{"id":"b19d1cf7-12d7-471c-92b1-3f11920ea71b","arxiv_id":"2506.22222","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A deep-learning study reports improved automated segmentation of Type B aortic dissection components on the ImageTBAD dataset, with best true-lumen Dice 0.92 and best false-lumen-thrombosis Dice 0.47, but model selection used test data.","lead":"This paper tests four deep-learning pipelines for automatically outlining the true lumen, false lumen, and blood clots in CT scans of Type B aortic dissections. On a 100-patient public dataset, the best pipeline reports higher Dice scores than a published baseline, though the comparison and model selection have methodological flaws.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-set leakage in Section 2.3—architecture selection used testing performance—makes the reported Dice improvements over Yao et al. unreliable as unbiased estimates.","rationale":"Test-set leakage is a well-known source of optimistic performance estimates. Here it is not a suspicion but an explicit statement in Section 2.3, and the later decision to bypass the FLT classifier in Method 3 reinforces that the test set informed methodological choices. The reported numbers may still represent the true performance if the authors were lucky, but the experiment as described does not establish unbiased generalization. The paper has useful components—a 100-case dataset, four implemented pipelines, and a comparison of GDL vs DCEL—but the headline 'outperforming Yao et al.' cannot be evaluated from the current protocol. A re-analysis with validation-only selection (or nested cross-validation) would resolve the issue. My recommendation aligns with the reader's REJECT: the concern is central, and the verdict does not need to change.","tokens_in":7741,"tokens_out":3174,"duration_ms":33143,"concrete_test":"Ask the authors to re-run the complete selection procedure using only the 10-case validation fold: pick the segmentation network (3D U-Net vs Swin-UnetR), loss (GDL vs DCEL), pipeline (1–4), and whether to include the FLT classification branch based solely on validation DC/True FLT DC; then compute test metrics once on the untouched 10-case test set. If the resulting True FLT DC is below the reported 0.47, or if a method other than Method 3 is chosen on validation, the advertised superiority over Yao et al. is not supported by the current evaluation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that Method 3 (and other proposed pipelines) substantially outperform Yao et al. on TL, FL, and FLT Dice—rests on test-set scores that were not obtained with a cleanly held-out test set. Section 2.3 states: 'We employed two segmentation networks... selecting them based on their testing performance.' Section 3.1 then reports that the FLT classification network 'did not perform well' and 'the classification step was bypassed in Method 3,' a decision evidently made after examining test results. Because the choice of network, pipeline, and even whether to use the FLT classification branch was influenced by the test set, the numbers in Table 1 (e.g., True FLT DC 0.47 vs Yao et al.'s 0.25) are optimistic estimates of generalization. The comparison to Yao et al. is also not controlled: Table 1 lists 'Yao et al.'s Testing' with FLT DC 0.15, whereas the original Yao et al. paper reported 0.25, and the authors modified Yao's preprocessing (removed downsampling), so the baseline is their re-implementation rather than a published number. The load-bearing issue is test-set leakage, not the baseline mismatch alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes four deep learning pipelines for automated segmentation of true lumen (TL), false lumen (FL), and false lumen thrombosis (FLT) in Type B aortic dissection from CTA images. Using the ImageTBAD dataset (100 cases), the authors compare single-step, sequential, sequential multi-task, and ensemble architectures built on 3D U-Net and Swin-UnetR, with DenseNet classifiers in the multi-task pipeline. They report Dice coefficients and Hausdorff distances, claim that their Method 3 achieves the best FLT segmentation (True FLT Dice 0.47) and that their methods substantially outperform the prior state of the art reported by Yao et al.","tokens_in":7909,"tokens_out":8257,"duration_ms":74521,"significance":"If the reported results were unbiased, the paper would provide a useful comparison of modern segmentation architectures on a public aortic dissection dataset and a potential step forward in FLT