DeepRetinaNet: An Automated AI-Based Framework for Retinal Disease Diagnosis

Abstract

Automated retinal disease diagnosis leveraging cutting-edge computer vision methodologies supports clinicians in the early identification of pathological conditions. This investigation delivers a novel framework, DeepRetinaNet for automating retinal disease diagnosis. The developed DeepRetinaNet model has two stages of novelties, including vessel extraction followed by disease identification. In the vessel extraction stage, the green channel, known for its heightened sensitivity to retinal vascular structures, is extracted from the source images. Subsequently, the vessel extraction network: RetiSegNet, processes these green channel images to extract retinal vessels, generating binary vessel maps. During the fusion phase, the original fundus images are combined with the extracted vessel maps to produce fused representations, encapsulating enriched spatial details from both sources. In the identification stage, these fused images are utilized to train the proposed classification framework: STDeepNet, which incorporates Modified Identity (MI), Modified Convolution (MCONV) blocks, and Long Short-Term Memory (LSTM) layers to effectively identify the diseases. The efficacy of the developed technique is corroborated using visual illustration and objective analysis. Also, the efficiency of the designed framework is verified on six benchmark datasets. The proposed framework demonstrates superior performance compared to 49 state-of-the-art methods, achieving notable accuracy in retinal disease diagnosis.