Automated Cone Cell Identification on Adaptive Optics Scanning Laser Ophthalmoscope Images Based on TV-L1 Optical Flow Registration and K-Means Clustering

Cone cell identification is essential for diagnosing and studying eye diseases. In this paper, we propose an automated cone cell identification method that involves TV-L1 optical flow estimation and K-means clustering. The proposed algorithm consists of the following steps: image denoising based on TV-L1 optical flow registration, bias field correction, cone cell identification based on K-means clustering, duplicate identification removal, identification based on threshold segmentation, and merging of closed identified cone cells. Compared with manually labelled ground-truth images, the proposed method shows high effectiveness with precision, recall, and F1 scores of 93.10%, 94.97%, and 94.03%, respectively. The method performance is further evaluated on adaptive optics scanning laser ophthalmoscope images obtained from a healthy subject with low cone cell density and subjects with either diabetic retinopathy or acute zonal occult outer retinopathy. The evaluation results demonstrate that the proposed method can accurately identify cone cells in subjects with healthy retinas and retinal diseases.