Pith. sign in

REVIEW 5 major objections 4 minor 21 references

Small Lesions-aware Bidirectional Multimodal Multiscale Fusion Network for Lung Disease Classification

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read MMCAF-Net claims that making small 3D lesions visible and aligning image/table dimensions improves lung-cancer subtype classification beyond six multimodal baselines.

desk verdict Plausible fusion architecture with a real public-dataset test, but the single-split evaluation and an AUROC that trails one baseline undermine the 'surpassing SOTA' claim. read the letter →

arxiv 2508.04205 v1 pith:DZV4UAKD submitted 2025-08-06 cs.CV

classification cs.CV
keywords multimodalfusionsmalllesiondetectionlungcancerclassification3Dmedicalimagingcross-attentionmultiscalefeaturesPET-CTtabularclinicaldata
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

The paper sets out to show that two distinct failure modes in lung cancer diagnosis — small lesions that standard 3D encoders overlook and dimensional mismatch between volumetric images and tabular clinical records — can be addressed in one network. It proposes MMCAF-Net, whose image branch wraps the PENet encoder in a feature pyramid with an efficient 3D multi-scale attention module (E3D-MSCA) and whose fusion branch uses multiscale cross-attention with bidirectional scale weighting to merge image and table features. On the Lung-PET-CT-Dx dataset, the paper reports the best accuracy, F1, specificity, positive predictive value, and negative predictive value among seven multimodal methods, while acknowledging an AUROC 1.6 points below MMTM. A sympathetic reader would care because the target task, distinguishing adenocarcinoma from squamous cell carcinoma in small PET/CT lesions, is clinically hard, and the paper's design targets the lesion-size and alignment problems directly rather than adding model capacity alone.

What carries the argument

The central machinery is the MMCAF-Net architecture, and the load-bearing component is E3D-MSCA — a 3D extension of a multi-scale convolutional attention module made of a channel attention block, a spatial attention block, and a depth-wise convolution fusion block. It sits inside a feature pyramid over the PENet encoder, with a Bidirectional Feedback Propagation Unit fusing two pyramid scales. Tabular features are encoded by a Kolmogorov-Arnold Network, a network built from learnable univariate spline activations. The fusion module, also called MSCA, projects image and table features through three inverted-pyramid dimension levels, applies multi-head cross-attention at each level, and then p

What would settle it

Run the same comparison on ten different train/validation/test splits that keep the class proportions, and report confidence intervals for ACC, F1, AUROC, PPV, and NPV. If the 10-15 point edges over MMTM and PECon shrink to within the interval width or flip sign in some splits, the claimed superiority is not established; computing the same metrics on the original 251:61 class distribution without oversampling, at a matched operating point, would also test whether the gains are a threshold or prior artifact.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that MMCAF-Net classifies lung adenocarcinoma versus squamous cell carcinoma on CT/PET images plus tabular clinical data better than six recent multimodal baselines. The reported numbers are ACC 0.791, F1 0.690, specificity 0.857, sensitivity 0.667, PPV 0.714, NPV 0.828, and AUROC 0.786; MMTM reaches 0.802 AUROC, so the paper's superiority is on the threshold-based metrics, where the margins are roughly 10 percentage points in ACC and F1 and 15 in PPV. The ablation studies are the paper's evidence for mechanism: adding E3D-MSCA to PENet improves AUROC by 10 points over SAM+E3D-MSCA and 15 points over PENet, and the MSCA fusion module improves ACC by 5-12 po

Load-bearing premise

The load-bearing premise is that the single train/validation/test split — with minority-class training oversampled from 34 to 198 cases and only 15 SCC cases in the test set — gives a representative and stable estimate of performance.

Editorial extensions

If this is right

  • On the reported split, MMCAF-Net achieves ACC 0.791, F1 0.690, specificity 0.857, sensitivity 0.667, PPV 0.714, and NPV 0.828, higher than all six compared methods on those metrics.
  • The E3D-MSCA image encoder is the largest single contributor: replacing PENet raises AUROC by 10-15 points and F1 by 10-14 points in the ablation.
  • The MSCA fusion module contributes separately: its ablation gains of 5-12 points in ACC over cross-attention, CLIP fusion, and late fusion support the paper's claim that handling dimensional mismatch matters.
  • The high specificity and NPV relative to baselines imply fewer false positives in screening-oriented use, which is the paper's stated practical benefit.
  • AUROC remains 0.786 versus 0.802 for MMTM, so the paper's superiority claim is metric-specific rather than universal.

