A comparative study of machine learning and deep learning models for predicting diabetic retinopathy severity in retinal fundus image datasets

Diabetic retinopathy (DR) remains a prominent cause of visionimpairment, necessitating timely and accurate grading of disease severity from retinal fundus images for effective screening and referral. Early detection and precise classification of DR severity are essential to mitigate the risk of vision loss. This study evaluates the performance of eight machine learning (ML) and five deep learning (DL) techniques using two publicly accessible datasets to classify DR into five severity levels. A standardized workflow is used for image quality assessment, normalization, and preprocessing, enhancing the detection of vascular structures while employing a hybrid approach to mitigate class imbalance. Key performance metrics, including prediction accuracy, scalability, stratified validation, and robustness to domain shifts, are comprehensively assessed. Using a smaller dataset of 495 retinal fundus images, traditional ML models had limited success in fine-grained severity classification, with XGBoost yielding the best accuracy. DL models faced similar challenges due to limited and imbalanced data. However, when tested on a larger dataset of 3, 662 images, both ML and DL models showed significant performance improvements. XGBoost and LR excelled at low-granularity tasks, while EfficientNetB0 achieved the highest accuracy across severity levels. This underscores that model performance significantly improves with larger, balanced datasets for DR prediction. This research emphasizes the challenges in DR severity prediction, particularly due to class imbalance and limitations of current datasets. It underscores the pressing need for larger, more balanced datasets to enhance the robustness and effectiveness of DL models for multi-class DR severity prediction. This work closely aligns with the research objective of providing a comparative analysis of both ML and DL approaches tailored explicitly for multi-class diabetic retinopathy severity prediction, rather than a broad evaluation of algorithms.