{"id":"57dd6bcd-be8d-430a-a96b-b7313c45b42b","arxiv_id":"2505.03838","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"IntelliCardiac is a web platform that segments cardiac MRI structures with a 3D U-Net and classifies five cardiac conditions with a random forest plus SVM pipeline, reporting 92.6% Dice and 98% accuracy on ACDC.","lead":"IntelliCardiac is a web-based tool that automatically outlines heart chambers in cardiac MRI scans and then labels each scan as one of five heart conditions. The authors report 92.6% average segmentation accuracy and 98% disease classification accuracy on a public benchmark, and the tool is designed for patients and doctors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 98% classification accuracy is at risk of test-set-guided refinement; the paper does not show a validation split and describes the SVM's two MWT features inconsistently.","rationale":"The paper's central claim is that IntelliCardiac jointly segments and classifies cardiac MRI with 92.6% Dice and 98% classification accuracy, outperforming prior integrated methods. For that claim to hold, the classification accuracy must be a genuinely held-out result. The reader's weakest assumption identifies exactly the load-bearing risk: the two-stage classifier's refinement design may have been selected using the 50-patient test set. My reading agrees, and I find additional supporting evidence in the paper itself: the two MWT features used by the advanced SVM are described inconsistently in Section III-C2 versus Section IV-B, and no validation split or feature-selection protocol is reported. The small test set (50 patients) makes the difference between 92% and 98% just three corrected errors, so the result is highly sensitive to any tuning that used test-set information. The segmentation numbers, the web platform, and the ablation study are useful engineering contributions and are not the focus of this concern. A single concrete check, namely retuning the classifier only on a held-out validation split before evaluating on the 50 test patients, would settle whether the 98% result is reproducible as a true hold-out. Until then, the appropriate verdict is conditional, consistent with the reader's assessment.","tokens_in":9578,"tokens_out":4920,"duration_ms":61202,"concrete_test":"Hold out a validation set from the 100 training patients (e.g., 20 patients) before any classifier design. Use only that split to select the two MWT features, the SVM kernel and hyperparameters, and the decision to refine only initial MINF/DCM predictions; then evaluate the fixed pipeline on the untouched 50-patient ACDC test set. If the resulting accuracy drops materially below 98%, the reported number reflects test-set-guided selection rather than a held-out measurement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central 'outperforming existing state-of-the-art' claim rests on a 98% classification accuracy on the 50-patient ACDC test set. That number requires that the two-stage classifier, and especially the expert SVM refinement, was designed without access to test-set outcomes. The paper does not establish this. Section IV-B reports that the initial random forest obtained 92% accuracy and the final pipeline 98%, after an SVM refined only the MINF/DCM predictions using two myocardial wall thickness features. With 50 patients, 92% to 98% is three corrected errors, so even mild selection on the test set can move the headline number substantially. No validation split, feature-selection criterion, or SVM hyperparameter tuning procedure is reported. The description of the two features is internally inconsistent: Section III-C2 says the SVM uses the maximum of slice-wise mean MWT and the mean of slice-wise standard deviations at end-systole, while Section IV-B and the text near Figure 5 say it uses the mean of slice-wise standard deviations and the standard deviation of slice-wise mean MWT at end-diastole. If the features, the refinement rule, or the SVM hyperparameters were chosen after inspecting the 50 test cases, the 98% figure is not a held-out accuracy, and the comparison in Table III is invalid for the classification claim. The segmentation result (92.56% average Dice) is less vulnerable, but the classification claim is the load-bearing part of the paper's central contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents IntelliCardiac, a web-based platform for cardiac cine-MRI segmentation and disease classification. It combines a 3D residual U-Net segmentation model with a two-stage classifier (random forest followed by an RBF-SVM expert refinement) and evaluates both tasks on the ACDC dataset (100 training, 50 test patients). The authors report a mean Dice score of 92.56% for segmentation and 98% classification accuracy across five diagnostic categories, and they claim these results outperform prior methods that integrate segmentation and classification. The paper also describes the platform's architecture, user roles, and deployment