REVIEW 5 major objections 4 minor 6 references
Diagnosis of diabetic retinopathy using machine learning & deep learning technique
T0 review · 5 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read A YOLOv8 detector that counts retinal lesions, followed by an SVM classifier, grades diabetic retinopathy severity with 84% accuracy.
desk verdict Sincere student project, not a research contribution: the 84% accuracy is unsupported and the pipeline is a standard combination of existing tools. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central mechanism is the lesion-count feature vector. YOLOv8, a real-time object-detection network trained on manually drawn bounding boxes around lesions, converts each fundus image into counts of lesion types. Those counts are then scaled and fed to an SVM, whose separating hyperplane assigns one of the five DR severity grades. The machinery transforms high-dimensional image data into a compact, interpretable numerical summary that a classical classifier can separate.
What would settle it
Run the trained YOLO+SVM pipeline on an independent, expertly labeled fundus-image test set with verified severity grades and compare predicted grades to ground truth; if accuracy falls substantially below 84%, or if the detector's lesion counts do not correlate with severity, the reported performance was an artifact of the training annotations or data split.
Extended reading notes
Core claim
The central claim is that fundus images can be reduced to a small numerical summary—counts of each lesion type—which is enough for a classical classifier to assign diabetic retinopathy severity. In the authors' own words, YOLOv8 locates regions of interest such as the optic disc, optic cup, and lesions, and then SVM classifies these into DR stages based on the presence or absence of pathological signs. Their reported numbers are a training score of 91%, testing score of 84%, accuracy 84%, F1 81%, and precision 82% for the SVM stage, with YOLO achieving 78% accuracy, 74% F1, and 72% precision. The authors conclude that this method can be applied for retinal fundus disease triage, especially in remote areas.
Load-bearing premise
The pipeline's performance rests on the assumption that the manually drawn lesion boxes used to train the detector are accurate and that the public dataset's severity labels are correct; the paper itself states that the dataset contains many poor-quality, incorrectly labelled images, and no cleaning or annotation-validation step is described.
Editorial extensions
If this is right
- If the 84% accuracy holds on independently cleaned data, the pipeline could serve as a first-pass triage tool in primary-care and telemedicine settings, flagging images that need specialist review.
- The two-stage design means large end-to-end severity-labeled datasets are not required; only lesion bounding boxes and count-based labels are needed for training.
- The SVM's decisions are tied to named lesion types, so a clinician could see which pathological signs drove the predicted grade.
- Because each image is reduced to a short vector of counts, the method is computationally light enough to run on modest hardware, which matters for remote clinics.
Reading between the lines
- The reported 84% accuracy comes from a test split of the same public dataset used for training, which the paper admits contains poor-quality and incorrectly labelled images; real-world generalization to expertly graded fundus images is likely lower.
- A natural extension is to test the same lesion-count vector with other classifiers, such as random forests or gradient boosting, which might improve the grading stage without retraining the detector.
- The detector's 78% lesion accuracy implies some lesions are missed; severity grades relying on lesion counts may be less reliable for mild DR, where a single missed microaneurysm can change the grade.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript describes a two-stage pipeline for grading diabetic retinopathy (DR) from fundus images: a custom YOLOv8 object detector is trained on bounding-box annotations (made with makesense.ai) to locate lesions, the detected lesions are counted per image to form a numeric feature vector, and a support vector machine (SVM) is trained on those features to predict DR severity on a five-class scale. The authors report 84% test accuracy, 82% precision, and 81% F1 for the SVM, and 78% accuracy, 72% precision, and 74% F1 for the YOLO detector. The paper includes a GUI for image input and prediction, but the experimental description omits dataset size, class distribution, training hyperparameters, validation protocol, leakage prevention, and comparisons with prior work. The conclusion cautiously notes that clinical testing has not yet been performed.
