{"id":"f4680789-4e76-4b54-937d-59a1d27b07b3","arxiv_id":"2607.28978","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A hybrid 3D CNN + 3D MLP-Mixer classifies chest CT volumes into high/low COVID-19 likelihood with 79.5% accuracy, about 4.7 points above a 3D CNN baseline.","lead":"This paper trains a hybrid deep-learning model that combines a 3D CNN with a 3D MLP-Mixer to sort chest CT volumes into high- or low-likelihood COVID-19 cases, reporting 79.5% accuracy on a 1,205-volume Japanese dataset. It claims the hybrid architecture outperforms a simple 3D CNN baseline by about 4.7 percentage points.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's 302 test cases contradict the stated 80/20 split of 1205 volumes (241 expected); the central accuracy comparison is not reproducible without correcting this.","rationale":"The reader's weakest assumption correctly identifies the evaluation setup, and my reading sharpens it: the arithmetic inside the paper does not close, not merely a missing patient-level split. The proposed and baseline accuracies are single-run numbers computed from an under-specified and internally inconsistent protocol, so the central comparative claim is currently unverified. This is addressable by a corrected, patient-level split with variance reporting, which is why the verdict remains CONDITIONAL rather than being rejected. I credit the authors for a concrete architecture and a confusion matrix that reproduces the stated accuracy; the concern is about reproducibility, not integrity.","tokens_in":4771,"tokens_out":4702,"duration_ms":42892,"concrete_test":"Run a patient-level 80/20 split of the 1205 volumes (expect 241 test volumes) and recompute Table 1 and the 3D CNN baseline on the same split. If the test set is 302, the stated dataset or split must be corrected; if the proposed accuracy minus baseline falls below about 1-2 points or within one standard deviation over 5 random seeds, the claimed advantage is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 reports 1205 CT volumes and an 80/20 train/test split, which implies 241 test volumes. Table 1's confusion matrix sums to 302 (168+18+44+72), so at least one of the dataset size, split ratio, or confusion matrix entries is wrong. Because the proposed accuracy (79.5%) is computed from these numbers, the headline comparison rests on an internally inconsistent evaluation. The baseline is described only as tested on 'the same dataset,' without explicitly stating it used the identical split, so the 4.7-point advantage may compare mismatched test sets. Additionally, the text says training and testing 'cases' were randomly split, not patients; if one patient contributes multiple volumes across both sets, the reported accuracy is optimistically biased. No error bars or repeated splits are reported, so the 4.7-point gap cannot be distinguished from split-to-split variance. The architecture itself is coherent and the confusion matrix does reproduce 79.5%, but the evidence as reported does not establish the improvement claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid classification model for chest CT volumes that combines a 3D CNN feature extractor with a 3D version of the MLP-Mixer. The 3D CNN extracts local volumetric features; these are then divided into non-overlapping 4x4x4 patches and fed into a 3D MLP-Mixer with L=2 layers, DS=1024, DC=256, followed by a two-class output. Preprocessing removes axial slices without lung regions and resizes volumes to 192x192x128. The method is evaluated on 1205 chest CT volumes with an 80/20 train/test split, reporting 79.5% classification accuracy, compared with 74.8% for a 3D CNN with simple MLP layers on the same dataset.","tokens_in":5054,"tokens_out":3336,"duration_ms":27999,"significance":"If the reported improvement is valid, the paper makes a useful architectural contribution to COVID-19 CT volume classification: it demonstrates that a 3D MLP-Mixer head, replacing simple MLP layers after a 3D CNN, can improve classification by exploiting global feature mixing. The motivation is clear and the confusion matrix arithmetic is internally consistent with the reported 79.5%. However, the experimental evidence as presented is not reproducible because of a serious inconsistency between the stated split and the reported test-set size, and because the baseline and evaluation protocol are underspecified. The significance of the 4.7-point improvement therefore cannot be assessed from the manuscript in its current form.","major_comments":[{"comment":"The