Automated Quality Assessment and Image Selection of Ultra-Widefield Fluorescein Angiography Images through Deep Learning

PURPOSE: Numerous angiographic images with high variability in quality are obtained during each ultra-widefield fluorescein angiography (UWFA) acquisition session. This study evaluated the feasibility of an automated system for image quality classification and selection using deep learning. METHODS: The training set was comprised of 3543 UWFA images. Ground-truth image quality was assessed by expert image review and classified into one of four categories (ungradable, poor, good, or best) based on contrast, field of view, media opacity, and obscuration from external features. Two test sets, including randomly selected 392 images separated from the training set and an independent balanced image set composed of 50 ungradable/poor and 50 good/best images, assessed the model performance and bias. RESULTS: In the randomly selected and balanced test sets, the automated quality assessment system showed overall accuracy of 89.0% and 94.0% for distinguishing between gradable and ungradable images, with sensitivity of 90.5% and 98.6% and specificity of 87.0% and 81.5%, respectively. The receiver operating characteristic curve measuring performance of two-class classification (ungradable and gradable) had an area under the curve of 0.920 in the randomly selected set and 0.980 in the balanced set. CONCLUSIONS: A deep learning classification model demonstrates the feasibility of automatic classification of UWFA image quality. Clinical application of this system might greatly reduce manual image grading workload, allow quality-based image presentation to clinicians, and provide near-instantaneous feedback on image quality during image acquisition for photographers. TRANSLATIONAL RELEVANCE: The UWFA image quality classification tool may significantly reduce manual grading for clinical- and research-related work, providing instantaneous and reliable feedback on image quality.