Abstract:
This research presents a deep learning based diagnostic system designed to enhance the early detection of Retinitis Pigmentosa using Fundus Autofluorescence imaging. Retinitis Pigmentosa is a rare inherited retinal disorder that progresses slowly and is often diagnosed only after substantial vision damage has occurred. Clinical diagnosis at early stages remains difficult due to subtle biomarkers, high variability across patient images, and limited availability of well annotated datasets. Conventional deep learning techniques struggle in such low data conditions, and many existing diagnostic models lack interpretability, which reduces their value for clinical decision making.
The study introduces a data efficient approach that integrates advanced preprocessing, a lightweight convolutional architecture, and a fuzzy logic reasoning layer to generate clinically meaningful predictions. Preprocessing steps such as normalization, noise filtering, and contrast enhancement were used to improve the visibility of early metabolic abnormalities in FAF images. The model was trained on a curated dataset balanced across normal and RP affected cases, with data augmentation applied to reduce overfitting and increase variability. The prediction output of the deep learning model was converted into medically interpretable stages through a fuzzy classification mechanism that presents results as Healthy, Mild, Moderate, or Severe along with confidence scores.
A complete diagnostic platform named Retina Vision was implemented using Flask for backend APIs, React for the user interface, and MongoDB for secure data storage. The system supports real time prediction, patient record management, and automated report generation. Initial evaluation indicates promising performance and reinforces the importance of combining deep learning with interpretability strategies for rare disease detection. The proposed solution contributes toward more accessible, consistent, and clinically reliable screening of Retinitis Pigmentosa, especially in environments with limited specialist expertise.
