Automated Quality Optimized Phase Selection in Ultra-widefield Angiography Using Retinal Vessel Segmentation with Deep Neural Networks

Abstract

Purpose : Manual selection of optimal ultra-widefield fluorescein angiography (UWFA) frame for clinical evaluation or analysis is subjective and time-consuming. We evaluated the performance of a retinal vessel area based automated arteriovenous (early) and late phase frame selection tool.

Methods : A total of 13980 UWFA sequences from 463 sessions were used to evaluate the performance of early-late frame selection tool. Retinal vessel areas are extracted by a deep learning algorithm from all available UWFA frames. Automated identification of the image with maximum retinal vessel area was used for selection of the optimum early phase frame for each session. Late phase frame is selected from the images acquired minimum 4 minutes after dye injection that most closely mirrored the vascular area from the selected early image. A trained image analyst evaluated the automatically selected pairs for individual and combined successes of early and late phase frame selection.

Results : Automated selection tool successfully identified appropriate images for both phases in 400 out of 463 visits (86.4%). Success rate for identifying early and late images individually were 91.8% (425/463) and 94.6% (438/463) respectively. Of the 38 images that were considered unsuccessful, the algorithm was unable to identify an early image in 6 sessions; in the remaining 32 visits the images selected were not considered as optimal because of image quality issues (e.g., non-central FOV), Of the sessions where the algorithm did not detect an optimal late image, the algorithm failed to identify a late image in 10 images; in the remaining 15, a superior late phase image was available based on image focus and FOV.

Conclusions : We demonstrated the feasibility of automated UWFA frame selection using deep learning assisted retinal vessel area measurements. Clinical application of this tool might significantly reduce manual image selection and UWFA evaluation times.

This is a 2020 Imaging in the Eye Conference abstract.

 

Automatically selected early (A) and late (B) phase angiographic frames and corresponding retinal vessel masks (C, D) created with deep neural networks.