{"id":"ef6899cc-837f-4234-a2eb-5a885974880b","arxiv_id":"2411.15656","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A YOLOv8 plus Self-ONN cascade segments lumbar vertebrae and discs with 91.03% Dice, but only T1-weighted scans are used, so the machine-agnostic claim is not established.","lead":"This paper proposes a two-stage pipeline that first locates the lumbar spine in MRI slices with YOLOv8 and then segments individual vertebrae and intervertebral discs with a Self-ONN based decoder paired with a DenseNet121 encoder. The authors report a Dice score of 91.03% on a multi-scanner T1-weighted dataset, but their machine-agnostic and first-of-its-kind claims outrun the evidence presented.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Scanner-disjoint evaluation is missing: subject-wise folds let the same MRI scanner appear in both train and test, so the 'machine-agnostic' claim is not actually tested.","rationale":"The reader's REJECT verdict is well-founded, but the most load-bearing flaw is not only that segmentation is restricted to T1-weighted images. Even within T1, the cross-validation design does not establish scanner-agnostic generalization because folds are subject-disjoint but not scanner-disjoint. This is a concrete, correctable experimental-design issue that directly undermines the paper's central claim. The reader's mention of T1-only and selection bias are related, but the scanner-leakage problem is more decisive: it can inflate the headline numbers on the exact axis the paper claims. A scanner-disjoint evaluation is a straightforward check that would settle whether the model genuinely generalizes across machines. If the numbers hold under scanner-disjoint validation, the central claim would be substantially strengthened; if they drop, the paper's main contribution would need to be reframed. Thus the verdict remains REJECT pending that evidence, and I recommend no change to the reader's verdict.","tokens_in":16966,"tokens_out":4212,"duration_ms":39295,"concrete_test":"Re-run the 10-fold cross-validation with scanner-disjoint folds: assign all subjects scanned on each of the 12 scanners to the same fold, or use leave-one-scanner-out, then retrain the same YOLOv8 + Self-ONN pipeline and report mean IoU, DSC, and sensitivity. Also report the original scanner-to-fold assignment to confirm whether the same scanner appears in both training and test folds. If the mean DSC/IoU drops substantially or a held-out scanner performs poorly, the 'machine-agnostic' claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the pipeline is 'machine-agnostic,' i.e., generalizes across MRI scanners. The only evidence is a 10-fold cross-validation described in Section 4.2 as subject-preserved ('no subject would be present in different folds'). But subject preservation does not prevent scanner leakage. The dataset has 12 scanners and 34 subjects, so a single scanner typically contributes multiple subjects; those subjects can be assigned to different folds. The same scanner's intensity and artifact characteristics can therefore appear in both training and test folds, allowing the model to memorize scanner-specific appearance rather than learn anatomy. That is exactly the shortcut that would inflate cross-scanner generalization scores. The abstract's emphasis on '12 scanners' is used as evidence for machine-agnostic performance, but no analysis shows whether scanners overlap across folds. Section 5 only concedes 'potential challenges in generalization to different modalities of images'; scanner-wise generalization is asserted without a scanner-disjoint evaluation. Even within the T1-only segmentation setup (Section 3.2), the reported mean IoU of 83.66% and DSC of 91.03% could be substantially inflated by scanner leakage, so the headline claim is not supported as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a two-stage deep learning pipeline for lumbar spine MRI segmentation: a YOLOv8-medium detector localizes the lumbar region, and a DenseNet121 encoder with a Self-ONN decoder segments five lumbar vertebrae and four intervertebral discs into nine classes. The authors evaluate on a public multi-scanner, multi-modal dataset (12 scanners, 34 subjects) using subject-preserved 10-fold cross-validation, reporting a mean IoU of 83.66%, sensitivity of 91.44%, and DSC of 91.03% for the segmentation stage, along with a YOLOv8 ROI detection mAP of 0.916. The paper claims the system is 'machine-agnostic', generalizing across scanners and addressing the challenge of diverse MRI modalities.","tokens_in":17140,"tokens_out":7578,"duration_ms":60656,"significance":"If the reported performance holds under a stricter evaluation, the cascaded design and the use of Self-ONN for spine segmentation could be a useful engineering contribution. The use of a public multi-scanner dataset, subject-preserved cross-validation, and comparisons with several decoder architectures are commendable. However, the central claims are not yet substantiated: the segmentation model is trained only on T1-weighted scans, and the cross-validation does not prevent scanner overlap between training and test folds. Therefore, the 'machine-agnostic' and multi-modal generalization