details.","tokens_in":9896,"tokens_out":3467,"duration_ms":43448,"significance":"If the reported numbers are reproducible, the contribution is valuable: an open-source, web-based tool that addresses segmentation and classification together, with a clearly described interface and a 4D processing pipeline, evaluated on the public ACDC benchmark. The two-stage refinement idea, using clinically motivated myocardial wall thickness features to separate DCM from MINF, is reasonable and potentially useful. However, the scientific claims currently rest on sparse empirical evidence: no error bars or repeated runs, no described validation split for classifier design, and an inconsistent specification of the refinement features. With these added, the paper could provide a credible and useful reference for integrated cardiac MRI analysis platforms.","major_comments":[{"comment":"The two myocardial wall thickness (MWT) features used by the expert SVM are described inconsistently. Section III-C2 states the SVM uses the maximum of slice-wise mean wall thickness and the mean of slice-wise standard deviations measured during the end-systolic phase, while Section IV-B and the text around Figure 5 state the features are the mean of slice-wise standard deviations of MWT and the standard deviation of slice-wise mean MWT at end-diastole. This is a load-bearing discrepancy: it makes the classifier design unreproducible and prevents verification of the reported 98% accuracy. The authors must specify the exact feature definitions, the cardiac phase used, and the feature-selection rule that led to these two features.","section":"§III-C2 vs §IV-B"},{"comment":"No validation split or model-selection protocol is reported for the two-stage classifier. The initial classifier is described as obtaining 92% accuracy and the final pipeline 98% on the 50-patient test set; on 50 patients, that is a difference of about three corrected predictions. If the random forest hyperparameters, the choice of the refinement features, the MINF/DCM-only refinement rule, or the SVM hyperparameters (C, gamma) were selected after inspecting the test-set predictions, the 98% figure is not a held-out accuracy. The paper should report the validation procedure, the hyperparameter ranges considered, and the decision rule for selecting the refinement features.","section":"§IV-B (Classification Results)"},{"comment":"The claim of 'outperforming existing state-of-the-art methods' is not supported by the comparison set. Table III compares against Wolterink (2018), Wibowo (2022), Zheng (2019), and Khened (2018), all of which predate many recent ACDC classification results, and no statistical significance tests or confidence intervals are reported for the classification accuracy difference. The paper should either include current state-of-the-art methods that perform both segmentation and classification on ACDC, or temper the claim to 'outperforms the compared methods.'","section":"Table III and Classification Comparison"},{"comment":"The segmentation results are reported as point estimates without standard deviations, ranges, or patient-level variability. Given that the mean Dice differences between IntelliCardiac and the nearest comparator (nnFormer, EMCAD) are small (0.4–0.5 points), the absence of variability measures makes it difficult to judge whether the improvements are meaningful. The authors should report error bars or per-fold/per-patient statistics and describe how the segmentation model was selected with respect to validation data.","section":"Table II and Segmentation Results"}],"minor_comments":[{"comment":"The phrase 'bridges the hole between AI research and real-world deployment' should be 'bridges the gap between AI research and real-world deployment.'","section":"Abstract/Introduction"},{"comment":"In the decoder description, 'linked with the unsampled feature maps' should be 'linked with the upsampled feature maps.'","section":"§III-B2"},{"comment":"The sentence 'We have a contribution lies in the dynamically weighted Focal Dice Loss' should be rewritten, for example as 'Our contribution includes the dynamically weighted Focal Dice Loss.'","section":"§III-B3"},{"comment":"The sentence 'This immediately address class imbalance and improve segmentation uniformity, this encourages the model to give priority to minority or underperforming buildings' has subject-verb agreement issues and should read 'This addresses class imbalance and improves segmentation uniformity, encouraging the model to give priority to minority or underperforming structures.'","section":"§III-B3"},{"comment":"The phrase 'avoids minimizes overfitting' is ungrammatical; it should be 'avoids overfitting' or 'minimizes overfitting.'","section":"§III-C2"},{"comment":"The caption says 'Confusion matrices before (right) and after (left) expert refinement,' but the usual convention is to present the initial state on the left and the refined state on the right; please clarify the ordering or reorder the