OCT-CAD: An Automated Deep Learning Diagnostic Tool for Eye Diseases from Imbalanced OCT Scans

This work addresses the challenge of class imbalance in Optical Coherence Tomography (OCT). We propose OCT-CAD framework, a two-stream Convolutional Neural Network (CNN) for eye diseases classification. The proposed framework consists of two complimentary CNNs. One CNN is focused on classifying tail classes, and the other is focused on head classes. In the proposed framework, we utilize Key Point Sensitive Loss (LKPs) and Cross-Entropy Loss (LCE) to train the proposed framework. Then, the features extracted from the two CNNs are processed using PCA, DCT, and feature selection techniques to enhance the representation. The experimental results show that a stacking classifier that combined random forest and gradient boost achieves 100% accuracy on the OCTID dataset, demon-strating the effectiveness of the proposed OCT-CAD framework in addressing class imbalance, improving diagnostic performance and surpassing existing state-of-the-art techniques.