Significance. If the 84% accuracy figure were rigorously established, the proposed integration of YOLOv8 lesion counting with SVM severity classification could be a useful low-cost triage tool, particularly for remote settings. The high-level pipeline is reasonable: lesion types and counts are clinically meaningful features for DR grading, and the GUI demonstrates a practical deployment path. However, the paper contributes no reproducible code, no dataset, no experimental protocol, no statistical uncertainty quantification, and no comparison against existing DR grading methods. The claimed novelty is not substantiated; the contribution is an integration of existing components whose performance is currently supported only by an undocumented empirical number. As a result, the significance is contingent on experimental details that are not present in the manuscript.
major comments (5)
- [Dataset Description] The dataset is not described with the numbers needed to interpret the result: no total image count, no number of subjects, no per-class distribution across the five DR grades, and no train/validation/test partition sizes are given. The manuscript itself states that 'Kaggle contains many images with poor quality and incorrect labelling,' yet no cleaning, relabeling, exclusion, or quality-control protocol is described. Without this information, the reported 84% test accuracy cannot be interpreted as a statement about generalization to unseen fundus images.
- [Training Work-Flow] The number of images annotated with makesense.ai and used to train YOLOv8 is not stated, nor are the YOLOv8 hyperparameters, number of epochs, input resolution, augmentation scheme, or the set of lesion classes annotated. The procedure then applies the custom detector to 'the whole diabetic retinopathy image dataset' to generate the feature matrix, and the SVM is trained on an 80/20 split of that matrix. If any image used for SVM testing was also seen by YOLO during training or validation, the derived lesion counts could reflect information leakage rather than independent generalization; no mechanism to prevent such overlap is described.
- [Result] The Results section reports only four aggregate scalar metrics for the SVM (91% training accuracy, 84% testing accuracy, 81% F1, 82% precision) and three for YOLO (78% accuracy, 74% F1, 72% precision). No confusion matrix, per-class precision/recall, confidence intervals, cross-validation, or external test set is reported. DR grading is typically class-imbalanced, so aggregate accuracy alone can be misleading; the central claim that the model 'can correctly predict the grade of diabetic retinopathy up to 84% of the testing images' requires per-class evidence and an explicit statement of how the test set was constructed.
- [Method And Materials / Proposed System] The paper repeatedly calls the method 'novel,' but it provides no comparison with any baseline or prior DR grading system, and the literature survey does not report baseline accuracies from the cited papers. A novelty claim without comparative evaluation is an overclaim; the manuscript needs at least a comparison with a simple classifier on the same features, a standard CNN classifier, and published benchmark results on the same dataset.
- [Training Work-Flow] The SVM description is not reproducible: the text says 'we train our SVM model on the training data by finding the optimal hyperparameters,' but it never specifies the kernel type, regularization parameter C, gamma, feature scaling method, feature selection criterion, or dimensionality reduction technique. Similarly, the claimed YOLO metrics are given without the evaluation protocol used to obtain them. These omissions are load-bearing because the 84% accuracy figure cannot be verified or reproduced.
minor comments (4)
- [Conclusion] The sentence 'We have test it on our dataset taken from online website yet the clinical test is to be performed' is duplicated verbatim in the conclusion and should be corrected and reduced to one statement.
- [Literature Survey] The literature survey contains several incomplete or garbled sentences, for example 'r results, the problem was that it was quite slow' and the reference citations occasionally run into the text; the section needs careful editing and full bibliographic entries for all cited works.
- [Throughout] There are minor language and typographical errors, including 'upto 84%' in the Results, 'lessions' in the literature survey, 'loose' for 'lose' in the lesion descriptions, and inconsistent figure references (Figure 10 appears before its caption context). A thorough proofread is needed.
- [Dataset Description] The abstract and introduction state the method can be applied 'for retinal fundus disease triage, especially in remote areas,' but no runtime, hardware, or deployment analysis is provided; this claim should either be supported or softened.