claims are not supported by the evidence presented. The paper also contains a contradictory novelty claim and several technical errors that need correction.","major_comments":[{"comment":"The segmentation model is trained and evaluated exclusively on T1-weighted images (T1 non-contrast, T1 contrast-enhanced, T1 fat-saturated), while the abstract and introduction claim that the method addresses 'diverse MRI modalities' and is 'machine-agnostic'. The paper explicitly states that T2, STIR, and Dixon scans were excluded because of limited sample size, and the model is never tested on those modalities. This is a load-bearing overstatement: the claim should be restricted to T1-weighted multi-scanner segmentation, or the authors must provide evidence of performance on the other modalities.","section":"Section 3.2, Table 2"},{"comment":"The 10-fold cross-validation is subject-preserved, as stated in Section 4.2, but it is not scanner-disjoint. Because multiple subjects in the dataset are scanned on the same scanner (Table 1 reports 34 subjects and 12 scanners), the same scanner's intensity characteristics can appear in both training and test folds. This allows the model to exploit scanner-specific appearance rather than learn general anatomy, potentially inflating the reported DSC and IoU. To support the 'machine-agnostic' claim, the authors should run a scanner-disjoint cross-validation or, at minimum, report per-scanner performance and the degree of scanner overlap across folds.","section":"Section 4.2, Table 3"},{"comment":"The ROI detector is trained on only 2 subjects (9 scans, 160 slices) and validated/tested on one subject each. With such a small training set, the reported mAP of 0.916 and the downstream segmentation performance are not convincing evidence of robust localization across 12 scanners. The authors should either increase the ROI training set (if possible) or explicitly discuss the statistical limitations of this split.","section":"Section 4.2, Table 3"},{"comment":"The claim that this is 'the first of its kind' in segmenting individual lumbar vertebral bodies and intervertebral discs is contradicted by the same paper's Table 7, which lists Yilizati-Yilihamu et al. (2023) as segmenting 10 vertebrae and 9 IVDs individually. The novelty statement should be revised to a precise and defensible claim, e.g., regarding the cascaded Self-ONN approach on this dataset, rather than being the first to segment individual lumbar structures.","section":"Section 1 and Section 4.6, Table 7"},{"comment":"The preprocessing removes 20% of slices from the beginning and end of each scan (heuristically chosen) and excludes 53 corrupted scans (about 25% of the dataset). Both decisions may preferentially remove hard or ambiguous cases and thus bias the reported metrics upward. No analysis is provided to assess this bias. The authors should either quantify the impact of these exclusions (e.g., by reporting metrics with and without them, if feasible) or explicitly acknowledge this limitation in the discussion.","section":"Section 3.2 and Section 4.1"}],"minor_comments":[{"comment":"The accuracy formula is incorrect: it should be (TP+TN)/(TP+FP+TN+FN), not (TP+FP)/total. Please correct the definition.","section":"Equation (9)"},{"comment":"The Dice similarity coefficient formula has an error in the denominator: it should be |X|+|Y| (the sum of the two set sizes), not their union. Please fix the equation.","section":"Equation (14)"},{"comment":"The experimental setup paragraph contains an out-of-place sentence about 'a sizable dataset of pediatric respiratory sounds', which appears to be a copy-paste error from another manuscript. This should be removed or corrected.","section":"Section 4.4"},{"comment":"The term 'Lumber vertebrae' is a typo; it should be 'Lumbar vertebrae'.","section":"Keywords and Title"},{"comment":"To enable reproduction, please provide full training details: loss function (e.g., Dice loss, cross-entropy), optimizer, learning rate schedule, batch size, number of epochs, input resolution, and augmentation parameters.","section":"Section 3.5"},{"comment":"The comparison with prior work would be strengthened by reporting confidence intervals or per-class performance, especially because the dataset is small and folds vary in size.","section":"Section 4.6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central claims are notably stronger than the evidence. The 'machine-agnostic' and multi-modal claims are unsupported because only T1 images are segmented and the cross-validation is not scanner-disjoint. The novelty claim is contradicted by the paper's own related-work table. However, the underlying engineering may have merit if evaluated honestly. I recommend major revision with the expectation that the claims are scaled back and a scanner-aware evaluation is added, if feasible. Given the small dataset, the authors should also provide a more thorough analysis of the limitations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline is that this is a competent engineering paper whose central claims are two sizes too big. The cascade—YOLOv8m ROI detector plus DenseNet121 encoder and Self-ONN decoder—is new on this public multi-scanner dataset, and the reported 91.03% DSC/83.66% IoU are honestly measured with subject-preserved 10-fold CV. The ceiling analysis and ablations are a plus; they show the ROI stage costs about 2.7 DSC and that preprocessing matters. That part is useful and reproducible.