Enhancing pathological feature discrimination in diabetic retinopathy multi-classification with self-paced...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) diagnosis is challenging due to subtle lesions and reliance on expert assessment.
• Deep learning (DL) shows promise but is often hindered by low-quality data and small sample sizes.
• Existing DL methods require improvement for robust DR detection in clinical settings.

Purpose of the Study:

• To develop a novel deep learning framework for accurate and efficient diabetic retinopathy detection.
• To address data limitations and enhance feature extraction in DR image analysis.
• To improve classification consistency and performance for early DR diagnosis.

Main Methods:

• Implemented a self-paced progressive learning approach, introducing training samples from simple to complex.
• Utilized randomized multi-scale image reconstruction for advanced data augmentation and feature extraction.
• Employed ensemble learning with Kullback-Leibler (KL) divergence-based collaborative regularization for classification consistency.

Main Results:

• Achieved an AUC of 0.9907 in 4-class classification on an integrated dataset, a 2.2% improvement over ResNet-50.
• Demonstrated high recall (97.65%) and precision (96.54%) for the No-DR class and excellent performance for the Severe class.
• Showcased superior classification on limited data and robust localization of subtle lesions using multi-scale progressive learning.

Conclusions:

• The proposed deep learning framework significantly enhances diabetic retinopathy detection accuracy and robustness.
• The method effectively overcomes data quality and sample size limitations in DR diagnosis.
• This framework shows strong potential for practical clinical deployment in diabetic retinopathy screening.