Training Configuration Effects on Deep Learning Generalization for Multi-Class Retinal Pathology Classification

Deep learning models for fundoscopic image classification demonstrate variable performance depending on preprocessing strategies, architectural modifications, and training configurations. Understanding how these factors influence internal validation, cross-dataset generalization, and cross-pathology detection remains critical for developing robust diagnostic systems. Three modified ResNet-18 architectures were trained on 4,217 fundoscopic images (normal, cataract, glaucoma, and diabetic retinopathy) with stratified partitioning. Model 1 employed class-specific preprocessing, no class weighting, stochastic gradient descent with momentum optimization, and 1,260 iterations. Model 2 implemented unified enhanced preprocessing, capped class weighting, 30-layer freezing, Adam optimization with learning rate decay, and 2,520 iterations. Model 3 utilized unified preprocessing with stochastic augmentation, unlimited class weighting, 30-layer freezing, and 2,520 iterations. Performance was evaluated through internal validation, external validation on 700 independent images, and binary classification on 2,190 images representing three unseen conditions. Internal test accuracy reached 91.94%, 86.49%, and 69.67% for Models 1, 2, and 3, respectively, with Model 2 demonstrating optimal stability. External validation revealed substantial heterogeneity: cataract classification achieved 91–98% accuracy, diabetic retinopathy 99–100%, glaucoma 1.5–43%, and normal images 0–100%. Binary classification on three unseen retinal conditions showed variable performance, with Models 1 and 2 achieving 66–93% accuracy and Model 3 achieving 100% accuracy across all categories. All models demonstrated 100% precision for abnormality detection. Training configurations substantially influence model behavior across evaluation scenarios. Comprehensive protocols incorporating internal validation, cross-dataset testing, and cross-pathology assessment are necessary to characterize diagnostic system performance and identify failure modes not apparent from standard train-test splits.