stated 80/20 split of 1,205 CT volumes implies 241 test volumes (0.2 × 1205 = 241), but the confusion matrix in Table 1 sums to 302 (168+18+44+72). The reported 79.5% accuracy is arithmetically consistent with the matrix ((168+72)/302 = 240/302 = 79.5%), so at least one of the dataset size, the split ratio, or the test-set counts is incorrect. This is load-bearing: the headline accuracy and the comparison with the baseline both depend on this number. Please correct the discrepancy and recompute all reported results.","section":"Section 3, Table 1"},{"comment":"The paper states that 'separations of training and testing cases were randomly performed,' but it does not state that the split was performed at the patient level. If a patient contributes multiple CT volumes, volumes from the same patient can appear in both training and testing, which would inflate the reported accuracy. Please report the number of patients, the distribution of volumes per patient, and confirm that no patient appears in both training and testing sets.","section":"Section 3"},{"comment":"The baseline 3D CNN model is severely underspecified: no architecture, number of layers, number of channels, pooling strategy, or training detail is given beyond '3D CNN part and simple MLP layers.' The text says it was trained and tested 'using the same dataset,' but does not explicitly confirm it used the identical split. Without these details, the 4.7-point improvement over the baseline cannot be evaluated. Please specify the baseline architecture precisely and confirm that it was evaluated on the exact same training/testing partition.","section":"Section 3, baseline description"},{"comment":"Only a single random split is reported, with no confidence interval, no repeated-split variance, and no statistical significance test. The reported 4.7-point gap may be within split-to-split variability for a dataset of this size. Please report the mean and standard deviation over multiple random splits, or at least a bootstrap confidence interval, so the improvement can be properly assessed.","section":"Section 3, experimental protocol"}],"minor_comments":[{"comment":"There is a grammatical error: 'Development of a computer aided diagnosis (CAD) system for COVID-19 is pressing demanded' should be 'is urgently demanded' or 'is a pressing need.'","section":"Abstract / Section 1"},{"comment":"The notation is unclear: 'A set of all patches generated from V 3D feature volumes, S 3D patches are generated from each 3D feature volume, can be represented as X ∈ R^{SV×C}.' Please define V and S explicitly, and clarify whether S is the number of patches per volume or the total number of patches.","section":"Section 2.3.2"},{"comment":"The preprocessing relies on a lung segmentation method from the authors' previous work (reference [6]) without independent validation. Since an inaccurate lung mask could affect the extracted sub-volume and downstream classification, please comment on the segmentation accuracy or provide a visual/quantitative validation.","section":"Section 2.2"},{"comment":"The description of the 3D MLP-Mixer does not specify whether residual connections, layer normalization, dropout, or other regularization techniques from the original MLP-Mixer are included. Such details are necessary for reproducibility.","section":"Section 2.3.1 / 2.3.2"},{"comment":"The training hyperparameters report minibatch size, epochs, and learning rate, but not weight decay, learning-rate schedule, or any data augmentation. Please provide the full training protocol.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"The inconsistency between the stated 80/20 split and the confusion matrix sum (302 vs 241) is a fundamental evaluation error that should not have survived the authors' own checks. The paper is more suitable for a workshop/proceedings venue, but it needs a re-analysis of the experimental protocol, a corrected split, patient-level separation, and a properly specified baseline before it can be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful idea here is simple and real: take the MLP-Mixer, make it 3D, and bolt it onto a 3D CNN feature extractor so you get global mixing without the ViT data appetite. That is a legitimate engineering contribution, and the paper explains well why it might matter for COVID CT, where ground-glass opacities are scattered and both local and global context matter. The confusion matrix in Table 1 is arithmetically consistent with the claimed 79.5%, so the method itself is not internally broken.