Automated Segmentation of Microvasculature in Ophthalmic Neovascular Diseases

Abstract

Purpose : Segmentation of vessel patterns and types, especially neovascularization associated with ischemia and/or tumor growth, provides crucial information for the diagnosis and treatment for a number of ophthalmic neovascular diseases, including branch retinal vein occlusion (BRVO). We developed and tested an automated deep learning-based retinal vessel segmentation algorithm that segments the major vessels, as well as micro-vessels, in areas of neovascularization in fundus fluorescein angiography (FFA) images.
<!--[if !supportLineBreakNewLine]-->
<!--[endif]-->

Methods : We applied Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance the contrast of the microvasculature in raw FFA images. Two deep learning-based vessel segmentation models, including a coarse model that segments the major blood vessels and a fine model which segments the microvasculature in neovascularization areas, were then developed. Both models have the same architecture as ResUNet. A total of 498 FFA images, obtained from patients with ophthalmic neovascular diseases such as BRVO or central retinal vein occlusion (CRVO), were used to train the models. The coarse model was trained in a supervised manner using the raw FFA images as inputs, while the fine model was trained in a weakly-supervised manner using the FFA images after CLAHE enhancement as inputs. Finally, we superimposed the segmentation results of the two models together.
<!--[if !supportLineBreakNewLine]-->
<!--[endif]-->

Results : Our deep learning approach reliably and accurately segmented the major vessels as well as the microvasculature in neovascularization areas in FFA images. Quantitatively, the coarse model achieves average foreground, background and overall accuracy of 0.976, 0.999 and 0.989 on a testing dataset of FFA images with high-quality manual segmentation of major blood vessels.

Conclusions : The incorporation of our approach into the clinical field will aid in the diagnosis and management of ophthalmic neovascular disease and improve the efficiency of providing point of contact diagnoses, saving patients and providers significant time and resources while obtaining accurate and robust diagnoses.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.