Circularity Check
No substantive circularity: the 84% accuracy is an empirical result from a supervised split, not a derivation that reduces to its inputs; the main risk is unreported data leakage, which is a validity concern rather than circularity.
full rationale
The paper's central claim is an SVM testing accuracy (84%) obtained from an 80/20 train/test split of a numerical dataset whose features are lesion counts produced by a custom YOLOv8 detector. This is an empirical measurement, not a self-referential derivation: the SVM is not defined in terms of the reported accuracy, and no fitted parameter is renamed as a prediction. There are no load-bearing self-citations, no imported uniqueness theorems, and no ansatz smuggled in via citation. The only nearby concern is methodological leakage: the text says 'we apply object detection on the whole diabetic retinopathy image dataset' after YOLO training, which could mean SVM test images had their features generated by a detector that may have seen those same images during annotation-based training. However, the paper does not specify the overlap, so one cannot exhibit a concrete reduction of the 84% figure to a training input. Under the hard rule requiring quotation and explicit reduction, this is a reproducibility/validity weakness, not demonstrable circularity. The 'novel method' phrasing is an overclaim, not circular reasoning. Hence the appropriate score is 1, reflecting a minor concern about the self-containedness of the test procedure rather than any definitional circularity.
Assumptions & free parameters
free parameters (2)
- SVM regularization and kernel hyperparameters =
not reported
- YOLOv8 training hyperparameters =
not reported
assumptions (2)
- domain assumption Lesion type and count are sufficient features to grade DR severity.
- domain assumption Kaggle fundus images, despite acknowledged poor quality and incorrect labels, are adequate for training and testing the models.
Cite this review
Pith. "Pith review of Diagnosis of diabetic retinopathy using machine learning & deep learning technique." pith.science (2026). https://pith.science/paper/HD5HJBGS
@misc{pith2026241116250,
author = {Pith},
title = {Pith review of: Diagnosis of diabetic retinopathy using machine learning & deep learning technique},
year = {2026},
howpublished = {\url{https://pith.science/paper/HD5HJBGS}},
note = {Machine review of arXiv:2411.16250}
}
read the original abstract
Fundus images are widely used for diagnosing various eye diseases, such as diabetic retinopathy, glaucoma, and age-related macular degeneration. However, manual analysis of fundus images is time-consuming and prone to errors. In this report, we propose a novel method for fundus detection using object detection and machine learning classification techniques. We use a YOLO_V8 to perform object detection on fundus images and locate the regions of interest (ROIs) such as optic disc, optic cup and lesions. We then use machine learning SVM classification algorithms to classify the ROIs into different DR stages based on the presence or absence of pathological signs such as exudates, microaneurysms, and haemorrhages etc. Our method achieves 84% accuracy and efficiency for fundus detection and can be applied for retinal fundus disease triage, especially in remote areas around the world.
Figures
Figures from the paper (3 more)
Reference graph
Works this paper leans on
-
[5]
“Multiclass SVM -Based Automated Diagnosis of Diabetic Retinopathy,
A. P. a. D. Jeyakumari, ““Multiclass SVM -Based Automated Diagnosis of Diabetic Retinopathy,” in Int. Conf. on Comm.& Signal Proc., India, April 3 to 5 2013
work page 2013
-
[1]
L. W. S. S. S. Qingqing Xu1, “A systematic literature review of predicting diabetic retinopathy, nephropathy and neuropathy in patients with type 1 diabetes using machine learning”
-
[2]
K. S. Mann and S. Kaur, “Segmentation of retinal blood vessels using artificial neural networks for early detection of diabetic retinopathy,” 27 january 2017. [Online]. Available: https://pubs.aip.org/aip/acp/article/1836/1/020026/586205/Segmentation- of-retinal-blood-vessels- using
work page 2017
-
[3]
L. B. S. R. GeethaRamani, “Retinal blood vessel segmentation employing image processing and data mining techniques for computerized retinal image analysis. Retinal blood vessel segmentation in fundus images,” Jour. of Biocybernetics & Biomed.Engg, 2016
work page 2016
-
[4]
J. G. N. L. H. L.-A. A. S. E. M. Larsen, “Automated detection of,” Investigative Ophthalmology &d Vision Sci, vol. 44, 2003
work page 2003
-
[6]
Model based method for retinal blood vessel detection,
F. V. H. L. A. V. K.A. Vermeer, “Model based method for retinal blood vessel detection,” Comps. in Bio.& Med, vol. 34, p. No.3, 2004
work page 2004
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.