic liver disease","prognosis prediction","CT imaging","feature fusion","medical image analysis"],"falsifier":"Re-run the released code on the same 184 patients but with the $\\lambda$ values chosen on a held-out validation fold or an external cohort; if TMI-CLNet no longer beats HFBSurv by roughly 3.67 percentage points in accuracy and 0.0208 in AUC, the paper's central claim would be refuted.","tokens_in":6608,"feed_emoji":"🩺","tokens_out":6072,"duration_ms":52564,"temperature":0.7,"pith_summary":"The paper sets out to show that chronic liver disease prognosis improves when three heterogeneous data sources—CT images, handcrafted radiomic features, and clinical laboratory/text information—are fused in a single network rather than used alone or combined by simple concatenation. It proposes TMI-CLNet, whose central bet is that fusion fails unless the network first cleans each modality internally, then lets each modality read the adjacent modality's key-value content cyclically, and finally aligns all three feature spaces with a dedicated loss. On a private cohort of 184 patients, the network reports 83.12% accuracy, 0.7805 F1, and 0.8223 AUC, outperforming the best compared method by 3.67 percentage points in accuracy, 0.0572 in F1, and 0.0208 in AUC. A sympathetic reader would care because accurate prognosis guides early treatment, and the architecture is offered as a template for other triple-modal medical tasks.","feed_headline":"Triple-modal fusion hits 83.12% accuracy for liver prognosis","feed_subtitle":"Combining CT images, clinical labs, and radiomic features beats single-modality baselines on a 184-patient cohort.","key_machinery":"The load-bearing machinery is a three-part interaction loop. The Intra-Modality Aggregation (IMA) module is a 16-head multi-head self-attention block that consolidates each modality's own features before fusion. The Triple-Modal Cross-Attention Fusion (TCAF) module computes, for each modality $i$, cross-attention against the adjacent modality $j = i \\bmod 3 + 1$ through $F^i_{\\text{hidden}} = \\operatorname{softmax}(Q^i K^{j\\top}/\\sqrt{d_k}) V^j$, concatenates the attended output with the original modality features, and concatenates the three results into a global representation. The Triple-Modal Feature Fusion (TMFF) loss is a weighted sum of pairwise Similarity Distribution Matching losses, $L_{\\text{multi}} = \\lambda (L^{it}_{\\text{SDM}} + L^{rt}_{\\text{SDM}})/2 + (1-\\lambda) L^{ir}_{\\text{SDM}}$, added to the cross-entropy task loss. The claim is that these three components together remove intra-modality redundancy, capture cross-modal relationships, and align the three feature spaces so that a 1D DenseNet-121 classification head can predict good versus bad prognosis.","core_discovery":"The paper's central claim is that TMI-CLNet, a triple-modal network for CT imaging, radiomic features, and clinical information, achieves the best prognosis classification on its liver disease dataset, and that each of its three interaction components contributes to the gain. The Intra-Modality Aggregation module uses 16-head self-attention to reduce redundancy inside each modality; the Triple-Modal Cross-Attention Fusion module computes, for each modality, attention over the adjacent modality's key-value pairs in a cycle and concatenates the attended features with the original modality features; and the Triple-Modal Feature Fusion loss aligns image-text, radiomics-text, and image-radiomics pairwise similarity distributions. In the head-to-head comparison, TMI-CLNet reaches 83.12% accuracy, 84.38% precision, 74.28% recall, 0.7805 F1, and 0.8223 AUC, exceeding every unimodal and multimodal baseline on all metrics. The ablation results attribute the largest gains to using IMA and TCAF together (accuracy increase of 7.64 percentage points) and to the TMFF loss with weight $\\lambda = 0.6$ (accuracy increase of 6.45 percentage points and AUC increase of 0.0687 over no TMFF loss).","pith_inferences":["Editorial inference: the core design cycles one step to the next modality rather than computing all pairwise attention maps, reducing the number of cross-modal interactions from six to three; on tasks with more modalities this linear cycle may scale better, but whether it loses information versus full pairwise attention is untested here.","Editorial inference: the reported improvements should be re-checked with nested cross-validation or a separate test fold, because the $\\lambda$ sweep and model selection happen on the same five folds; a reader can test this by re-running the released code with a fixed hold-out split.","Editorial inference: the framework likely transfers to other prognosis tasks with the