\n\nThe problem is the evaluation does not tie out. Section 3 says 1205 volumes and an 80/20 split, which gives 241 test cases. Table 1 sums to 302. One of those numbers is wrong, and the baseline is underspecified: 'the same dataset' does not say same split, same preprocessing, or even the same label set. No error bars, no repeated splits, and the split is by volume, not by patient, which risks leakage. Any one of these is fixable; together they mean the 4.7-point advantage over the 3D CNN is not currently established.\n\nTo give credit where it is due, the architecture is described clearly, the hyperparameters are listed, and the paper is honest that 79.5% is below clinical use. The authors do not oversell the clinical impact. The preprocessing step uses a lung segmentation from the same group without independent validation, which is a minor concern because both models likely receive the same preprocessing, so the comparison would be biased only if the segmentation interacts with the classifier in unexpected ways.\n\nBottom line: this is a modest but legitimate architectural contribution, and the issues are all addressable. I would send it to a serious referee, not desk-reject it, but I would ask the authors to correct the split arithmetic, report variance across repeated splits, use a patient-level split, and specify the baseline precisely. If those fixes land, the paper becomes useful. As it stands, the accuracy claim should not be taken at face value. For my own work, I might mention the 3D MLP-Mixer idea, but I would not cite the 79.5% number until the evaluation is cleaned up.","headline":"A legit but modest 3D MLP-Mixer+CNN architecture; the evaluation has an arithmetic inconsistency that keeps the headline 4.7-point gain from being trusted as-is.","tokens_in":5538,"tokens_out":1826,"would_cite":false,"duration_ms":15388,"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 that a hybrid model combining a 3D CNN with a 3D MLP-Mixer classifies COVID-19 chest CT volumes with 79.5% accuracy, outperforming a conventional 3D CNN baseline.","keywords":["COVID-19","chest CT","3D CNN","MLP-Mixer","volume classification","computer-aided diagnosis","hybrid model","deep learning"],"falsifier":"Re-run the same experiment with a strict patient-level split, ensuring no CT volume from a patient appears in both training and testing sets; if the hybrid model's accuracy falls to or below the 3D CNN baseline (74.8%), the claimed benefit of the 3D MLP-Mixer for COVID-19 classification is refuted.","tokens_in":4697,"feed_emoji":"🩺","tokens_out":2547,"duration_ms":20998,"temperature":0.7,"pith_summary":"This paper tries to establish that a hybrid neural network, which pairs a 3D convolutional neural network with a 3D adaptation of the MLP-Mixer, classifies chest CT volumes as high or low likelihood of COVID-19 more accurately than a standard 3D CNN. The motivation is that COVID-19 pneumonia produces both local signs (ground-glass opacities) and global patterns spread across large lung areas, so classification should use both local and volume-wide features. On a dataset of 1,205 CT volumes, the proposed hybrid model reached 79.5% accuracy, about 4.7 percentage points higher than a 3D CNN baseline. If the result holds, it offers a practical CAD approach for COVID-19 diagnosis that works with limited training data.","feed_headline":"Hybrid 3D model hits 79.5% accuracy on COVID-19 CT scans","feed_subtitle":"Combining 3D CNN local features with MLP-Mixer global mixing beats a plain 3D CNN by about five points.","key_machinery":"The key mechanism is a 3D extension of the MLP-Mixer, a Vision-Transformer-like architecture built entirely from MLPs. In this hybrid model, a 3D CNN first extracts a 3D feature volume from the input CT sub-volume (192x192x128 voxels, limited to lung-containing slices). The feature volume is divided into non-overlapping 4x4x4 patches, each reshaped into a 64-dimensional vector; the resulting Sx64 matrix is fed into the 3D MLP-Mixer with L=2 layers, token-mixing width 1024 and channel-mixing width 256. The 3D MLP-Mixer alternates between mixing information across patches (global) and across channels (local), enabling classification that weighs both local and volume-wide features. Additionally","core_discovery":"The central claim is that a hybrid 3D CNN + 3D MLP-Mixer architecture improves COVID-19 classification from chest CT volumes over a plain 3D CNN. The 3D CNN extracts local features from overlapping spatial regions, while the 3D MLP-Mixer, applied