Classification of COVID-19 cases from chest CT volumes using hybrid model of 3D CNN and 3D MLP-Mixer

Kensaku Mori; Masahiro Hashimoto; Masahiro Oda; Shigeki Aoki; Tong Zheng; Toshiaki Akashi; Yoshito Otake; Yuichiro Hayashi

REVIEW 4 major objections 5 minor 7 references

Classification of COVID-19 cases from chest CT volumes using hybrid model of 3D CNN and 3D MLP-Mixer

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 2607.28978 v1 pith:35V6YF64 submitted 2026-07-31 cs.CV

Masahiro Oda , Tong Zheng , Yuichiro Hayashi , Yoshito Otake , Masahiro Hashimoto , Toshiaki Akashi , Shigeki Aoki , Kensaku Mori This is my paper · ORCID

classification cs.CV

keywords COVID-19chestCT3DCNNMLP-Mixervolumeclassificationcomputer-aideddiagnosishybridmodeldeeplearning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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

What would settle it

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.

Watch

Extended reading notes

Core claim

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

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

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.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the 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.

The 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.

To 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.

Bottom 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.

Referee Report

4 major / 5 minor

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.

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 (4)

[Section 3, Table 1] 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 3] 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 3, baseline description] 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 3, experimental protocol] 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.

minor comments (5)

[Abstract / Section 1] 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 2.3.2] 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 2.2] 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 2.3.1 / 2.3.2] 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 3] 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.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation; the accuracy claim rests on a held-out empirical evaluation, with only minor non-load-bearing self-citation.

full rationale

This paper reports a supervised learning experiment rather than an analytic derivation, so most circularity patterns do not apply. The central claim ('Classification accuracy ... 79.5% ... higher than ... 3D CNN models') is supported by training on 80% and testing on 20% of 1205 CT volumes, not by fitting the test set or by defining the metric in terms of the result. The MLP-Mixer portion is an engineering extension (3D patches of 4x4x4, C=64) and does not smuggle in the claimed accuracy. The only same-author citation is reference [6] for lung-region preprocessing; this is a preprocessing step, not the classification accuracy or the comparison, so it is minor and not load-bearing in the circularity sense. The manuscript's substantive weakness is a reporting inconsistency: an 80/20 split of 1205 volumes implies about 241 test volumes, while Table 1's confusion matrix sums to 302 (168+18+44+72); the text also does not state a patient-level split. These are correctness/reproducibility risks, not circularity: nothing in the paper's equations reduces the claimed result to its own inputs. Accordingly, circularity score is low.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard deep-learning assumptions plus dataset/labeling assumptions that are not independently validated. The free parameters are hand-chosen architecture and training choices with no sensitivity analysis; no new physical entities are introduced.

free parameters (6)

Input volume scaling = 192×192×128 voxels
Chosen normalization for all CT volumes; not justified or ablated.
3D patch size = 4×4×4
Patch size for tokenization in the 3D MLP-Mixer; determines C=64; chosen by hand.
Number of mixer layers L = 2
Hyperparameter chosen by hand; no sensitivity analysis.
Token-mixing width DS = 1024
Hyperparameter chosen by hand; no sensitivity analysis.
Channel-mixing width DC = 256
Hyperparameter chosen by hand; no sensitivity analysis.
Training hyperparameters = minibatch 5, epochs 50, Adam LR 5e-6
Selected without reported tuning procedure or sensitivity analysis.

assumptions (4)

standard math Adam optimization and backpropagation converge to a model that generalizes from the training split
All trained deep models rely on this standard assumption.
domain assumption Lung segmentation method [6] correctly identifies lung regions in all CT volumes
Preprocessing removes non-lung slices; if segmentation fails, the input to the classifier is corrupted. This method is self-cited and not independently evaluated in this paper.
domain assumption Radiologist-assigned high/low likelihood labels are correct and consistent across institutions
Ground truth for training and testing; no inter-observer agreement or label reliability analysis is reported.
domain assumption Random split separates patients or volumes with no leakage between train and test
Paper says random split of volumes, not patient-level; no statement preventing same-patient overlap in both sets.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Classification of COVID-19 cases from chest CT volumes using hybrid model of 3D CNN and 3D MLP-Mixer." pith.science (2026). https://pith.science/paper/35V6YF64

@misc{pith2026260728978,
  author       = {Pith},
  title        = {Pith review of: Classification of COVID-19 cases from chest CT volumes using hybrid model of 3D CNN and 3D MLP-Mixer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/35V6YF64}},
  note         = {Machine review of arXiv:2607.28978}
}

read the original abstract

This paper proposes an automated classification method of COVID-19 chest CT volumes using improved 3D MLP-Mixer. Novel coronavirus disease 2019 (COVID-19) spreads over the world, causing a large number of infected patients and deaths. Sudden increase in the number of COVID-19 patients causes a manpower shortage in medical institutions. Computer-aided diagnosis (CAD) system provides quick and quantitative diagnosis results. CAD system for COVID-19 enables efficient diagnosis workflow and contributes to reduce such manpower shortage. In image-based diagnosis of viral pneumonia cases including COVID-19, both local and global image features are important because viral pneumonia cause many ground glass opacities and consolidations in large areas in the lung. This paper proposes an automated classification method of chest CT volumes for COVID-19 diagnosis assistance. MLP-Mixer is a recent method of image classification using Vision Transformer-like architecture. It performs classification using both local and global image features. To classify 3D CT volumes, we developed a hybrid classification model that consists of both a 3D convolutional neural network (CNN) and a 3D version of the MLP-Mixer. Classification accuracy of the proposed method was evaluated using a dataset that contains 1205 CT volumes and obtained 79.5% of classification accuracy. The accuracy was higher than that of conventional 3D CNN models consists of 3D CNN layers and simple MLP layers.

