Deep Learning Segmentation of Non-perfusion Area from Color Fundus Images and AI-generated Fluorescein Angiography

Abstract

Purpose : Deep learning models can segment the non-perfusion area (NPA) in retinal vein occlusion (RVO) using fluorescein angiography (FA) or color fundus images; however, their comparative effectiveness remains unclear. In addition, some generative AI models can translate color fundus images into FA-like images, yet their diagnostic benefits are unknown. We performed a retrospective, observational study to evaluate and compare deep learning models that use fluorescein angiography (FA) images, color fundus images, and AI-generated synthetic FA images.

Methods : We retrospectively collected 403 sets of images from 319 RVO patients and separated them into training (330 images), validation (38 images), and test (35 images). U-Net models with Monte Carlo dropout were trained using three different types of images: (1) FA images, (2) color fundus images, and (3) a combination of color fundus and synthetic FA images. Prediction accuracy and uncertainty were evaluated using Dice scores and Monte Carlo dropout, respectively, tested by the Wilcoxon signed-rank test.

Results : The FA-based model showed the highest median Dice score, yet the other two models also presented comparable performance. We found no statistical significance in median Dice scores between the models. However, the Color model showed higher uncertainty than the others (p<0.05).

Conclusions : Deep learning models can detect NPAs from color fundus images with reasonable accuracy, albeit with compromised prediction stability. Synthetic FA helps reduce misleading uncertainty estimates through its image enhancement.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Accuracy of NPA prediction with different input sources
Dice score and sensitivity of models with different input. The red line indicates median, and the whiskers show 1.5 times IQR.

 

Monte Carlo dropout uncertainty
Median standard deviation (SD) in each image, measuring the degree of uncertainty in general. The SD was acquired from 100 predictions using Monte Carlo dropout.