A multi-task deep-learning system for assessment of diabetic macular ischemia on optical coherence tomography angiography images.

Purpose We aimed to develop and test a deep-learning (DL) system to perform image quality and diabetic macular ischemia (DMI) assessment on OCTA images. Methods This study included 7,194 OCTA images with diabetes mellitus for training and primary validation, and 960 images from three independent datasets for external testing. A trinary classification for image quality assessment and presence or absence of DMI for DMI assessment were labelled on all OCTA images. Two DenseNet-161 models were built for both tasks for OCTA images of superficial and deep capillary plexus, respectively. External testing was performed on three unseen datasets in which one dataset using the same model of OCTA device as of the primary dataset, and two datasets using another brand of OCTA device. We assessed the performance by using the area under the receiver operating characteristic curves with sensitivities, specificities, accuracies, and the area under the precision-recall curves with precision. Results For the image quality assessment, analyses for gradability and measurability assessment were performed. Our DL system achieved the AUROCs >0.948 and AUPRCs >0.866 for the gradability assessment, AUROCs >0.960 and AUPRCs >0.822 for the measurability assessment, and AUROCs>0.939 and AUPRCs >0.899 for the DMI assessment across three external validation datasets. Grad-CAM demonstrated the capability of our DL system paying attention to regions related to DMI identification. Conclusions Our proposed multi-task DL system might facilitate the development of a simplified assessment of DMI on OCTA images among individuals with DM at high risk for visual loss.