A Harmonized Multi-Source Dataset with Baseline Deep Learning Validation for Staging Diabetic Retinopathy

Accurate automated grading of diabetic retinopathy (DR) significantly depends on the quality of retinal fundus images. Inferior-quality pictures, resulting from inadequate lighting, motion blur, distortions, or incomplete retinal coverage, may obscure minor lesions and diminish the accuracy of model predictions. This study constructs a harmonized multi-source dataset using a multi-dimensional image quality assessment framework for multi-class DR staging. Retinal images are collected from IDRiD, Messidor-2, SUSTech-SYSU, APTOS 2019, DeepDRiD-v1.1, and Zenodo DR V03 datasets. The proposed pipeline includes preprocessing, image quality assessment using technical quality and medical relevance indicators, dataset-specific statistics, and adaptively thresholded using DR severity-aware percentiles derived from stratified samples with weighting to match diagnostic needs. Baseline deep learning models were trained for three hierarchical DR classification schemes to validate the dataset. Experimental results show that the quality-filtered merging of datasets improves model generalization accuracy by 3-7% compared to the normal merging of datasets. This work provides a benchmark dataset and baseline performance results to facilitate future research in DR staging and medical image classification.