Interpretable Deep Learning For Automated Retinal Damage Detection in Optical Coherence Tomography

This work portrays developing and evaluating deep learning models for automatically identifying retinal injuries using Optical Coherence Tomography (OCT) images. In particular, this study optimizes three pre-trained Convolutional Neural Networks (CNNs), VGG16, EfficientNetB3, and ResNet50, using transfer learning to classify retinal images into four categories, Drusen, Normal, Diabetic Macular Edema (DME), and Choroidal Neovascularization (CNV). In addition to high-performance classification, the study integrates Explainable Artificial Intelligence (XAI) methods, utilizing Gradient-weighted Class Activation Mapping (Grad-CAM) to highlight the retinal regions influencing the model’s predictions visually. This offers critical interpretability for clinical settings. The Key performance metrics indicators include accuracy, precision, and F1-score and help to evaluate the models. With the available selected models, EfficientNetB3 achieved the highest performance with an accuracy of 93%, outperforming VGG16 and ResNet50. The application of Grad-CAM highlights the increase in transparency and prediction trust of this model's decision-making process. This study demonstrated the profound learning model efficacy, particularly EfficientNetB3, for accurate retinal injury detection and provides valuable insights into model interpretability for clinical use.