panels.","section":"Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for an imaging-applications venue, and the open-source platform is a positive element. The main concern is the unverified classifier design: the inconsistent feature descriptions and the absence of a validation split make the headline classification number non-reproducible. I would ask the authors to resolve the inconsistency and document the model-selection protocol, ideally with error bars, before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about IntelliCardiac. First, the platform itself is a genuine engineering contribution: a working web-based cardiac MRI tool with a public GitHub repo, an end-to-end workflow from upload to segmentation to diagnosis, and a sensible two-stage classifier that first runs a random forest on volumetric features, then refines ambiguous MINF/DCM cases with a small SVM. Second, the headline classification number - 98% on the ACDC 50-patient test set - is not backed by a protocol that rules out test-set-guided tuning. There is no validation split, no hyperparameter selection procedure, and no discussion of whether the SVM refinement features were chosen after looking at test outputs. The jump from 92% to 98% is exactly three corrected errors, so even mild selection on the test set could produce this gap. I read the stress-test concern carefully and it holds: Section III-C2 says the SVM uses the maximum of slice-wise mean MWT and the mean of slice-wise standard deviations at end-systole, while Section IV-B and Figure 5 say the mean of slice-wise standard deviations and the standard deviation of slice-wise mean MWT at end-diastole. That is a factual contradiction about the core of the classification module.\n\nWhat is genuinely good: the segmentation architecture is reasonable, the dynamically weighted Focal Dice loss plus ROI cropping is a plausible combination, and the ablation table showing the effect of the loss and ROI is a nice touch. The 92.56% mean Dice is competitive with nnU-Net and nnFormer, though the reported gains are within the noise you would expect without error bars or repeated runs. The platform's design for patients and doctors, with shared reports and feedback, is thoughtful and clinically motivated.\n\nThe soft spots beyond the classification protocol: the comparison in Table III omits the strongest integrated methods from the ACDC challenge (most notably Isensee et al.'s own feature-based classifier, which is cited but never benchmarked). Claiming to outperform state-of-the-art on that basis is overreach. Also, 98% on 50 patients is a small sample; a confidence interval or per-class sensitivity/specificity would be more honest than a single accuracy number.\n\nMy recommendation: send it to peer review, but with a request for major revision. The segmentation side is probably salvageable and the platform deserves a public description. The classification claim needs a proper validation protocol, a consistent description of the SVM features, and a realistic comparison. This paper is for researchers building applied medical imaging pipelines, and for clinical informatics readers who care about deployable AI. I would not cite it myself until the classification evaluation is cleaned up, but it is not a desk-reject. The thinking behind the platform is clear; the reporting is just sloppy.","headline":"The platform is real and the segmentation numbers are plausible, but the 98% classification accuracy is not trustworthy as a held-out result because the paper never shows a validation split and describes the SVM's features inconsistently.","tokens_in":783,"tokens_out":1721,"would_cite":false,"duration_ms":40079,"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":"IntelliCardiac argues that a web-based cardiac MRI platform can segment heart structures at 92.6% mean Dice and classify five disease categories at 98% accuracy, outperforming prior combined segmentation-and-classification systems on the…","keywords":["cardiac MRI","image segmentation","disease classification","two-stage classifier","web-based platform","ACDC dataset","3D U-Net","myocardial wall thickness"],"falsifier":"Re-run the classification module on the same 50-patient ACDC test split while withholding all test labels until after the SVM refinement features are selected on a training or validation split; if final accuracy then fails to exceed the initial Random Forest's 92%, the reported 98% is explained by test-set-guided tuning rather than genuine generalization.","tokens_in":9429,"feed_emoji":"🫀","tokens_out":7562,"duration_ms":78168,"temperature":0.7,"pith_summary":"The paper tries to establish that a single web-based platform can take a 4D cardiac MRI scan, automatically segment the right ventricle, left ventricle, and myocardium, and then classify the patient into one of five diagnostic categories with accuracy sufficient for clinical screening. On the public ACDC dataset the authors report a mean Dice score of 92.6% for segmentation and 98% accuracy for the