Reading between the lines

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

  • The paper does not test transfer to other small-lesion volumetric tasks with tabular records; the same encoder and fusion blocks could plausibly be dropped into pulmonary embolism or nodule follow-up, but that remains an open experiment.
  • With only 15 squamous-cell carcinoma cases in the test set, a repeated stratified split experiment is the natural check on the claimed 10-15 point margins; the paper reports a single split.
  • The metric pattern — higher ACC, F1, PPV, and NPV but not AUROC than MMTM — suggests threshold placement, not just ranking, drives the reported gains; comparing at matched operating points would separate the two.
  • Training with oversampled minority data changes the prior; reporting metrics on the original 251:61 class distribution would give prevalence-realistic PPV and NPV.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper proposes MMCAF-Net, a multimodal fusion network for lung disease classification on the Lung-PET-CT-Dx dataset. The network combines a 3D image encoder (PENet augmented with an Efficient 3D Multi-Scale Convolutional Attention module, E3D-MSCA, and a Bidirectional Feedback Propagation Unit), a KAN-based tabular encoder, and a Multiscale Cross-Attention (MSCA) fusion module with a Bidirectional Scale Fusion (BSF) step. The authors report binary classification results for adenocarcinoma versus squamous cell carcinoma, claiming improved diagnostic accuracy and state-of-the-art performance. The manuscript includes quantitative comparisons with six prior multimodal methods, ablation studies for the image encoder and fusion module, and qualitative examples of challenging cases.

Significance. If the reported results were reliable, the contribution would be practically relevant: small lesion-aware 3D feature extraction combined with cross-modal fusion is a meaningful direction for PET/CT plus tabular-data diagnosis. A positive aspect is that the code is released, which supports reproducibility. However, the evidence in the paper is not sufficient to establish the central claim. The evaluation rests on a single small test split with no uncertainty quantification; the reported AUROC is not the best in Table 1; and several experimental details needed for an apples-to-apples comparison are missing. The architectural ideas are reasonable and the writing is generally clear, but the experimental validation must be substantially strengthened before the claims can be accepted.

major comments (5)
  1. [§3.1, Table 1] The evaluation is based on a single train/validation/test split with a very small test set: the squamous cell carcinoma test portion contains only 15 samples and the total test set appears to be around 48 patients. No confidence intervals, bootstrap estimates, repeated-seed results, or significance tests (e.g., McNemar) are reported. The 10–15 percentage-point gaps in ACC/F1/PPV over baselines could easily be split noise. Moreover, the AUROC of MMCAF-Net (0.786) is below MMTM (0.802), so the abstract's claim of 'surpassing current state-of-the-art methods' is not supported even by the reported numbers. Please provide multi-split or bootstrapped evaluation and a statistical comparison.
  2. [§3.1, Dataset] The paper states the dataset contains 355 cases but reports only 251 adenocarcinoma and 61 squamous cell carcinoma cases, leaving 43 cases unaccounted for. It is not stated whether cases without both CT/PET and tabular data were excluded, whether other histologies were dropped, or how the final binary cohort was selected. Since the test-set composition is central to interpreting every metric, this curation step must be described explicitly.
  3. [§3.1, Oversampling] Random oversampling increases the SCC training set from 34 to 198 by duplicating minority samples. This can lead to overfitting to the duplicated instances, and the paper provides no sensitivity analysis for this factor or for the other hyperparameters (number of slices, input size, learning-rate schedule). The small 15-sample SCC test set makes any conclusion about sensitivity on the minority class especially fragile. Please report results without oversampling, with class-weighting, or with multiple oversampling ratios, and include class-stratified error bars.
  4. [§3.2, Table 1] The comparison with baselines is not demonstrably fair. No training details are provided for PECon, MedFuse, Drfuse, MMTM, PEfusion, or daft: no information on which backbone is used, whether the same 12-slice input and oversampling protocol are applied, what hyperparameters are chosen, or whether the same train/validation/test split is used. Without this, the claim that MMCAF-Net 'outperforms all comparative approaches' is not substantiated. Please specify the common experimental protocol or, failing that, present results under the same split and training conditions for all methods.
  5. [Ablation study, Tables 2 and 3] The ablation presentation is unclear about what the complete model is. In Table 2, the row 'PENet+E3D-MSCA+drop' reaches AUROC 0.712, but the final model in Table 1/Table 3 has AUROC 0.786; 'drop' is never defined. Table 3 labels the proposed component as 'MSCA_Fusion' but the reported numbers are identical to the full MMCAF-Net in Table 1, which also includes E3D-MSCA and BSF. A clear ablation tree (full model minus each module) with the same evaluation protocol is needed to attribute the gains.
minor comments (4)
  1. [Eq. (10)] The attention formula appears to be missing the scaling division: it should be Attention(Q_h, K_h) = Q_h K_h^T / sqrt(C) rather than the printed expression. Please also clarify the tensor shapes in Eqs. (4)–(9).
  2. [Fig. 1] The figure contains the word 'Frozen' next to PENet, but the text and implementation details never state whether PENet weights are frozen or fine-tuned. Please clarify.
  3. [§3.1, Implementation details] The description says the model was trained for 50 epochs with SGD and reports learning rate and weight decay, but does not state the momentum, learning-rate schedule, or validation-based model selection criterion. These details affect reproducibility.
  4. [§4, Conclusion] The conclusion says the method 'outperforms competing methods' without qualification, while the results in Table 1 show a lower AUROC than MMTM. Please soften the wording to match the evidence, e.g., 'on most metrics'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported results are empirical measurements on an external dataset, and the self-citations are inspirational rather than load-bearing.

