Deep-Learning Based Quality Assessment in Adaptive Optics Ophthalmoscopy Images

Abstract

Adaptive optics ophthalmoscopy (AOO) is an imaging modality that offers superior resolution compared to conventional fundus photography. It is widely used in the clinical study of diseases affecting retinal vessel morphology, where biomarkers are computed from vessel diameters extracted via image segmentation. However, the reliability of these biomarkers depends heavily on segmentation accuracy, which is often compromised by the variable quality of AOO images, commonly affected by local defocus. In this work, we propose a patch-based deep learning method that assigns a continuous quality score in the range [0, 1] to each image patch, producing local quality maps. These maps highlight high-quality regions suitable for robust biomarker extraction. Our model achieves a mean absolute error of 0.145 on our private dataset. We further extend our method to flicker sequences for best-frame selection, achieving high Top-3 accuracy and enhancing segmentation performance for external vessel diameters and wall-to-lumen ratio estimation.