Diabetic retinopathy classification based on multipath CNN and machine learning classifiers

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic Retinopathy (DR) diagnosis relies on manual fundoscopy by eye care professionals.
• Early detection and precise grading of DR are crucial for effective patient management.
• Current methods may lack the speed and accuracy needed for widespread screening.

Purpose of the Study:

• To develop an automated system for Diabetic Retinopathy grading using deep learning and machine learning.
• To extract and categorize features from fundus images based on DR severity.
• To enhance early DR detection and improve patient care pathways.

Main Methods:

• Utilized a Multipath Convolutional Neural Network (M-CNN) for extracting global and local image features.
• Employed machine learning classifiers, including Support Vector Machine (SVM), Random Forest, and J48, for DR severity categorization.
• Evaluated the model on diverse public datasets: IDRiD, Kaggle (DR detection), and MESSIDOR.

Main Results:

• The M-CNN combined with the J48 classifier demonstrated superior performance.
• Achieved an average accuracy of 99.62% for Diabetic Retinopathy grading.
• The system showed high effectiveness in both DR grading and early disease detection.

Conclusions:

• The proposed automated method offers a highly accurate and efficient solution for Diabetic Retinopathy grading.
• This AI-driven approach has the potential to significantly aid in early DR detection and management.
• The M-CNN and J48 classifier combination provides a robust tool for clinical application in ophthalmology.