full rationale

This paper is an empirical deep-learning evaluation, not a first-principles derivation. The central claim (MMCAF-Net outperforms comparison methods on Lung-PET-CT-Dx) is supported by Table 1, which reports metrics computed on a held-out test split. The paper states the data split explicitly: 'We use a random oversampling strategy to supplement the training set from 34 samples to 198 samples, and the number of verification sets and test sets remains the same, which are 12 and 15 respectively.' Thus the reported ACC/F1/PPV values are not derived from the fitted training parameters by construction; they are independent test-set measurements. The only notable self-citation is in Section 2.2: 'In prior work, Yu et al. utilized a multi-scale group aggregation bridge for feature fusion [19] [20]. Inspired by their approach, we propose a MSCA fusion module.' This is an inspirational citation to the authors' prior work, but it does not supply a uniqueness theorem, a fitted parameter, or a definition that forces the reported outcome. The paper even reports an adverse result relative to its own headline claim: 'While MMCAF-Net falls short of mmtm by 1.6% in AUROC, it surpasses mmtm by approximately 10% in both ACC and F1 scores,' which is an honest metric-level comparison, not circularity. No equation equates a predicted quantity to an input by definition, and no fitted parameter is renamed as a prediction. Accordingly, the derivation chain is self-contained against an external benchmark, and no circular step is present.

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

The claims rest on four domain assumptions about label quality, split representativeness, baseline fairness, and slice selection. The hand-chosen experimental settings are listed as free parameters because they affect the central comparison and are not swept or justified with sensitivity analyses. No physical entities or new conserved quantities are introduced.

free parameters (4)
  • Minority class oversampling factor = SCC training count 34 to 198
    Hand-chosen augmentation of the training set; no sensitivity analysis and duplicated samples may bias training and metrics.
  • Slices per sample = 12
    Chosen without stated justification; the effect on classification is not measured.
  • Input spatial size = 192x192
    Resized from DICOM with normalization; no scale or resolution ablation.
  • Optimization schedule = 50 epochs, SGD, lr=0.0001, weight decay=0.01, batch size=4
    Reported in Section 3.1; no tuning procedure or stability check.
assumptions (4)
  • domain assumption Ground truth labels in Lung-PET-CT-Dx are correct and clinically reliable.
    Section 3.1 relies on the dataset's tumor classification labels without discussing label noise or expert verification.
  • domain assumption A single random split with SCC oversampling yields a representative validation and test set.
    Section 3.1 reports only SCC split counts (34/12/15) and no stratification or repeated-seed protocol, so metric differences could be split artifacts.
  • domain assumption Baseline methods were trained and evaluated under the same preprocessing, split, and oversampling conditions as MMCAF-Net.
    Section 3.2 lists comparison numbers but does not state whether baselines were retrained by the authors on identical folds.
  • domain assumption Processing 12 slices per 3D volume captures the lesion-relevant information.
    Section 3.1 says '12 slices per input sample' but does not define how slices are selected, so the small-lesion sensitivity claim is not directly testable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Small Lesions-aware Bidirectional Multimodal Multiscale Fusion Network for Lung Disease Classification." pith.science (2026). https://pith.science/paper/DZV4UAKD