segmentation, which is clinically relevant for surveillance and treatment planning. The paper's strengths include the use of a public dataset, the systematic comparison of multiple pipeline designs, and the evaluation of two loss functions. However, the central performance claim is not currently established because the evaluation protocol appears to use the test set for model selection, the baseline comparison to Yao et al. is not controlled, and the experimental design is internally inconsistent. The contribution is therefore more of a proof-of-concept than a reliable advance until these methodological issues are resolved.","major_comments":[{"comment":"Section 2.3 states that the two segmentation networks and two classification networks were 'selecting them based on their testing performance.' This is a direct admission that the test set was used for model selection, which invalidates the test-set Dice coefficients in Table 1 as unbiased estimates of generalization. This concern is compounded by Section 3.1, where the classification step in Method 3 was 'bypassed' after observing poor classification performance, a decision that also appears to be informed by test-set results. Because the paper's central claim—that Method 3 substantially outperforms Yao et al.—rests on these test-set numbers, the reported improvements are likely optimistic. The authors should redo the evaluation with model selection based solely on validation data and report metrics from a test set that is used only once.","section":"2.3"},{"comment":"The baseline comparison to Yao et al. is not controlled. Section 3.1 says the authors 'implemented the approach of Yao et al. on our dataset also by removing some suboptimal preprocessing steps,' yet Section 4 compares Method 3 to 'their original method's performance of 0.25 ± 0.31 DC for True FLT.' Table 1 lists a row labeled 'Yao et al.'s Testing' with FLT DC 0.15 and True FLT DC 0.25, which is not the original published baseline (the original paper reported FLT DC of 0.25). Since the authors altered the preprocessing (removing downsampling), the reimplementation is a different method, and the comparison to published numbers that were obtained under different conditions is invalid. A valid comparison would reimplement the original method exactly, or clearly label the modified version and avoid quoting the original published numbers.","section":"3.1, Table 1, Section 4"},{"comment":"The experimental design is internally inconsistent. The Abstract and Section 2.2 describe a single split into training (n=80), validation (n=10), and testing (n=10), but Section 2.2 also states 'We implemented cross-validation with 5 folds to maximize the utilization of this small dataset.' It is not explained how a fixed 80/10/10 split relates to 5-fold cross-validation, nor whether the results in Tables 1-3 are from a single held-out test set or aggregated across folds. This ambiguity affects the interpretation of the reported standard deviations and the generalizability of the claims; the authors should specify the exact evaluation protocol used to produce each table.","section":"2.2"},{"comment":"The architecture of Method 3 as evaluated is unclear. Section 2.3 describes a sequential multi-task pipeline in which a classification network first determines whether FLT is present and conditionally invokes an FLT-specific segmentation network. Section 3.1 then reports that 'the classification step was bypassed in Method 3' because the classifier underperformed. The paper does not state whether the Method 3 results in Tables 1 and 2 correspond to the full described pipeline or to a modified version without classification. If the classification step was removed, the method should be described and named accordingly, and the results should be interpreted as belonging to that modified architecture.","section":"2.3 vs 3.1"}],"minor_comments":[{"comment":"The Abstract reports 'Dice Coefficients of ... 0.47 ± 0.25 for FLT,' but Table 1 shows 0.47 for True FLT and 0.38 for overall FLT. Please report the metric consistently and clearly label True FLT versus FLT in the Abstract and Key Points.","section":"Abstract"},{"comment":"The text says 'we maintained a constant ratio of FLT cases within each fold,' but the exact ratio and the number of FLT-positive cases per fold are not provided. Please quantify the stratification.","section":"2.2"},{"comment":"The preprocessing description states that images were 'cropped to focus on the foreground,' but the criterion for identifying foreground is not defined. Please specify the cropping procedure.","section":"2.2"},{"comment":"The Swin-UnetR DCEL row in Table 3 exactly matches the Method 1 testing row in Table 1. Please clarify whether Method 1 is Swin-UnetR trained with DCEL, so that the relationship between the two tables is explicit.","section":"Table 3"},{"comment":"Reference [26] (Kingma and Ba) is missing the year; reference [8] (Isensee et al.) is missing the year and volume/page details. Please complete all bibliographic entries.