Deep-learning-based automatic computer-aided diagnosis system for diabetic retinopathy.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• The increasing prevalence of diabetic retinopathy (DR) necessitates efficient diagnostic tools.
• Computer-aided diagnosis (CAD) systems offer potential for early detection and management of DR.
• Deep neural networks (DNNs) show promise in analyzing complex biomedical imaging data.

Purpose of the Study:

• To develop and evaluate an optimal DR CAD solution using the AlexNet deep neural network (DNN).
• To enable early disease detection and diagnosis decision-making for diabetic retinopathy.
• To achieve a high accuracy five-class classification of DR stages.

Main Methods:

• Utilized AlexNet, a convolutional neural network (CNN), for feature extraction in DR detection.
• Implemented a multi-level optimization including pre-processing, Gaussian Mixture Model (GMM) segmentation, and Region of Interest (ROI) localization.
• Employed Principle Component Analysis (PCA) and Linear Discriminant Analysis (LDA) for feature selection, followed by Support Vector Machine (SVM) classification.

Main Results:

• The proposed AlexNet DNN-based DR system achieved a classification accuracy of 97.93% with LDA feature selection and FC7 features.
• With PCA feature selection, the system achieved an accuracy of 95.26%.
• Comparative analysis showed the AlexNet DNN approach outperformed the Spatial Invariant Feature Transform (SIFT) technique (94.40% accuracy).

Conclusions:

• The AlexNet DNN-based CAD system demonstrates superior performance for diabetic retinopathy classification.
• LDA feature selection provides better results compared to PCA for this DR detection task.
• The developed system holds significant potential for improving early DR diagnosis.