@misc{pith2026250804205,
  author       = {Pith},
  title        = {Pith review of: Small Lesions-aware Bidirectional Multimodal Multiscale Fusion Network for Lung Disease Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DZV4UAKD}},
  note         = {Machine review of arXiv:2508.04205}
}
read the original abstract

The diagnosis of medical diseases faces challenges such as the misdiagnosis of small lesions. Deep learning, particularly multimodal approaches, has shown great potential in the field of medical disease diagnosis. However, the differences in dimensionality between medical imaging and electronic health record data present challenges for effective alignment and fusion. To address these issues, we propose the Multimodal Multiscale Cross-Attention Fusion Network (MMCAF-Net). This model employs a feature pyramid structure combined with an efficient 3D multi-scale convolutional attention module to extract lesion-specific features from 3D medical images. To further enhance multimodal data integration, MMCAF-Net incorporates a multi-scale cross-attention module, which resolves dimensional inconsistencies, enabling more effective feature fusion. We evaluated MMCAF-Net on the Lung-PET-CT-Dx dataset, and the results showed a significant improvement in diagnostic accuracy, surpassing current state-of-the-art methods. The code is available at https://github.com/yjx1234/MMCAF-Net

Figures

Figures reproduced from arXiv: 2508.04205 by the authors.

Figure 1
Figure 1. Overview of the Proposed MMCAF-Net. The framework consists of an [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Detailed structure of E3D-MSCA. It comprises 3D CAB, 3D SAB, and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Detailed structure of MSCA module. 3 Experiments 3.1 Dataset and Implementation Dataset. We used a publicly available dataset provided by The Cancer Imaging Archive, called Lung-PET-CT-Dx [10], which includes CT or PET) scans and tabular clinical information for 355 cases. The dataset provides tumor classifi￾cation labels for each patient. Most of the PET/CT data is stored in DICOM format, and all data has undergone… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visual Representation of Classification Performance for Challenging [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 14 canonical work pages

  1. [1]

    In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

    Chen, X., Pan, J., Dong, J.: Bidirectional multi-scale implicit neural representa- tions for image deraining. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 25627–25636 (2024)

  2. [2]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Guo, Z., Gan, H.: Cpp-net: Embracing multi-scale feature fusion into deep unfold- ing cp-ppa network for compressive sensing. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 25086–25095 (2024)

  3. [3]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Hager, P., Menten, M.J., Rueckert, D.: Best of both worlds: Multimodal contrastive learning with tabular and imaging data. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 23924–23935 (2023)

  4. [4]

    In: Machine Learning for Healthcare Conference

    Hayat, N., Geras, K.J., Shamout, F.E.: Medfuse: Multi-modal fusion with clini- cal time-series data and chest x-ray images. In: Machine Learning for Healthcare Conference. pp. 479–503. PMLR (2022)

  5. [5]

    NPJ Digital Medicine 3(1), 61 (2020)

    Huang, S.C., Kothari, T., Banerjee, I., Chute, C., Ball, R.L., Borus, N., Huang, A., Patel, B.N., Rajpurkar, P., Irvin, J., et al.: Penet—a scalable deep-learning model for automated diagnosis of pulmonary embolism using volumetric ct imaging. NPJ Digital Medicine 3(1), 61 (2020)

  6. [6]

    Scientific Reports10(1), 22147 (2020)

    Huang, S.C., Pareek, A., Zamanian, R., Banerjee, I., Lungren, M.P.: Multimodal fusion with deep neural networks for leveraging ct imaging and electronic health record: a case-study in pulmonary embolism detection. Scientific Reports10(1), 22147 (2020)

  7. [7]