same data triad of imaging, quantitative features, and clinical text, such as other organ cancers, but the 184-patient single-center cohort makes the effect size uncertain until multi-center validation appears."],"forward_implications":["If TMI-CLNet's reported gains are real, fusing CT, radiomics, and clinical data is better than any single modality for this prognosis task, with recall improving from 64.32% for the best baseline to 74.28%.","The ablation results imply that neither cross-attention nor self-attention alone is enough: adding both IMA and TCAF raises accuracy by 7.64 percentage points over the no-module baseline.","The TMFF loss contributes independently: setting $\\lambda = 0.6$ raises accuracy by 6.45 percentage points and AUC by 0.0687 over omitting the loss.","Because the method outperforms HFBSurv and MMD, which were designed for heterogeneous and incomplete multimodal medical data, the architecture is a plausible template for other three-modality medical prediction tasks."],"supporting_citations":[{"why":"Supplies the pretrained 3D ResNet-50 used as the CT image encoder.","marker":"[5]"},{"why":"Supplies the pretrained BioBERT model used to encode clinical text.","marker":"[6]"},{"why":"Provides the Similarity Distribution Matching loss that the TMFF loss adapts from two modalities to three.","marker":"[7]"},{"why":"Serves as the sequential image-plus-radiomics fusion baseline.","marker":"[12]"},{"why":"Is the strongest prior multimodal baseline whose accuracy, F1, and AUC TMI-CLNet claims to beat.","marker":"[13]"},{"why":"Is the universal multimodal fusion baseline designed for incomplete modalities, used for comparison.","marker":"[14]"}],"fun_headline_variants":["Triple-modal network hits 83.12% for liver prognosis","Fusing CT, labs, radiomics improves liver outcome prediction","TMI-CLNet outperforms single-modality models for liver disease","Three-way data fusion lifts liver prognosis accuracy to 83%","Cross-attention fusion boosts chronic liver disease prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole reported advantage rests on the assumption that 5-fold cross-validation on a single private 184-patient cohort, with the TMFF weight $\\lambda$ chosen by sweeping on those same folds, gives an unbiased estimate of the model's true predictive advantage.","fun_headline_variants_meta":{"raw":{"variants":["Triple-modal network hits 83.12% for liver prognosis","Fusing CT, labs, radiomics improves liver outcome prediction","TMI-CLNet outperforms single-modality models for liver disease","Three-way data fusion lifts liver prognosis accuracy to 83%","Cross-attention fusion boosts chronic liver disease prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000171,"raw_usage":{"total_tokens":1310,"prompt_tokens":1020,"completion_tokens":290,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":204}},"tokens_in":636,"tokens_out":290,"duration_ms":3484,"temperature":1.0,"reasoning_tokens":204,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T18:01:37.725500+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the released code on the same 184 patients but with the $\\lambda$ values chosen on a held-out validation fold or an external cohort; if TMI-CLNet no longer beats HFBSurv by roughly 3.67 percentage points in accuracy and 0.0208 in AUC, the paper's central claim would be refuted.","supporting_citations":[{"cited_title":"Approval was granted by the Ethics Committee of Longgang Central Hospital of Shenzhen (2024.5.8/No.2024052)","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained 3D ResNet-50 used as the CT image encoder."},{"cited_title":"2023A1515012587)","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained BioBERT model used to encode clinical text."},{"cited_title":"An update on viral hepatitis b and c in mexico: Advances and pitfalls in eradication strategies,","cited_arxiv_id":null,"evidence_quote":"Provides the Similarity Distribution Matching loss that the TMFF loss adapts from two modalities to three."},{"cited_title":"Biobert: a pre-trained biomedical language representa- tion model for biomedical text mining,","cited_arxiv_id":null,"evidence_quote":"Serves as the sequential image-plus-radiomics fusion baseline."},{"cited_title":"Cross-modal implicit rela- tion reasoning and aligning for text-to-image person re- trieval,","cited_arxiv_id":null,"evidence_quote":"Is the strongest prior multimodal baseline whose accuracy, F1, and AUC TMI-CLNet claims to beat."},{"cited_title":"Deep residual learning for image recognition,","cited_arxiv_id":null,"evidence_quote":"Is the universal multimodal fusion baseline designed for incomplete modalities, used for comparison."}],"review_version":1}