Detection of Retinitis Pigmentosa Using Hybrid Deep Learning Architecture

Abstract

Retinitis Pigmentosa (RP) is a hereditary vision disorder where the photosensitive cells of the retinal degenerate thus leading to blindness. RP victims commonly experience night vision issues, then constrict side eye vision, which may result in tunnelling eyesight and even total blindness. The early detection is necessary as it assists in the planning of treatment and prevents further development of the infection. This paper presents a new and improved method utilizing a Vision Transformer (ViT-B16), that analyses the whole picture of retina to more accurately discover patterns, along with EfficientNet-B4, a CNN algorithm that retrieves crucial information from pictures. It is recommended to use a multi-modal images training that supports Ultra-Wide Field (UWF), Fundus Autofluorescence (FAF), including colour fundus pictures. The rotation plus brightness modification procedure is carried out by data augmentation to expand the quantity of dataset in order to aid the algorithm in learning more efficiently. Gradient weighted class activation mapping(Grad-CAM) was applied to identify the regions of retinal surface which made the biggest contributions to algorithm's forecasting. Additionally, RP effectiveness was separated into Early, Moderate, and Severe levels with a stage estimation method utilizing Grad-CAM activation intensity. Through testing and validation, the architecture exhibits exceptionally high precision, F1-Score, and accuracy. It provides a scalable and effective method of detecting Retinitis Pigmentosa through significantly increased accuracy through the use of multi-modal image training. The previous RP identification availed by this method allows appropriate medication in good time that improves the overall quality of life of the affected individuals.