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"I recommend that the editor require the authors to address the test-set selection issue explicitly. If the statement in Section 2.3 is accurate, the reported test metrics are not trustworthy, and the conclusions may need to be revised after a properly blinded evaluation. The paper also has multiple internal inconsistencies (cross-validation vs. fixed split, baseline labeling, Method 3 description) that suggest the manuscript was not carefully checked. With a true held-out test set and a controlled baseline, the comparison would be valuable, but as submitted the central claim is not established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the paper's headline numbers—Dice 0.91/0.88/0.47 for TL/FL/FLT—are not trustworthy as unbiased estimates because the authors selected networks \"based on their testing performance\" (Section 2.3) and then bypassed the FLT classification step after seeing test results (Section 3.1). That is test-set leakage, and it inflates the reported scores.\n\nWhat is genuinely useful: the paper compares four segmentation pipelines on the public ImageTBAD dataset, including a cascade with a dedicated FLT branch, and it includes a small loss-function ablation (GDL vs DCEL). The writing is clear, and the authors openly list limitations—small dataset, mislabeled cases, hardware constraints—which is more than many papers do.\n\nThe other major soft spot is the baseline. Table 1 labels \"Yao et al.'s Testing\" but these are not the published numbers: the authors reimplemented Yao's method with modified preprocessing (removing downsampling) and report FLT Dice of 0.15 in the table while the abstract and discussion cite 0.25 as Yao's result. The comparison is therefore not controlled, and part of the claimed improvement may just be a different evaluation setup. Also, Section 2.2 mentions 5-fold cross-validation, but the results come from a single 80/10/10 split; the relationship is never explained. With only 10 test cases, the standard deviations are wide, and True FLT Dice is computed on roughly 6-7 cases.\n\nThe central argument does not hold as written. A reader in the aortic-dissection imaging subfield could still learn something from the pipeline taxonomy and the loss-function comparison, but the quantitative claims should not be cited. If the authors re-run the evaluation with architecture selection on validation data only, run a fully controlled baseline reimplementation, and clarify the split, this could become a modest but citable contribution. As it stands, I would not send it to external peer review. I'd give a clear, encouraging reject-and-resubmit response.","headline":"Test-set leakage in model selection and an uncontrolled baseline make the reported Dice gains over Yao et al. unreliable, though the pipeline comparison has some utility.","tokens_in":8505,"tokens_out":4279,"would_cite":false,"duration_ms":44384,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A multi-task deep learning pipeline segments Type B aortic dissection on CT with Dice 0.91–0.92 for true lumen, 0.87–0.89 for false lumen, and 0.47 for thrombosis, beating prior published results.","keywords":["Type B aortic dissection","automated segmentation","computed tomography angiography","deep learning","false lumen thrombosis","medical image segmentation","ensemble model","Swin-UnetR"],"falsifier":"Run Method 3 on a fresh pre-registered test set that is completely untouched during model and loss selection, with the FLT-presence classifier left in the loop, and compare True FLT Dice to the reported 0.47; if the classifier's roughly 50% accuracy causes most thrombosed cases to bypass the dedicated network, the effective FLT Dice on real deployment will fall well below the headline number.","tokens_in":7456,"feed_emoji":"🫀","tokens_out":9473,"duration_ms":83013,"temperature":0.7,"pith_summary":"This paper tries to show that automated segmentation of Type B aortic dissection from CT angiography can be