\n\nThe soft spots are in the framing, not the code. 'Machine-agnostic' is only supported for T1-weighted images: T2, STIR, and Dixon are used in ROI detection but the segmentation model never sees them. So at best you have scanner-agnostic T1 segmentation. Worse, the CV is subject-preserved but not scanner-disjoint. With 34 subjects and 12 scanners, a scanner can contribute subjects to both training and test folds, which lets the model memorize scanner-specific intensity/artifact patterns. The reported numbers can therefore overstate cross-scanner generalization. The paper needs a scanner-disjoint split or per-scanner held-out analysis before that claim flies.\n\nThe 'first of its kind' novelty claim is contradicted by their own Related Works: SAFNet (Yilizati-Yilihamu 2023) already segments individual vertebrae and discs, and Spine-GAN (Han 2018) segments multiple spinal structures. That sentence should be deleted. The preprocessing—dropping 20% of slices and 53 corrupted scans—could bias accuracy upward, but the authors disclose it; I'd call that a minor issue, not a fatal one. The comparison table mixes datasets and class definitions, and the only same-dataset comparison (Kim et al. 2024) is within 1 point on DSC while segmenting only discs, so 'excellent performance' relative to SOTA is not really established.\n\nWho gets value: groups working on lumbar MRI segmentation who want a baseline on the multi-scanner dataset. It deserves a serious referee—the empirical work is real and the dataset choice is useful—but the decision should be major revision, with scanner-disjoint evaluation and scaled-back claims.","headline":"Competent lumbar MRI segmentation pipeline whose 'machine-agnostic' and 'first of its kind' claims outrun the evidence; scanner-disjoint evaluation needed.","tokens_in":17769,"tokens_out":3454,"would_cite":false,"duration_ms":30714,"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":"A cascaded YOLOv8 plus Self-ONN model segments each lumbar vertebra and disc across 12 MRI scanners.","keywords":["lumbar spine MRI segmentation","intervertebral disc segmentation","vertebral body segmentation","Self-ONN","generative neurons","YOLOv8","machine-agnostic","multi-scanner MRI"],"falsifier":"Run the trained cascade on held-out T2, STIR, and Dixon volumes, or on a scanner absent from the 12 used, without any slice trimming, and measure per-class Dice and IoU; a substantial drop would falsify the machine-agnostic claim.","tokens_in":16718,"feed_emoji":"🩻","tokens_out":6713,"duration_ms":56457,"temperature":0.7,"pith_summary":"This paper tries to establish that lumbar spine MRI segmentation can be made machine-agnostic: one automated pipeline can locate and outline each lumbar vertebral body (L1–L5) and intervertebral disc (L1-2 through L4-5) in images from many different scanners. The proposed method is a cascade: a YOLOv8 medium detector first crops the lumbar region from each slice, and a Self-ONN-based encoder-decoder network with a DenseNet121 encoder then produces nine-class segmentation masks. On a 34-subject, 12-scanner dataset, evaluated with subject-wise 10-fold cross-validation, the authors report a mean Intersection over Union of 83.66%, a Dice similarity coefficient of 91.03%, and a sensitivity of 91.44%. They also report that the cascaded design matters: using the detected ROI improves Dice by roughly 7 points and IoU by roughly 10 points over running the segmenter on full images, and a ceiling analysis with perfect ROIs reaches 93.73% Dice. If the claims hold, automated per-level spine measurement and diagnosis support become feasible across scanner vendors without retraining.","feed_headline":"Cascaded AI segments each lumbar vertebra and disc across 12 MRI scanners","feed_subtitle":"ROI-first Self-ONN design hits 91% Dice, a step toward automated spine diagnosis.","key_machinery":"The load-bearing mechanism is the cascade itself plus the generative neurons of Self-ONN. The YOLOv8 medium model localizes the lumbar spine (mAP 0.5:0.95 of 0.916), and the segmenter is a DenseNet121 encoder feeding a five-block Self-ONN decoder with skip connections and no bottleneck. A Self-ONN generative neuron replaces the fixed linear convolution of a CNN with a learnable composite nodal operator expressed as a Taylor polynomial of order Q; the neuron output is the sum of Q convolutions of the input raised to powers 1 through Q, so each neuron can synthesize nonlinear operators during backpropagation. This is what the paper credits for capturing tissue boundaries that low-contrast MRI makes indistinct. The ground-truth masks themselves are formed by fusing manual annotations with the STAPLE algorithm, and the evaluation uses subject-preserved 10-fold cross-validation to prevent data leakage.","core_discovery":"The central discovery is that a two-stage architecture—ROI extraction followed by segmentation—lets a Self-ONN decoder with a DenseNet121 encoder