Toward Multi-Modal Deep Learning for Pulmonary Disease Classification: A Texture-Based Machine Learning Pilot Study on Public Chest X-Ray Data

Yogisri Pujitha Chinthoti

REVIEW 3 major objections 4 minor 13 references

Toward Multi-Modal Deep Learning for Pulmonary Disease Classification: A Texture-Based Machine Learning Pilot Study on Public Chest X-Ray Data

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper shows that handcrafted texture and gradient features, evaluated with patient-level cross-validation on a small public chest X-ray set, beat the majority-class baseline by only a few points, and positions that honest baseline as th

desk verdict A small, honest pilot with proper patient-level CV and a reproducible baseline; the signal is real but likely confounded by source artifacts, and the over-baseline claim lacks uncertainty estimates. read the letter →

arxiv 2607.27286 v1 pith:FM7PVE3U submitted 2026-07-29 eess.IV cs.CV

Yogisri Pujitha Chinthoti This is my paper

classification eess.IVcs.CV

keywords chestX-rayclassificationCOVID-19pneumoniatexturefeaturesHOGGLCMsupportvectormachinepatient-levelcross-validation

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 pilot study tries to establish whether cheap, handcrafted image features can separate COVID-19 from other pneumonias on a small public chest X-ray collection well enough to serve as an honest baseline. Using histogram-of-oriented-gradients and gray-level co-occurrence-matrix texture descriptors with a support vector machine, and evaluating with patient-level 5-fold cross-validation to prevent leakage, the paper reports 75.4% mean accuracy and 0.755 AUC. These numbers sit only modestly above the 71.6% majority-class baseline and the 0.500 chance AUC, so the author reads them as evidence of a small but real signal, not as clinical readiness. The paper's larger point is that a transparent, reproducible baseline like this should precede any claim about a multi-modal deep learning system for the same task.

What carries the argument

The load-bearing mechanism is a 1,792-dimensional feature vector per image that concatenates histogram-of-oriented-gradients (HOG) descriptors — 9 orientation bins over 16×16-pixel cells with 2×2 blocks — with gray-level co-occurrence matrix (GLCM) statistics (contrast, homogeneity, energy, correlation computed at two distances and three angles) and basic intensity measures (mean, standard deviation, interquartile range). These features are fed to class-balanced classical classifiers, with the RBF SVM performing best. The evaluation mechanism is equally central: patient-level grouped 5-fold stratified cross-validation, using the patient identifier as the grouping variable so that no patient'

What would settle it

Train the same SVM pipeline on an independent, multi-institutional chest X-ray set with verified COVID-19 and pneumonia labels, then test with patient-level splits; if AUC falls to near 0.5 while the majority-class baseline remains around 0.7, the pilot's apparent signal is dataset-specific rather than generalizable.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that texture and gradient features extracted from chest radiographs carry enough discriminative signal to distinguish COVID-19 from other pneumonia etiologies in a small public dataset, but only slightly more than always predicting the majority class. The best classifier, an RBF SVM on concatenated HOG, GLCM, and intensity features, reaches 75.4% mean accuracy and 0.755 AUC under patient-level grouped stratified 5-fold cross-validation, against a 71.6% accuracy majority baseline and 0.500 AUC. The author explicitly frames this as a small effect that justifies, but does not by itself validate, a proposed multi-modal convolutional-transformer arch

Load-bearing premise

The load-bearing premise is that the public chest X-ray collection used here, assembled from heterogeneous case reports, carries labels and images that reflect genuine disease categories rather than source-institution or scanner artifacts.

Editorial extensions

If this is right

A reproducible baseline now exists for this task on public chest X-ray data, so future deep learning studies can measure improvement against a known number rather than a guessed one.
Patient-level grouping in cross-validation is shown to be practical on a 668-image dataset; the same discipline should be expected in larger medical imaging studies.
The modest separation (AUC 0.755 vs 0.500, accuracy 75.4% vs 71.6%) implies that handcrafted features alone do not solve the task, strengthening the case for richer learned representations.
The paper's scoped multi-modal CNN-transformer proposal is explicitly future work; its success hinges on access to larger, multi-institutional, ethically governed datasets and external validation.