    In: Forty-first Inter- national Conference on Machine Learning (2024) 10 J

    Jiang, J.P., Ye, H.J., Wang, L., Yang, Y., Jiang, Y., Zhan, D.C.: Tabular insights, visual impacts: transferring expertise from tables to images. In: Forty-first Inter- national Conference on Machine Learning (2024) 10 J. Yu et al

  8. [8]

    In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

    Joze, H.R.V., Shaban, A., Iuzzolino, M.L., Koishida, K.: Mmtm: Multimodal trans- fer module for cnn fusion. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 13289–13299 (2020)

Show all 21 references
  1. [9]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015–4026 (2023)

  2. [10]

    7937/TCIA.2020.NNC2-0461 (2020)

    Li, P., Wang, S., Li, T., Lu, J., HuangFu, Y., Wang, D.: A large-scale ct and pet/ct dataset for lung cancer diagnosis (lung-pet-ct-dx).https://doi.org/10. 7937/TCIA.2020.NNC2-0461 (2020)

  3. [11]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Liu, S., Ma, Y., Zhang, X., Wang, H., Ji, J., Sun, X., Ji, R.: Rotated multi-scale interaction network for referring remote sensing image segmentation. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 26658–26668 (2024)

  4. [12]

    arXiv preprint arXiv:2404.19756 (2024)

    Liu, Z., Wang, Y., Vaidya, S., Ruehle, F., Halverson, J., Soljačić, M., Hou, T.Y., Tegmark, M.: Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756 (2024)

  5. [13]

    In: International Conference on Medi- cal Image Computing and Computer-Assisted Intervention

    Ma, T., Dai, X., Zhang, S., Wen, Y.: Pivit: Large deformation image registration with pyramid-iterative vision transformer. In: International Conference on Medi- cal Image Computing and Computer-Assisted Intervention. pp. 602–612. Springer (2023)

  6. [14]

    In: International Conference on Medi- cal Image Computing and Computer-Assisted Intervention

    Pölsterl, S., Wolf, T.N., Wachinger, C.: Combining 3d image and tabular data via the dynamic affine feature map transform. In: International Conference on Medi- cal Image Computing and Computer-Assisted Intervention. pp. 688–698. Springer (2021)

  7. [15]

    In: International Conference on Machine Learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748–8763. PmLR (2021)

  8. [16]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Rahman, M.M., Munir, M., Marculescu, R.: Emcad: Efficient multi-scale convolu- tional attention decoding for medical image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11769– 11779 (2024)

  9. [17]

    In: International Workshop on Machine Learning in Medical Imaging

    Sanjeev, S., Al Khatib, S.K., Shaaban, M.A., Almakky, I., Papineni, V.R., Yaqub, M.: Pecon: Contrastive pretraining to enhance feature alignment between ct and ehr data for improved pulmonary embolism diagnosis. In: International Workshop on Machine Learning in Medical Imaging...

  10. [18]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Yao, W., Yin, K., Cheung, W.K., Liu, J., Qin, J.: Drfuse: Learning disentangled representation for clinical multi-modal fusion with missing modality and modal inconsistency. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 16416–16424 (2024)

  11. [19]

    In: 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)

    Yu, X., Elazab, A., Ge, R., Jin, H., Jiang, X., Jia, G., Wu, Q., Shi, Q., Wang, C.: Ich-scnet: Intracerebral hemorrhage segmentation and prognosis classification network using clip-guided sam mechanism. In: 2024 IEEE International Conference on Bioinformatics and Biomedicine (...

  12. [20]

    arXiv preprint arXiv:2506.23121 (2025)

    Yu, X., Wang, C., Jin, H., Elazab, A., Jia, G., Wan, X., Zou, C., Ge, R.: Crisp- sam2: Sam2 with cross-modal interaction and semantic prompting for multi-organ segmentation. arXiv preprint arXiv:2506.23121 (2025)

  13. [21]

    In: International Conference on Medi- cal Image Computing and Computer-Assisted Intervention

    Zhu, V., Ji, Z., Guo, D., Wang, P., Xia, Y., Lu, L., Ye, X., Zhu, W., Jin, D.: Low- rank continual pyramid vision transformer: Incrementally segment whole-body or- Small Lesions-aware BMMFN for Lung Disease Classification 11 gans in ct with light-weighted adaptation. In: Inter...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.