Three-Dimensional Semantic Segmentation of Diabetic Retinopathy Lesions and Grading Using Transfer Learning.

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision impairment, necessitating early detection and accurate grading.
• Current diagnostic methods for DR can be time-consuming and subjective, highlighting the need for automated solutions.

Purpose of the Study:

• To develop and evaluate learning-based techniques for automated segmentation and multi-classification of diabetic retinopathy lesions.
• To enhance the accuracy and efficiency of DR diagnosis using deep learning models.

Main Methods:

• Utilized a pre-trained Xception model for deep feature extraction, followed by Deeplabv3 for semantic segmentation of DR lesions.
• Developed a multi-classification model by fusing features from efficient-net-b0 and squeeze-net, optimized using the Marine Predictor Algorithm (MPA).
• Employed neural network and K-Nearest Neighbors (KNN) classifiers for grading DR lesions into stages 0, 1, 2, and 3.

Main Results:

• Achieved effective semantic segmentation of DR lesions through optimized hyperparameter selection for the segmentation model.
• Successfully performed multi-classification of DR lesions with high accuracy, outperforming existing state-of-the-art methods.
• Validated the proposed method on diverse open-access datasets, including DIARETDB1, e-ophtha-EX, IDRiD, and Messidor.

Conclusions:

• The proposed learning-based approach demonstrates superior performance in segmenting and classifying diabetic retinopathy lesions.
• This AI-driven method offers a promising tool for early and accurate diagnosis of DR, potentially reducing vision loss.
• The study highlights the effectiveness of feature fusion and advanced machine learning algorithms in medical image analysis for ophthalmology.