Reading between the lines

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

A testable extension not run in the paper: hold out entire source institutions or case-report batches rather than only patients; if AUC collapses toward 0.5, much of the reported signal is source artifact, not disease pattern.
Pooling viral, bacterial, and fungal pneumonias into one 'other pneumonia' class may blur the decision boundary; splitting by etiology, where labels allow, could change the measured separability in either direction.
The proposed CNN-transformer fusion is only sketched; the empirical payload of the paper is the baseline itself, so the architecture should be treated as a hypothesis generator, not a result.

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.

Desk Editor's Note

read the letter

Dear colleague,

This is a small, honest pilot that delivers a reproducible baseline for COVID-19 versus other pneumonia on the Cohen CXR dataset. Nothing here is new methodologically — HOG/GLCM plus classical classifiers has been done on this dataset many times — but the execution is careful: patient-level 5-fold CV (408 patients, 668 images), class-balanced weighting, a majority-class baseline, and a clear limitations section. The code is public.

The best result is SVM at 75.4% accuracy and 0.755 AUC, versus 71.6% accuracy and 0.5 AUC for the majority baseline. That is a modest effect, and the paper says so. It also explicitly labels the proposed CNN-transformer architecture as future work and does not let it contaminate the results.

The soft spots. First, no confidence intervals or significance tests are reported. The 3.8-point accuracy margin over the majority baseline could easily be noise; the AUC gap is larger but still needs a variance estimate. Second, the stress-test concern about acquisition artifacts is on target. The Cohen dataset aggregates case reports from many institutions, scanners, and views. Patient-level grouping prevents the same patient leaking across folds, but it does not control for source-institution or view artifacts that HOG/GLCM can encode. The authors acknowledge heterogeneity but never test for this confound, so claiming the features capture disease-specific signal is under-supported. Third, the pneumonia class pools viral, bacterial, and fungal cases, which further blurs any real signal.

None of this kills the paper. It is a pilot, and the authors are appropriately cautious — they explicitly say the results are not clinically meaningful. The central argument, that this provides a transparent baseline for later work, holds. The stronger claim about texture features is fragile.

For whom? Someone working on this dataset or wanting an example of correct patient-level CV in a small study. Not a must-read, but a decent reproducibility exercise. I'd send it to peer review if the venue wants to encourage thorough, honest pilot studies; otherwise it's a fine preprint. If you cite it, cite it as a baseline, not as evidence about COVID-19 imaging.

Referee Report

3 major / 4 minor

Summary. The paper reports a pilot machine-learning study distinguishing COVID-19 from other pneumonia on chest X-rays using classical texture and gradient features (HOG, GLCM, basic intensity statistics) with logistic regression, random forest, and RBF SVM. Experiments use 668 PA/AP radiographs from 408 patients in the Cohen et al. COVID-19 Image Data Collection, with patient-level stratified 5-fold cross-validation. The best classifier, SVM, achieves mean accuracy 75.4% and AUC 0.755, compared with a 71.6% majority-class baseline and 0.500 chance AUC. The authors interpret this as evidence that handcrafted features carry modest discriminative signal, and on this basis outline a future multi-modal convolutional-transformer deep-learning architecture.

Significance. The manuscript has real strengths: it uses a fully public dataset and provides code for reproducibility; patient-level grouping prevents the most common form of data leakage; a majority-class baseline is reported; and the limitations section is candid about dataset size, heterogeneity, and lack of external validation. If the reported signal were robust and attributable to disease pathology, the paper would provide a useful reproducible baseline for a much-studied task. However, the central empirical claim rests on a small, unquantified accuracy gain over a trivial baseline, and the validation design does not rule out source-institution or acquisition-protocol artifacts. The proposed deep-learning architecture is explicitly future work and is not a contribution evaluated here. The significance is therefore conditional on addressing the artifact and uncertainty concerns.

major comments (3)