five-way classification after a two-stage classifier refines ambiguous cases. The significance, if true, is that integration need not cost accuracy: one tool can handle visualization, segmentation, diagnosis, and doctor-patient communication, whereas existing web platforms mostly offer visualization alone and prior combined AI pipelines trail these numbers. The paper also isolates which design choices carry the gain, namely ROI-focused preprocessing, a dynamically weighted Focal Dice Loss, and a specialist SVM that re-decides only MINF/DCM cases using two myocardial wall-thickness features.","feed_headline":"Cardiac MRI web platform claims 98% diagnostic accuracy","feed_subtitle":"A web-based tool segments heart chambers and sorts five disease classes, beating earlier combined pipelines.","key_machinery":"The central mechanism is a two-module pipeline. Segmentation uses a 3D residual U-Net with four ResBlock encoder layers, trilinear upsampling, skip connections, and a four-channel output; its loss is a Focal Dice Loss with per-class weights updated each epoch to the inverse of the previous epoch's Dice, so the model concentrates on underperforming classes. An ROI preprocessing step locates the left-ventricle center from temporal intensity variation and circular Hough detection, then crops a fixed 128x128 patch with logical-depth cropping to cover variable temporal depths. From the resulting ED/ES masks, the pipeline extracts 20 features, including volumes, ejection fractions, volume ratios, and myocardial wall thickness statistics. Classification is a two-stage decision: a Random Forest on all 20 features makes the initial five-way call, and an RBF SVM, trained only on the maximum of slice-wise mean wall thickness and the mean of slice-wise standard deviations at end-systole, rewrites predictions that were initially MINF or DCM. That refinement is what lifts accuracy from 92% to 98%.","core_discovery":"On the paper's own terms, the discovery is that combining a 3D residual U-Net segmentation module with a two-stage classification pipeline on top of segmentation-derived features yields state-of-the-art results on the ACDC benchmark: an average Dice score of 92.56% (RV 92.27%, Myo 90.33%, LV 95.09%) and a final classification accuracy of 98% across DCM, MINF, HCM, ARV, and normal cases. The classification gain comes from a specific refinement step: a Random Forest first labels each case, and an RBF-kernel SVM re-judges only the cases predicted as MINF or DCM using two myocardial wall-thickness statistics at end-systole, replacing those labels. The paper attributes the segmentation margin to logical-depth cropping, ROI detection via temporal-standard-deviation plus circular Hough voting, dynamically weighted Focal Dice Loss, and largest-connected-component post-processing. It reports these numbers beat the comparison methods listed in Tables II and III, including nnU-Net and nnFormer for segmentation and combined segmentation-classification baselines for accuracy.","pith_inferences":["The 98% figure is a single-dataset result on 50 ACDC test patients; before clinical use, the same fixed pipeline would need validation on an external cohort to test whether the two refinement features generalize.","The refinement recipe is transferable: any diagnostic pair with similar bulk-volume profiles but distinct localized shape statistics could use the same broad-classifier-plus-specialist-SVM pattern.","If the features were selected after inspecting test confusion, the honest comparison is the initial classifier's 92%, and the platform's real contribution would be integration rather than the reported accuracy gain.","The doctor-feedback feature could become a data-collection loop for retraining, but the paper does not measure whether that loop improves future models."],"forward_implications":["A complete MRI upload-to-report run takes about 3 seconds per patient, so the same pipeline can support interactive use rather than offline batch analysis.","The segmentation module's highest margin over baselines is in the right ventricle and myocardium, the structures whose geometry drives the downstream MINF/DCM distinction.","Combining segmentation and classification in one web platform means the same system can serve patients, doctors, and researchers, with secure sharing and doctor comment workflows.","The dynamically weighted Focal Dice Loss, which reweights classes by inverse previous-epoch Dice, is proposed as a general fix for class imbalance in 3D medical segmentation.","On ACDC, the full pipeline reports higher segmentation Dice and higher classification accuracy than the combined-task baselines listed in Table III."],"supporting_citations":[{"why":"Supplies the ACDC training (100 patients) and test (50 patients) data and the five disease categories used for both tasks.","marker":"[1]"},{"why":"Provides the feature-based classification approach from which the 20 segmentation-derived features are taken.","marker":"[10]"},{"why":"The Random Forest is