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

In computer-assisted diagnostics, assessing the quality of retinal images, especially for DR, is vital. While current Image Quality Assessment (IQA) methods lean on Transfer Learning (TL), their adaptability to specific IQA demands, especially for DR images, remains questionable due to the challenges of detecting detailed distortions. In this paper, we propose a novel framework termed Saliency-Aware Mutual Learning for Image Quality Assessment (SAM-IQA). This framework intricately learns the relationship between the representation of salient regions and the overall representation of fundus images. Specifically, we introduce a dual-branch network architecture that simultaneously extracts global features from distorted images and local features from their salient regions. This dual extraction promotes the learning of both coarse and fine-grained feature representations. To further enhance feature extraction, we integrate mutual learning techniques within this dual-branch network, facilitating the capture of high-level content presentation and low-level fusion quality features. This integration results in a more holistic quality assessment. Our evaluation of the DeepDRiD dataset demonstrates the efficacy of SAM-IQA. The method achieved an AUC of 81.5% (↑6.6% vs. previous state-of-the-art (SOTA) methods of 74.9%), outperforming existing IQA methods.