OcuMDNet: A lightweight CNN for robust multi-disease retinal diagnosis with cross-dataset reliability

Abstract

This study aims to develop a lightweight convolutional network for the classification of multiple retinal diseases using fundus images and to evaluate cross-dataset generalization under strict label alignment. We introduce OcuMDNet (Ocular Multi-Disease Net), a compact convolutional neural network (CNN) specifically designed for fundus imagery, incorporating batch normalization and dropout for regularization. A standardized processing pipeline is employed, which includes cropping and resizing images to 224 × 224 pixels, applying contrast-limited adaptive histogram equalization (CLAHE), and performing per-channel normalization. The training process utilizes the AdamW optimizer and incorporates early stopping to enhance model performance. We propose a label-aligned evaluation protocol: (i) assesses 4-class performance (Normal, diabetic retinopathy (DR), Glaucoma, age-related macular degeneration (AMD)) on a combined dataset assembled from public sources; (ii) reports disease-specific results based on the native labels of each dataset (DR: EyePACS, Messidor; Glaucoma: ORIGA; AMD: AREDS); and (iii) evaluates cross-dataset transfer for DR (training on EyePACS and testing on Messidor). Patient-level splits are implemented to prevent data leakage, and class counts are reported for each split. Performance metrics such as accuracy, macro-F1 score, and one-vs-rest ROC-AUC are calculated with 95 % confidence intervals using stratified bootstrap (n = 1000). Paired comparisons are conducted using McNemar's test for accuracy and DeLong's method for AUC, with multiplicity control applied. The OcuMDNet demonstrates strong performance on both combined and disease-specific benchmarks, maintaining robust discrimination in cross-dataset evaluations for DR while ensuring computational efficiency suitable for large-scale screening applications. Ablation studies confirm the significance of preprocessing steps and architectural choices. In conjunction with a label-aligned protocol, the OcuMDNet provides an accurate and efficient baseline for multi-disease fundus analysis, facilitating a transparent assessment of cross-dataset reliability. The code, scripts, and trained weights will be made available to support reproducibility.