the initial five-way classifier in the two-stage pipeline.","marker":"[2]"},{"why":"The RBF-kernel SVM is the expert refinement classifier for MINF/DCM cases.","marker":"[9]"},{"why":"The Focal Dice Loss is the base loss that the proposed dynamic weighting extends.","marker":"[24]"},{"why":"nnU-Net is a state-of-the-art segmentation baseline that the proposed model is compared against and reports to outperform.","marker":"[11]"},{"why":"A combined segmentation-and-classification baseline whose 90% accuracy the paper reports to beat.","marker":"[13]"},{"why":"A combined segmentation-and-classification baseline (89.30% Dice, 86% accuracy) the paper compares against.","marker":"[26]"},{"why":"A combined segmentation-and-classification baseline (90.89% Dice, 92% accuracy) the paper compares against.","marker":"[25]"},{"why":"A combined segmentation-and-classification baseline (88.33% Dice, 94% accuracy) the paper reports to outperform.","marker":"[28]"}],"fun_headline_variants":["IntelliCardiac hits 98% accuracy in cardiac MRI diagnosis","IntelliCardiac beats state-of-the-art with 98% cardiac MRI accuracy","Web platform IntelliCardiac: 98% accurate on five cardiac classes","Two-step AI classifier nails 98% of cardiac MRI cases","IntelliCardiac's hybrid model: 98% accuracy, 92.6% Dice on ACDC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 98% classification accuracy relies on the assumption that the two-stage design, particularly the choice of the two myocardial wall-thickness features and the rule that only initial MINF/DCM predictions get refined, was fixed before the 50-patient test set was examined, with no test information guiding feature selection.","fun_headline_variants_meta":{"raw":{"variants":["IntelliCardiac hits 98% accuracy in cardiac MRI diagnosis","IntelliCardiac beats state-of-the-art with 98% cardiac MRI accuracy","Web platform IntelliCardiac: 98% accurate on five cardiac classes","Two-step AI classifier nails 98% of cardiac MRI cases","IntelliCardiac's hybrid model: 98% accuracy, 92.6% Dice on ACDC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001452,"raw_usage":{"total_tokens":5873,"prompt_tokens":996,"completion_tokens":4877,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":4773}},"tokens_in":612,"tokens_out":4877,"duration_ms":36271,"temperature":1.0,"reasoning_tokens":4773,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:53:58.913968+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the classification module on the same 50-patient ACDC test split while withholding all test labels until after the SVM refinement features are selected on a training or validation split; if final accuracy then fails to exceed the initial Random Forest's 92%, the reported 98% is explained by test-set-guided tuning rather than genuine generalization.","supporting_citations":[{"cited_title":"Random Forests","cited_arxiv_id":null,"evidence_quote":"The Random Forest is the initial five-way classifier in the two-stage pipeline."},{"cited_title":"Hearst, S.T","cited_arxiv_id":null,"evidence_quote":"The RBF-kernel SVM is the expert refinement classifier for MINF/DCM cases."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Focal Dice Loss is the base loss that the proposed dynamic weighting extends."},{"cited_title":"nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation","cited_arxiv_id":null,"evidence_quote":"nnU-Net is a state-of-the-art segmentation baseline that the proposed model is compared against and reports to outperform."},{"cited_title":"Densely Connected Fully Convolutional Network for Short-Axis Cardiac Cine MR Image Segmentation and Heart Diagnosis Using Random Forest","cited_arxiv_id":null,"evidence_quote":"A combined segmentation-and-classification baseline whose 90% accuracy the paper reports to beat."},{"cited_title":"Wolterink, Tim Leiner, Max A","cited_arxiv_id":null,"evidence_quote":"A combined segmentation-and-classification baseline (89.30% Dice, 86% accuracy) the paper compares against."},{"cited_title":"Cardiac Disease Classification Using Two-Dimensional Thickness and Few-Shot Learning Based on Magnetic Resonance Imaging Image Segmentation","cited_arxiv_id":null,"evidence_quote":"A combined segmentation-and-classification baseline (90.89% Dice, 92% accuracy) the paper compares against."},{"cited_title":"Jaeger, Peter M","cited_arxiv_id":null,"evidence_quote":"Provides the feature-based classification approach from which the 20 segmentation-derived features are taken."},{"cited_title":"Maier-Hein, Peter M","cited_arxiv_id":null,"evidence_quote":"Supplies the ACDC training (100 patients) and test (50 patients) data and the five disease categories used for both tasks."},{"cited_title":"Explainable car- diac pathology classification on cine MRI with motion characterization by semi-supervised learning of apparent flow","cited_arxiv_id":null,"evidence_quote":"A combined segmentation-and-classification baseline (88.33% Dice, 94% accuracy) the paper reports to outperform."}],"review_version":1}