Extreme Points Guided Macular Hole Segmentation from Color Fundus Images

Timely diagnosis of macular hole is very crucial for saving or improving the patient’s visual acuity. Thus, in this paper, we develop a deep learning-based macular hole segmentation method. Our approach first estimates the four extreme points of the macular hole from the color fundus image using ExtremeNet. Then, guided by these extreme points, the Deep Extreme Cut network is employed to achieve precise segmentation of the macular hole. We compare the proposed method with two representative image segmentation methods (i.e., adaptive threshold and UNet) to verify its efficacy. The visual comparison results suggest that the macular holes segmented by our method are much closer to the ground truth than those segmented by the two comparison methods. The quantitative evaluation results show our method significantly improves the accuracy, sensitivity, specificity and dice coefficient by 0.19% (p<0.05), 13.73% (p=0.08), 0.18% (p<0.05) and 20.70% (p<0.05) respectively compared to the second-best method. Overall, our method is capable of automatically segmenting the macular hole from the patient’s color fundus image and exhibits great potential to apply in the clinical scenarios for helping the surgeons efficiently screen macular holes from a large amount of color fundus images.