Automated Optic Disc Finder and Segmentation Using Deep Learning for Blood Flow Studies in the Eye

In-vivo imaging with laser speckle flowgraphy (LSFG) enables optical measurement of an index of blood flow in the retina. LSFG can help observe blood flow changes in various diseases, including optic nerve problems such as glaucoma, ischemic optic neuropathy, and others. However, identification of the optic disc in LSFG images is particularly challenging because of limited contour information in the rendered blood flow maps. In this study, we adopted a state-of-the-art U-Net approach (nnU-Net) to automatically identify the optic disc region based on input two-channel LSFG composite blood flow maps and infrared light intensity images. Since the optic disc contour is not always obvious in LSFG, a trained neuro-ophthalmologist (Expert 1) traced the optic discs in color fundus photographs from the same eye; these masks were then registered into the LSFG domain. One hundred subjects (training/test dataset ratio: 70/30) were used in this study. The nnU-Net was trained to identify the optic disc just based on the LSFG composite and light intensity images. After training, we compared the difference between nnU-Net's output and Expert 1 with the difference between Expert 1 and a second clinician (Expert 2) in the test dataset. Both the Dice coefficient and the intersection over union (IoU) index showed the nnU-Net's predictions were significantly closer to Expect 1's tracing than Export 2's (p-values < 0.001). In summary, having a robust optic disc segmentation in LSFG can reduce tedious manual tracing and will also be a foundation for future developments of automated region-based feature extraction in LSFG.