Saliency-Aware Mutual Learning for enhanced retinal Image Quality Assessment in diabetic retinopathy diagnostics

Area of Science:

• Medical Imaging
• Computer-Assisted Diagnostics
• Artificial Intelligence

Background:

• Accurate assessment of retinal image quality is crucial for diabetic retinopathy (DR) diagnosis.
• Current Image Quality Assessment (IQA) methods using Transfer Learning (TL) struggle with the detailed distortions present in DR images.

Purpose of the Study:

• To propose a novel framework, Saliency-Aware Mutual Learning for Image Quality Assessment (SAM-IQA), for enhanced IQA of fundus images.
• To improve the adaptability and accuracy of IQA methods for DR detection.

Main Methods:

• Introduced a dual-branch network architecture to extract both global and local features from salient regions.
• Integrated mutual learning techniques to capture high-level content and low-level fusion quality features.
• Developed Saliency-Aware Mutual Learning for Image Quality Assessment (SAM-IQA).

Main Results:

• SAM-IQA achieved an Area Under the Curve (AUC) of 81.5% on the DeepDRiD dataset.
• Demonstrated a significant improvement of 6.6% compared to previous state-of-the-art methods (74.9% AUC).
• Outperformed existing IQA methods in assessing the quality of retinal images for DR.

Conclusions:

• The proposed SAM-IQA framework effectively enhances the quality assessment of fundus images, particularly for DR.
• The dual-branch architecture and mutual learning integration contribute to a more holistic and accurate IQA.
• SAM-IQA represents a significant advancement in computer-assisted diagnostics for DR.