[Section V, Table I; Section VI.A] The central claim that the SVM 'modestly exceeded' the majority baseline and that handcrafted features carry discriminative signal is not supported by uncertainty quantification. Only mean accuracy/AUC across folds are reported; no fold-wise standard deviations, confidence intervals, or significance tests are given. With 668 images and a 3.8-percentage-point accuracy difference, the result may be within cross-validation noise. Please report per-fold variability and perform a statistical comparison against the majority baseline (e.g., a paired test across folds or permutation test), and state whether the observed difference is statistically reliable.
[Section III; Section IV.C; Section VI.A] Patient-level grouping does not control for the strong association between label and source institution or acquisition protocol in the Cohen et al. collection, which aggregates images from diverse published case reports. COVID-19 and other-pneumonia images may come from different hospitals, scanners, or PA/AP view distributions; HOG and GLCM features are sensitive to resolution, contrast normalization, noise, and positioning geometry. The classifier may therefore be learning source artifacts rather than disease pathology. Please add a confound analysis: for example, report performance stratified by source institution or view, train/test on images from the same holding-out sources, include source/view as a covariate, or at minimum show the distribution of sources and views across the two classes. Without such a check, the interpretation in Section VI.A is not established.
[Section V, Table I] The random forest recall of 0.985 with precision of 0.743 indicates that the model is nearly always predicting the majority class, despite class-balanced weighting. The SVM's higher accuracy (0.754) is only slightly above the 0.716 majority baseline and may reflect decision-threshold or class-weight choices rather than a robust learned boundary. Please report precision-recall curves or threshold-independent metrics, and clarify whether the reported accuracy corresponds to the default 0.5 threshold or to an optimized threshold; if optimized, this should be incorporated into the uncertainty analysis.

minor comments (4)

[Section IV.B] The feature-vector dimensionality is stated as 1,792, but the contribution of each feature family is not broken down. A short derivation or table would improve reproducibility and help readers assess whether the SVM is operating in a very high-dimensional space relative to the sample size.
[Section V, Figures 1 and 2] The figures are referenced but are not included in the manuscript text. Please ensure they are embedded or clearly available; otherwise the pooled confusion matrix and ROC curve cannot be inspected.
[Section VI.C] The proposed multi-modal architecture is described only in broad terms. Since this is explicitly future work, a schematic or a more concrete module list would strengthen the motivation, but this is not required for the empirical claims.
[Throughout] There are minor typographical and formatting inconsistencies, such as inconsistent spacing after equations and the use of 'posteroanterior/anteroposterior' without defining PA and AP at first use. A careful proofread is recommended.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical benchmark with patient-level cross-validation and no fitted-theory/prediction loop.

full rationale

The paper reports a straightforward empirical evaluation: handcrafted HOG/GLCM features are extracted from a public dataset, three classical classifiers are trained, and performance is estimated by patient-level stratified 5-fold cross-validation. There is no derivation chain, no fitted parameter that is later 'predicted,' no uniqueness theorem, and no theory whose output is defined in terms of its inputs. The 'best mean accuracy of 75.4% and AUC of 0.755' are out-of-sample cross-validation estimates directly reported as such, with a majority-class baseline and explicitly stated limitations. The proposed multi-modal deep learning architecture is labeled future work and is not used to generate the reported numbers, so it cannot be circular. The skeptic's concern — that source-institution or view artifacts might explain part of the signal — is an external-validity threat, not a circularity: the reported result is what it is, and the paper itself acknowledges the lack of external validation and the heterogeneity of the Cohen et al. dataset. No self-citation is load-bearing; the only dataset citation is to the external, publicly available COVID-19 Image Data Collection, and no prior work by the author is invoked to justify the central claim. Therefore the appropriate circularity score is 0.

