Transfer Learning and Optimized Machine Learning Techniques for Multiclass Diabetic Retinopathy Classification Using...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic Retinopathy (DR) is a severe diabetes complication causing vision loss.
• Automated DR detection using CNNs shows promise but faces challenges with accuracy and data limitations.
• Current methods require improvement for reliable DR screening.

Purpose of the Study:

• To evaluate an integrated transfer learning framework with adaptive training for multiclass retinal image classification.
• To enhance the diagnostic accuracy and computational efficiency of DR detection systems.
• To address limitations of traditional approaches in DR screening.

Main Methods:

• Implemented a framework combining transfer learning, dimensionality reduction, and adaptive training.
• Utilized a ResNet50 backbone pretrained on ImageNet for feature extraction.
• Trained and evaluated the model on a large, publicly available dataset of retinal images.

Main Results:

• Achieved an overall accuracy of 84% and a maximum class-specific accuracy of 89%.
• Demonstrated high sensitivity (up to 97%) and an F1-score of 92%.
• Showcased reasonable classification performance under constrained data conditions.

Conclusions:

• The integrated framework shows feasibility for multiclass DR classification using transfer learning and adaptive strategies.
• The approach offers potential for computer-assisted retinal image analysis systems.
• Further validation on high-resolution clinical datasets is crucial for practical deployment.