Abstract
Purpose :
While there have been multiple automatic segmentation techniques for different types of fundus images, few target specifically optical coherence tomography angiography (OCTA) images due to the difficulty of manual annotation during the data collection process, despite its importance in the diagnosis and treatment of various ophthalmic diseases. We developed and tested a deep learning approach to automatically segment the vascular structures in OCTA scans without the need for a large amount of manually traced images for training.
Methods :
We pre-trained a deep learning-based image segmentation model, ResUNet, to automatically segment the axons and dendrites in retinal ganglion cell (RGC) images. A total of 110 RGC scans with manual annotations from human experts were used to develop this model, where 90 scans were used for training and the remaining 20 images were reserved as testing data. Next, transfer learning was applied by fine tuning the model on 4 manually annotated OCTA images, while 4 additional OCTA images were used for evaluation. Pre-processing and post-processing steps were applied to adjust the brightness and contrast of input images and segmentation outputs from the model.
Results :
Our model can effectively and automatically segment the structures in both RGC and OCTA images. Quantitatively, our segmentation model achieves average foreground, background and overall accuracy of 0.689, 0.998 and 0.997 for the RGC images, and 0.662, 0.968 and 0.955 on the OCTA images, respectively.
Conclusions :
Our model can successfully adapt the prior knowledge learned from axon and dendrite segmentation of RGC images to the segmentation of vascular structures in OCTA images.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.