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

The paper introduces no new physical or model entities. The multi-modal deep-learning architecture is a clearly-labeled future proposal, not an entity used in the experiments. The central claim depends on the standard ML assumptions about data labeling, patient grouping, and representativeness of the public dataset.

free parameters (4)

HOG cell size (16×16 px) = 16×16
Hand-chosen; not optimized. Affects gradient descriptor granularity.
GLCM distances and angles = two distances, three angles (unspecified)
Hand-chosen; not numerically specified, hindering exact replication.
SVM RBF hyperparameters (C, gamma) = unspecified
Not reported; the exact decision boundary cannot be reproduced.
Random forest tree count = 300
Hand-chosen; not optimized.

assumptions (4)

domain assumption Ground-truth labels in the Cohen et al. dataset are clinically correct.
Section III relies on the dataset's 'finding' labels to define COVID-19 vs. other pneumonia; mislabeling would bias results.
domain assumption Patient ID field correctly groups all images per patient.
Section IV uses patient ID for group CV; duplicate or misassigned IDs would leak information.
domain assumption The selected PA/AP images are a representative subsample for the task.
Section III filters from a case-report collection; the subset may carry source-specific artifacts.
domain assumption Standard preprocessing preserves diagnostically relevant texture information.
Section IV-A resizes to 128×128 and applies histogram equalization; this may discard details central to pneumonia patterns.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Toward Multi-Modal Deep Learning for Pulmonary Disease Classification: A Texture-Based Machine Learning Pilot Study on Public Chest X-Ray Data." pith.science (2026). https://pith.science/paper/FM7PVE3U

@misc{pith2026260727286,
  author       = {Pith},
  title        = {Pith review of: Toward Multi-Modal Deep Learning for Pulmonary Disease Classification: A Texture-Based Machine Learning Pilot Study on Public Chest X-Ray Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FM7PVE3U}},
  note         = {Machine review of arXiv:2607.27286}
}

read the original abstract

Automated classification of pulmonary disease from chest radiographs is a widely studied application of machine learning in medical imaging. This paper presents a pilot study evaluating classical texture- and gradient-based feature representations for distinguishing COVID-19 from other forms of pneumonia using the publicly available COVID-19 Image Data Collection (668 posteroanterior/anteroposterior radiographs from 408 patients). Using histogram of oriented gradients (HOG) and gray-level co-occurrence matrix (GLCM) texture descriptors with classical classifiers (logistic regression, random forest, and support vector machine), evaluated under patient-level 5-fold stratified cross-validation to prevent data leakage, we obtain a best mean accuracy of 75.4% and AUC of 0.755, modestly exceeding the 71.6% majority-class baseline. We report these results transparently, including their limitations, and use them to motivate and scope a proposed multi-modal deep learning architecture -- combining convolutional and transformer-based encoders across imaging modalities -- as a direction for future work requiring access to larger, multi-institutional, ethically sourced datasets.

Figures

Figures reproduced from arXiv: 2607.27286 by the authors.

**Figure 2.** Pooled ROC curve, SVM classifier, 5-fold CV (AUC = 0.755). [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

13 extracted references · 3 linked inside Pith

[1]

Deep learning applications in chest imaging,

G. Chassagnon, M. Vakalopoulou, N. Paragios, and M. P. Revel, “Deep learning applications in chest imaging,” Eur. J. Radiol., vol. 141, p. 109975, 2023

2023
[2]

ChestX-ray8: Hospital-scale chest X-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases,

X. Wang, Y. Peng, L. Lu, Z. Lu, M. Bagheri, and R. M. Summers, “ChestX-ray8: Hospital-scale chest X-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2017, pp. 2097–2106

2017
[3]

End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography,

D. Ardila et al., “End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography,” Nat. Med., vol. 25, no. 6, pp. 954–961, 2019

2019
[4]

Deep learning-enabled medical computer vision,

A. Esteva et al., “Deep learning-enabled medical computer vision,” NPJ Digit. Med., vol. 4, no. 1, p. 5, 2021

2021
[5]

TransUNet: Transformers make strong encoders for medical image segmentation,

J. Chen et al., “TransUNet: Transformers make strong encoders for medical image segmentation,” arXiv:2102.04306, 2021

arXiv 2021
[6]

Rethinking medical image classification: Computer vision foundation models and domain-specific fine-tuning,

