Pyramid Network With Quality-Aware Contrastive Loss for Retinal Image Quality Assessment.

Area of Science:

• Medical Imaging
• Computer Vision
• Ophthalmology

Background:

• Retinal image quality is crucial for accurate diagnosis, as low-quality images increase misdiagnosis risk.
• Current deep learning methods for retinal image quality assessment (RIQA) offer limited qualitative feedback, classifying images as 'Good,' 'Usable,' or 'Reject.'

Purpose of the Study:

• To develop a unified framework, QAC-Net, for comprehensive RIQA, providing both qualitative and quantitative quality scores.
• To enhance feature extraction for improved prediction accuracy in RIQA tasks.

Main Methods:

• QAC-Net employs a pyramid network structure for multi-scale feature learning and feature purification via consistency loss.
• A quality-aware contrastive (QAC) loss is utilized to improve feature representation by considering inter-image quality relationships.
• A new dataset of 2,300 distorted retinal images with subjective quality scores was created for quantitative evaluation.

Main Results:

• QAC-Net demonstrated competence in both qualitative and quantitative RIQA tasks.
• Experimental results on public and the newly constructed datasets confirmed the framework's considerable performance.

Conclusions:

• QAC-Net offers a robust solution for RIQA, addressing the limitations of existing methods by providing detailed quality feedback.
• The proposed framework has the potential to reduce misdiagnoses by enabling more precise assessment of retinal image quality.