Neutrosophic-inspired data augmentation and multi-architecture ensemble model for robust retinal disease classification

Abstract Classification of retinal diseases using OCT imaging is a problem of data imbalance and inter- and intra-observer variability, as well as generalization of deep learning methods in a clinical environment. Indeed, the present study proposes an innovative data augmentation approach merged with a multi-architecture ensemble model to increase the diagnostic performance and robustness of the system. The augmentation idea of the present work is based on Neutrosophic logic preserving clinical uncertainty by incorporating such operations as noise, blur, rotation, and contrast change. To extract diverse features, ConvNeXt-Tiny, DenseNet-121, and Swin Transformer-Tiny are used in parallel, and the features are fused at the representation level. On the OCT2017 dataset, the model’s accuracy reaches 96%; the classifiers have high class-specific F1-score: 0.97 for CNV and NORMAL, 0.95 for DME, and 0.86 for DRUSEN, which indicates the model’s high generality on the recognition of four classes of retinal diseases. Note that our study has placed our algorithm in conjunction with deep learning and traditional machine learning standards and reveals that our algorithm grossly performs higher than all earlier demonstrated algorithms in both aspects of precision and robustness. Moreover, the proposed ensemble design satisfies the requirement of its application in the clinical environment by optimizing the computational complexity and diagnostic performance. A balance adjustment was applied to the OCT2017 dataset through image selection until the total amounted to 5000 images distributed equally by class. When evaluated on this balanced dataset, the proposed model produced both 97.8% accuracy and 0.98 macro F1-score demonstrating superior and steady performance for every retinal disease category. They affirm the benefit of integrating the uncertainty-aware augmentation with the use of a diversified ensemble for the classification of the retinal diseases by OCT images where there are difficult or grey areas in the images.