Eyes wide open: a deep learning comparison of CNN models for retinal disease detection from fundus images

Abstract

Fluid interpretation of fundus photographs is impeded by class-specific visual heterogeneity coupled with clinical class imbalance. Subtle retinal pathology combined with marked class imbalance makes differential diagnosis challenging. Trained class classification is complicated further when clinically transferred neural models are evaluated using conventional metrics. We focus on a comparative analysis of five distinct convolutional neural network models classifying fundus photographs to single retinal diseases: Xception, ResNeXt-101, DenseNet-201, SE-ResNet, and NASNet-Mobile. We restricted our analysis to a fundus image dataset composed of nine clinically validated disease categories. Non-fundus and poor-quality images were manually excluded. To reduce the disparate class imbalance, we employed targeted photometric (zoom, contrast enhancement) and geometric (rotation, flipping) augmentation. Quantitative results show Xception to have the highest score with an accuracy of 94.07%, a precision of 94.11%, a recall of 94.07%, and an F1-score of 94.06%. DenseNet-201 and ResNeXt-101 followed closely with F1-scores of 93.21% and 91.02%, respectively. While SE-ResNet and NASNet-Mobile had overall lower classification accuracy, they showcased better balance. Parameter count, FLOPs, memory footprint, and inference latency were examined for further evaluation and practical viability. The Xception model seemed to have a good balance between predictive performance and efficiency, making it an appropriate model with limited resources, though ResNeXt-101 showed a higher computational demand. Additionally, an independent dataset of retinal images with three overlapping disease categories (n= 3,179) was used for external validation. With an accuracy of 92.03% and a macro-F1 score of 91.79%, Xception demonstrated a strong ability to generalize, demonstrating robustness under inter-dataset variability.