made substantially more accurate than prior published methods, especially for the difficult false lumen thrombosis component. The authors build four deep learning pipelines and report that a sequential multi-task model, which dedicates a separate network to thrombosis segmentation, achieves the highest Dice scores on the three anatomical classes, while an ensemble of two architectures gives the lowest boundary errors. If these numbers hold on new data, the pipeline would let clinicians extract true lumen and false lumen diameters and thrombosis extent from a scan in minutes instead of up to 1.5 hours of manual tracing. The central comparison is against the previous ImageTBAD baseline, whose reported Dice for true lumen, false lumen, and thrombosis were 0.78, 0.68, and 0.25.","feed_headline":"Multi-task network lifts aortic dissection segmentation to Dice 0.92","feed_subtitle":"Sequential multi-task model scores 0.91/0.88/0.47 for true lumen, false lumen, and thrombosis on 100 CTA scans.","key_machinery":"The central mechanism is the sequential multi-task pipeline (Method 3), which isolates the difficult false lumen thrombosis segmentation into its own network rather than asking one network to solve all classes at once. A companion ensemble (Method 4) averages the softmax probabilities of a Swin-UnetR transformer and a 3D U-Net, trading some structure for boundary precision. The pipelines rely on preprocessing that resamples each scan to 1.5 mm isotropic voxels and crops foreground rather than downsampling, preserving small thrombosis structures, and on Dice Cross Entropy Loss, which is reported to outperform Generalized Dice Loss for the transformer network.","core_discovery":"On its own terms, the paper's central claim is that its proposed pipelines—especially the sequential multi-task model (Method 3) and the ensemble model (Method 4)—outperform the prior state of the art for segmenting true lumen, false lumen, and false lumen thrombosis from Type B aortic dissection CTA scans. Method 3 uses a classification network to detect thrombosis, a dedicated segmentation network for the thrombosis when present, a second network for true and false lumen, and a final network that merges the outputs; the authors report test Dice of 0.91 for true lumen, 0.88 for false lumen, 0.38 for all-case false lumen thrombosis, and 0.47 for thrombosis measured only on cases that actually contain it. Method 4 averages the probability outputs of Swin-UnetR and 3D U-Net and achieves the lowest Hausdorff distances for true and false lumen, indicating the most spatially precise boundaries. The paper also reports that the thrombosis-presence classifier itself only reaches about 50% accuracy, so in practice the classification step was bypassed.","pith_inferences":["A fair replication should use a three-way split in which the test set is never touched until final evaluation; the methods text says networks were selected based on testing performance, so the reported test Dice may be optimistic.","Because the thrombosis-presence classifier reaches only about 50% accuracy, the reported 0.47 True FLT Dice applies only when the dedicated FLT network is actually triggered; in an automated deployment that keeps the classifier in the loop, many thrombosed cases would be missed.","The 1.5 mm resampling and cropping may under-resolve small peripheral thrombi; testing at higher input resolution or with patch-based inference could reveal whether 0.47 is a resolution ceiling or a model ceiling.","The same multi-task and ensemble designs could transfer to Type A dissection or other thrombus-bearing vessels if labeled training data are available."],"forward_implications":["If the reported Dice values generalize, clinicians could obtain true lumen, false lumen, and thrombosis segmentations from a CTA in minutes rather than the roughly 1.5 hours of manual tracing, enabling routine serial surveillance.","The sequential multi-task design raises the true-lumen thrombosis Dice from 0.32 for the plain two-network cascade to 0.47, suggesting that isolating the rare class in its own network is a workable route to improving segmentation of small structures.","Ensembling two architectures yields the lowest Hausdorff distances for true and false lumen, which matters for downstream morphological measurements like lumen diameter and thrombus volume.","With accurate automated masks, morphological parameters such as true lumen calibre, false lumen size, and thrombosis extent become feasible to derive routinely for