Automated Deep Learning-Based Region-of-Interest Segmentation in Ultra-widefield Angiography

Abstract

Purpose : Ultra-widefield fluorescein angiography (UWFA) is an imaging technique that allows near panretinal visualization of retinal vascular abnormalities. Emerging analytic technology has the potential to quantify features such as retinal vessel characteristics, microaneurysm count, and leakage burden. In order to facilitate the development of efficient semi-automated and automated systems, an accurate automated region-of-interest finder is needed. The purpose this analysis was to develop a deep learning (DL) model capable of defining the ROI at an accuracy similar to human performance (F1 > 0.8).

Methods : This was an IRB-approved image analysis study. 2869 UWFA images from 384 patients over three time points were annotated for the optimal ROI and utilized for this analysis. The data was split into 80% training data and 20% testing data. The training/testing split was carefully curated to confirm that the no patient’s timepoints were included in both testing or training. All data for a given patient was placed in either the training or testing group. Training was conducted in python 3.5 using tensorflow and a 41 layer, 7.7 million parameter UNET convolutional neural network with a kernel width of 3, and a batch size of 50. Patches were resized to 256x256 pixels. Patches were rotated to train the model to be insensitive to orientation.

Results : Of the 2869 images, 1431 (50%) were early images (i.e., obtained between 45 seconds and 90 seconds of fluorescein administration) and 1438 (50%) were late images (i.e., obtained after 5 minutes of fluorescein administration). The final model predicted the ROI on the testing data with an F1 score of 0.83 (precision = 0.76, recall = 0.92). The overall impression of clinical utility was quite high.

Conclusions : The ROI model generated with this analysis demonstrated promising results both based on quantitative performance measures and qualitative inspection of the results. As with many segmentation systems evaluation of overall initial systems may be difficult to encompass solely based on F-score performance alone due to the of what constitutes an ROI as ground truth and potential underlying human-based segmentation errors. Further evaluation of the model on novel images is needed to assess the usability of the model in research and eventually clinical practice.

This abstract was presented at the 2023 ARVO Imaging in the Eye Conference, held in New Orleans, LA, April 21-22, 2023.