Automated Classification of Ophthalmic Disorders Using Color Fundus Images

Abstract

This study proposes a novel methodology for classifying ocular diseases using convolutional neural networks (CNNs) and specialized loss functions. The proposed model architecture incorporates a convolutional layer, global average pooling, ReLU activation, and novel loss functions (FL and CILF) to improve classification performance. The CNN architecture consists of three main layers: the convolutional layer (ConvL), global average pooling layer (GAPL), and fully connected layer (FCL). Trained on RFCI images with dimensions 299 x 299 x 3, the model effectively captures low-level features such as edges and curves, enhancing visual recognition capabilities. Convolutional operations are applied systematically across the entire image, with filters learning weights during training to extract relevant features. Experimental evaluation is conducted using two publicly available Ocular Health Dataset (OHD) datasets, comparing the proposed model with established baseline models (DenseNet-169, EfficientNet-B7, ResNet-101, Inception-V3, and VGG-19). Additionally, an ablation study is performed to assess the effectiveness of the proposed model. Results, averaged over three cross-validation tests, demonstrate the model's efficacy in classifying ocular diseases, particularly for categories such as CATR, AMD, and GLU.