CNN-Based Quality Assessment for Retinal Image Captured by Wide Field of View Non-Mydriatic Fundus Camera.

In general, a high percentage of the retinal images captured by any non-mydriatic fundus cameras in telemedicine environment present inadequate quality for reliable diagnostics of retinal pathologies. An automatic quality assessment at the retinal image acquisition moment is indispensable for efficient screening program. In this paper, we present automatic quality assessment methods for retinal images captured by wide field of view (200° FOV) non-mydriatic fundus camera, using several CNN architectures with different configuration. We evaluate the performance of the eight off-the-shelf CNN architectures using sensitivity, specificity, accuracy, precision and Area Under Curve (AUC) of the Receiver Operating Characteristics (ROC) curve. The best performance is presented by the Vgg16 CNN with 100% of accuracy, and the Squeezenet presents very good performance with a lowest complexity.