treatment planning and follow-up."],"supporting_citations":[{"why":"Supplies the ImageTBAD dataset of 100 CTA scans and the prior baseline Dice values that the new pipelines are compared against.","marker":"[1]"},{"why":" Introduced the sequential multi-stage CNN approach for TBAD segmentation that the new cascaded pipelines build on.","marker":"[16]"},{"why":"Presented the cascaded 3D U-Net for TL, FL, and FLT segmentation whose cascade structure Method 3 modifies with a dedicated FLT network.","marker":"[17]"},{"why":"Showed that cropping image patches instead of downsampling avoids information loss, supporting the preprocessing choice used here.","marker":"[18]"},{"why":"Provides the 3D U-Net architecture used as one of the two base segmentation networks.","marker":"[21]"},{"why":"Provides the Swin-UnetR transformer architecture used as the other base segmentation network.","marker":"[22]"},{"why":"Defines the Generalized Dice Loss that the authors compared against their chosen Dice Cross Entropy Loss.","marker":"[24]"}],"fun_headline_variants":["Deep learning model segments aortic dissection with 0.91 Dice","Multi-task network outperforms prior on aortic dissection segmentation","Automated deep learning hits Dice 0.91 for aortic dissection lumens","Aortic dissection segmentation reaches 0.91 Dice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported test scores assume the test set was never used to pick models or settings; the methods text says networks were selected based on their testing performance, so that assumption may not hold.","fun_headline_variants_meta":{"raw":{"variants":["Deep learning model segments aortic dissection with 0.91 Dice","Multi-task network outperforms prior on aortic dissection segmentation","Automated deep learning hits Dice 0.91 for aortic dissection lumens","Aortic dissection segmentation reaches 0.91 Dice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001528,"raw_usage":{"total_tokens":6162,"prompt_tokens":1033,"completion_tokens":5129,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":649,"completion_tokens_details":{"reasoning_tokens":5059}},"tokens_in":649,"tokens_out":5129,"duration_ms":37512,"temperature":1.0,"reasoning_tokens":5059,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:08:36.280578+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Method 3 on a fresh pre-registered test set that is completely untouched during model and loss selection, with the FLT-presence classifier left in the loop, and compare True FLT Dice to the reported 0.47; if the classifier's roughly 50% accuracy causes most thrombosed cases to bypass the dedicated network, the effective FLT Dice on real deployment will fall well below the headline number.","supporting_citations":[{"cited_title":"ImageTBAD: A 3D Computed Tomography Angiography Image Dataset for Automatic Segmentation of Type-B Aortic Dissection","cited_arxiv_id":null,"evidence_quote":"Supplies the ImageTBAD dataset of 100 CTA scans and the prior baseline Dice values that the new pipelines are compared against."},{"cited_title":"Fully automatic segmentation of type B aortic dissection from CTA images enabled by deep learning","cited_arxiv_id":null,"evidence_quote":"Introduced the sequential multi-stage CNN approach for TBAD segmentation that the new cascaded pipelines build on."},{"cited_title":"Deep Learning-Based 3D Segmentation of True Lumen, False Lumen, and False Lumen Thrombosis in Type-B Aortic Dissection","cited_arxiv_id":null,"evidence_quote":"Presented the cascaded 3D U-Net for TL, FL, and FLT segmentation whose cascade structure Method 3 modifies with a dedicated FLT network."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Showed that cropping image patches instead of downsampling avoids information loss, supporting the preprocessing choice used here."},{"cited_title":"CT-based True- and False-Lumen Segmentation in Type B Aortic Dissection Using Machine Learning","cited_arxiv_id":null,"evidence_quote":"Provides the 3D U-Net architecture used as one of the two base segmentation networks."},{"cited_title":"3D U-Net: Learning Dense V olumetric Segmentation from Sparse Annotation; 2016","cited_arxiv_id":null,"evidence_quote":"Provides the Swin-UnetR transformer architecture used as the other base segmentation network."},{"cited_title":"Densely Connected Convolutional Networks; 2018","cited_arxiv_id":null,"evidence_quote":"Defines the Generalized Dice Loss that the authors compared against their chosen Dice Cross Entropy Loss."}],"review_version":1}