Figures

Figures reproduced from arXiv: 2607.28978 by the authors.

**Figure 2.** Axial slices of CT volumes. Infection regions are indicated by arrows. (a) cases identified as high likelihood [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

7 extracted references · 1 linked inside Pith

[1]

Coronavirus Update

“Coronavirus Update.”https://www.worldometers.info/coronavirus/. (Accessed: January 19, 2023)

2023
[2]

U., Abbara, S., Bhalla, S., Chung, J

Simpson, S., Kay, F. U., Abbara, S., Bhalla, S., Chung, J. H., Chung, M., Henry, T. S., Kanne, J. P., Kligerman, S., Ko, J. P., and Litt, H., “Radiological society of north america expert consensus statement on reporting chest CT findings related to COVID-19. endorsed by the society of thoracic radiology, the american college of radiology, and RSNA,”Radio...

2020
[3]

Correlation of chest CT and RT-PCR testing for coronavirus disease 2019 (COVID-19) in china: A report of 1014 cases,

Ai, T., Yang, Z., Hou, H., Zhan, C., Chen, C., Lv, W., Tao, Q., Sun, Z., and Xia, L., “Correlation of chest CT and RT-PCR testing for coronavirus disease 2019 (COVID-19) in china: A report of 1014 cases,” Radiology296(2) (2020)

2019
[4]

An image is worth 16x16 words: Trans- formers for image recognition at scale,

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Min- derer, M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby, N., “An image is worth 16x16 words: Trans- formers for image recognition at scale,”ICLR 2021(2021)

2021
[5]

MLP-Mixer: An all-mlp architecture for vision,

Tolstikhin, I., Houlsby, N., Kolesnikov, A., Beyer, L., Zhai, X., Unterthiner, T., Yung, J., Steiner, A., Keysers, D., Uszkoreit, J., Lucic, M., and Dosovitskiy, A., “MLP-Mixer: An all-mlp architecture for vision,” arXiv:2105.01601(2021)

arXiv 2021
[6]

Lung infection and normal region segmentation from CT volumes of COVID-19 cases,

Oda, M., Hayashi, Y., Otake, Y., Hashimoto, M., Akashi, T., and Mori, K., “Lung infection and normal region segmentation from CT volumes of COVID-19 cases,”Proceedings of SPIE Medical Imaging11597, 115972X–1–6 (2021)

2021
[7]

A multitask learning architecture for simultaneous segmentation of bright and red lesions in fundus images,

Playout, C., Duval, R., and Cheriet, F., “A multitask learning architecture for simultaneous segmentation of bright and red lesions in fundus images,”International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI)11071, 101–108 (2018). (a) (b) Figure 2. Axial slices of CT volumes. Infection regions are indicated by arrows. ...

2018

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

Coronavirus Update

“Coronavirus Update.”https://www.worldometers.info/coronavirus/. (Accessed: January 19, 2023)

2023

[2] [2]

U., Abbara, S., Bhalla, S., Chung, J

Simpson, S., Kay, F. U., Abbara, S., Bhalla, S., Chung, J. H., Chung, M., Henry, T. S., Kanne, J. P., Kligerman, S., Ko, J. P., and Litt, H., “Radiological society of north america expert consensus statement on reporting chest CT findings related to COVID-19. endorsed by the society of thoracic radiology, the american college of radiology, and RSNA,”Radio...

2020

[3] [3]

Correlation of chest CT and RT-PCR testing for coronavirus disease 2019 (COVID-19) in china: A report of 1014 cases,

Ai, T., Yang, Z., Hou, H., Zhan, C., Chen, C., Lv, W., Tao, Q., Sun, Z., and Xia, L., “Correlation of chest CT and RT-PCR testing for coronavirus disease 2019 (COVID-19) in china: A report of 1014 cases,” Radiology296(2) (2020)

2019

[4] [4]

An image is worth 16x16 words: Trans- formers for image recognition at scale,

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Min- derer, M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby, N., “An image is worth 16x16 words: Trans- formers for image recognition at scale,”ICLR 2021(2021)

2021

[5] [5]

MLP-Mixer: An all-mlp architecture for vision,

Tolstikhin, I., Houlsby, N., Kolesnikov, A., Beyer, L., Zhai, X., Unterthiner, T., Yung, J., Steiner, A., Keysers, D., Uszkoreit, J., Lucic, M., and Dosovitskiy, A., “MLP-Mixer: An all-mlp architecture for vision,” arXiv:2105.01601(2021)

arXiv 2021

[6] [6]

Lung infection and normal region segmentation from CT volumes of COVID-19 cases,

Oda, M., Hayashi, Y., Otake, Y., Hashimoto, M., Akashi, T., and Mori, K., “Lung infection and normal region segmentation from CT volumes of COVID-19 cases,”Proceedings of SPIE Medical Imaging11597, 115972X–1–6 (2021)

2021

[7] [7]

A multitask learning architecture for simultaneous segmentation of bright and red lesions in fundus images,

Playout, C., Duval, R., and Cheriet, F., “A multitask learning architecture for simultaneous segmentation of bright and red lesions in fundus images,”International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI)11071, 101–108 (2018). (a) (b) Figure 2. Axial slices of CT volumes. Infection regions are indicated by arrows. ...

2018