S. Matsoukas, A. B. Hernandez, and Y. Liu, “Rethinking medical image classification: Computer vision foundation models and domain-specific fine-tuning,” J. Digit. Imaging, vol. 35, pp. 1175–1192, 2022

2022
[7]

Contrastive learning of medical visual representations from paired images and text,

Y. Zhang, H. Jiang, Y. Miura, C. D. Manning, and C. P. Langlotz, “Contrastive learning of medical visual representations from paired images and text,” arXiv:2010.00747, 2020

arXiv 2010
[8]

PENet—a scalable deep-learning model for automated diagnosis of pulmonary embolism using volumetric CT imaging,

S. C. Huang et al., “PENet—a scalable deep-learning model for automated diagnosis of pulmonary embolism using volumetric CT imaging,” NPJ Digit. Med., vol. 3, no. 1, p. 61, 2020

2020

Show all 13 references

[9]

Swin transformer: Hierarchical vision transformer using shifted windows,

Z. Liu et al., “Swin transformer: Hierarchical vision transformer using shifted windows,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2021, pp. 10012–10022

2021
[10]

EfficientNetV2: Smaller models and faster training,

M. Tan and Q. Le, “EfficientNetV2: Smaller models and faster training,” in Proc. Int. Conf. Mach. Learn., 2021, pp. 10096–10106

2021
[11]

COVID-19 image data collection,

J. P. Cohen, P. Morrison, and L. Dao, “COVID-19 image data collection,” arXiv:2003.11597, 2020

2003 arXiv
[12]

Textural features for image classification,

R. M. Haralick, K. Shanmugam, and I. Dinstein, “Textural features for image classification,” IEEE Trans. Syst., Man, Cybern., vol. SMC-3, no. 6, pp. 610–621, 1973

1973
[13]

Histograms of oriented gradients for human detection,

N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2005, pp. 886–893

2005

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Deep learning applications in chest imaging,

G. Chassagnon, M. Vakalopoulou, N. Paragios, and M. P. Revel, “Deep learning applications in chest imaging,” Eur. J. Radiol., vol. 141, p. 109975, 2023

2023

[2] [2]

ChestX-ray8: Hospital-scale chest X-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases,

X. Wang, Y. Peng, L. Lu, Z. Lu, M. Bagheri, and R. M. Summers, “ChestX-ray8: Hospital-scale chest X-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2017, pp. 2097–2106

2017

[3] [3]

End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography,

D. Ardila et al., “End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography,” Nat. Med., vol. 25, no. 6, pp. 954–961, 2019

2019

[4] [4]

Deep learning-enabled medical computer vision,

A. Esteva et al., “Deep learning-enabled medical computer vision,” NPJ Digit. Med., vol. 4, no. 1, p. 5, 2021

2021

[5] [5]

TransUNet: Transformers make strong encoders for medical image segmentation,

J. Chen et al., “TransUNet: Transformers make strong encoders for medical image segmentation,” arXiv:2102.04306, 2021

arXiv 2021

[6] [6]

Rethinking medical image classification: Computer vision foundation models and domain-specific fine-tuning,

S. Matsoukas, A. B. Hernandez, and Y. Liu, “Rethinking medical image classification: Computer vision foundation models and domain-specific fine-tuning,” J. Digit. Imaging, vol. 35, pp. 1175–1192, 2022

2022

[7] [7]

Contrastive learning of medical visual representations from paired images and text,

Y. Zhang, H. Jiang, Y. Miura, C. D. Manning, and C. P. Langlotz, “Contrastive learning of medical visual representations from paired images and text,” arXiv:2010.00747, 2020

arXiv 2010

[8] [8]

PENet—a scalable deep-learning model for automated diagnosis of pulmonary embolism using volumetric CT imaging,

S. C. Huang et al., “PENet—a scalable deep-learning model for automated diagnosis of pulmonary embolism using volumetric CT imaging,” NPJ Digit. Med., vol. 3, no. 1, p. 61, 2020

2020