Robust Foveal Avascular Zone Segmentation and Anatomical Feature Extraction from OCT-A Handling Inter-Expert Variability

The Foveal Avascular Zone (FAZ) is commonly analyzed from OCT-A images to diagnose retinal diseases. When quantitative measures are required, ophthalmologists manually draw the contours of the FAZ and compute several anatomical features from these annotations. However, there is no clear consensus among experts on the FAZ definition in pathological cases. In this work, we developed a novel framework to automatically segment the FAZ based on three different expert annotations to be robust to inter-expert variability. A loss based on the Hausdorff Distance was used to specifically address complex pathological FAZ that are usually poorly segmented by automatic methods. We used a dataset of more than 200 images to train and test our model, and we achieved similar or even better segmentation performance than individual experts.