segment individual lumbar vertebrae and intervertebral discs at 91.03% Dice and 83.66% mean IoU, averaged over ten subject-disjoint folds, using T1-weighted images from 12 MRI scanners. This is presented as the first system to segment each lumbar vertebral body and each disc as separate classes rather than treating the spine as one or two structures. The authors also show that the ROI stage is the current bottleneck: when ground-truth ROIs replace detector outputs, Dice rises to 93.73%, whereas removing the ROI stage entirely drops Dice to 74.1%.","pith_inferences":["If the T1-only training set is representative, the machine-agnostic claim should be read as scanner-agnostic for T1; the paper's own limitation note concedes that other MRI modalities remain untested.","The heuristic removal of the first and last 20% of slices and the exclusion of 53 corrupted scans could bias the reported metrics upward, so a prospective test on full, unedited volumes would give a truer clinical picture.","The same ROI-then-Self-ONN recipe could transfer to other anatomical regions where background dominates the image and target structures vary across patients, such as cervical spine or knee cartilage segmentation.","With per-class labels, this type of model could feed downstream decision tools that report exactly which disc level is abnormal, rather than only flagging the presence of a lesion."],"forward_implications":["A single trained pipeline can label each lumbar vertebra and disc separately across scanner vendors, enabling automated level-specific measurements such as disc height, vertebral area, or stenosis indices.","The cascade improves segmentation by roughly 7 Dice points over whole-image segmentation, so localizing the anatomy before pixel classification is a practical strategy for low-contrast medical images.","Because the ceiling analysis shows 93.73% Dice with perfect ROIs, improving the detector or adding slice-level rejection is the clearest path to higher overall accuracy.","The Self-ONN decoder outperforms standard U-Net and U-Net++ decoders with the same DenseNet121 encoder on this dataset, suggesting learnable nonlinear operators help in this modality.","Subject-wise 10-fold cross-validation gives a more realistic estimate of generalization than random splits, and the reported numbers are the averaged per-fold results, not a single lucky test set."],"supporting_citations":[{"why":"Supplies the multi-scanner, multi-modal lumbar MRI dataset with 34 subjects and 12 scanners that the whole pipeline is trained and evaluated on.","marker":"(Khalil, et al., 2022)"},{"why":"Defines Self-ONN generative neurons, the core segmentation mechanism the proposed decoder is built on.","marker":"(Kiranyaz, et al., 2021)"},{"why":"Provides YOLOv8, the ROI detection model whose medium variant is used in the first cascade stage.","marker":"(Jocher, Chaurasia, & Qiu, 2023)"},{"why":"Supplies the DenseNet121 encoder used to extract features for the segmentation decoder.","marker":"(Huang, Liu, Van Der Maaten, & Weinberger, 2017)"},{"why":"Defines the U-Net skip-connection architecture that the Self-ONN decoder adapts and that serves as a baseline.","marker":"(Ronneberger, Fischer, & Brox, 2015)"},{"why":"STAPLE is used to fuse manual segmentation labels into the unified ground-truth masks.","marker":"(Warfield, Zou, & Wells, 2004)"},{"why":"The only prior work on the same dataset and the main point of comparison, though it segments all IVDs as one class.","marker":"(Kim, Park, Lee, & Lee, 2024)"}],"fun_headline_variants":["Cascaded AI labels each lumbar vertebra and disc on 12 MRI scanners","Two-stage Self-ONN achieves 91% Dice on lumbar MRI from 12 scanners","Per-vertebra lumbar MRI: cascaded AI scores 91% Dice across 12 scanners","ROI-first self-ONN segments each lumbar vertebra and disc (91% Dice)"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that T1-weighted slices, after trimming 20% from each volume and deleting 53 corrupted scans, fairly represent the multi-scanner clinical reality the model claims to handle; if the excluded modalities or edge slices behave differently, the machine-agnostic performance numbers could overstate real-world accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Cascaded AI labels each lumbar vertebra and disc on 12 MRI scanners","Two-stage Self-ONN achieves 91% Dice on lumbar MRI from 12 scanners","Per-vertebra lumbar MRI: cascaded AI scores 91% Dice across 12 scanners","ROI-first self-ONN segments each lumbar vertebra and disc (91% Dice)"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000659,"raw_usage":{"total_tokens":3004,"prompt_tokens":926,"completion_tokens":2078,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":1986}},"tokens_in":542,"tokens_out":2078,"duration_ms":14059,"temperature":1.0,"reasoning_tokens":1986,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:03:13.397998+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained cascade on held-out T2, STIR, and Dixon volumes, or on a scanner absent from the 12 used, without any slice trimming, and measure per-class Dice and IoU; a substantial drop would falsify the machine-agnostic claim.","supporting_citations":[],"review_version":1}