to non-overlapping 4x4x4 patches of these features, mixes information across all patches (token-mixing) and within each patch (channel-mixing) to capture global context. In the authors' evaluation with 1,205 CT volumes (80% training, 20% testing), the hybrid model achieved 79.5% accuracy, with a confusion matrix of 168 true high-likelihood, 72 true low-likelihood, 18 false positives, and 44 false negatives. The baseline 3D CNN with simple MLP layers","pith_inferences":["A likely next test is to evaluate the same hybrid architecture on an external, publicly available COVID-19 CT dataset with patient-level splitting; here, I infer the accuracy gap over a 3D CNN may shrink or widen depending on how much the MLP-Mixer's global mixing generalizes across scanner and population differences.","I infer that the hybrid model's advantage over a plain CNN will be most pronounced for cases with diffuse, bilateral ground-glass opacities rather than focal lesions, because the MLP-Mixer's token-mixing explicitly aggregates evidence from distant lung regions.","A practical extension the authors did not explore is using the 3D MLP-Mixer's intermediate feature vectors for localization or severity grading, not just binary classification; the token-mixing weights could be mapped back to spatial regions to highlight which lung areas drove the decision.","Since the paper uses only binary high/low likelihood, a natural follow-up would be to train the same architecture on multi-class labels (e.g., normal, COVID-19, other pneumonia), which would test whether the global-mixing benefit generalizes beyond a two-way decision."],"forward_implications":["If the hybrid model's 79.5% accuracy is reproducible, it provides a computer-aided diagnosis tool that can flag high-likelihood COVID-19 cases from chest CT volumes, helping triage during staff shortages.","The result suggests that MLP-Mixer-style global feature mixing can be successfully adapted from 2D images to 3D medical volumes, opening a path for applying MLP-Mixer to other volumetric imaging tasks.","The hybrid design avoids the large-data requirement of pure Vision Transformers or MLP-Mixers, implying that this architecture class is usable in medical domains where training sets are typically limited to thousands or hundreds of cases.","The 4.7-point accuracy gain over a 3D CNN baseline indicates that global feature integration carries diagnostic information beyond local texture, aligning with the clinical observation that viral pneumonia lesions span large lung regions.","The confusion matrix (168/186 high-likelihood and 72/116 low-likelihood correctly classified) gives a concrete operating point for future calibration if the model is deployed in a clinical workflow."],"fun_headline_variants":["3D CNN + MLP-Mixer hybrid beats plain 3D CNN on COVID CT","Hybrid 3D net scores 79.5% on COVID-19 CT classification","79.5% accuracy: Hybrid 3D CNN-MLP-Mixer for COVID-19 CT volumes","Local+global features in 3D hybrid boost COVID-19 CT detection","Hybrid 3D net hits 79.5% on COVID-19 CT, beats standard CNN"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the random 80/20 split of the 1,205 CT volumes produced test volumes that did not come from the same patients as training volumes; the paper does not state that a patient-level split was performed, so within-patient leakage could inflate the reported accuracy and the comparison to the baseline.","fun_headline_variants_meta":{"raw":{"variants":["3D CNN + MLP-Mixer hybrid beats plain 3D CNN on COVID CT","Hybrid 3D net scores 79.5% on COVID-19 CT classification","79.5% accuracy: Hybrid 3D CNN-MLP-Mixer for COVID-19 CT volumes","Local+global features in 3D hybrid boost COVID-19 CT detection","Hybrid 3D net hits 79.5% on COVID-19 CT, beats standard CNN"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001499,"raw_usage":{"total_tokens":5901,"prompt_tokens":844,"completion_tokens":5057,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":4938}},"tokens_in":588,"tokens_out":5057,"duration_ms":27178,"temperature":1.0,"reasoning_tokens":4938,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T16:09:29.260694+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same experiment with a strict patient-level split, ensuring no CT volume from a patient appears in both training and testing sets; if the hybrid model's accuracy falls to or below the 3D CNN baseline (74.8%), the claimed benefit of the 3D MLP-Mixer for COVID-19 classification is refuted.","supporting_citations":[],"review_version":1}