Abstract: In order to improve the segmentation accuracy of atrophic lesions of high myopia in fundus images, aiming at the problems of poor quality of the fundus images of different individuals and the difficulty of segmentation due to the blurred border between atrophic lesions and adjacent tissues, a segmentation algorithm of high myopia atrophy lesions based on multi-scale depth supervision is proposed. Firstly, an optimization algorithm is developed to make the fundus image organization structure clear and uniform in style, reducing the difficulty of distinguishing complex features. Because V-Net can only obtain lower segmentation accuracy, the MS-V-Net that combines multi-level and low-level features to form multi-scale feature learning can extract semantic information in images of different scales. More importantly, the deep supervision of each multi-scale module of MS-V-Net eventually forms a closely supervised MSS-V-Net. Compared with the original V-Net segmentation method, it improves the discrimination of important semantic information by the network and generalization ability. The experimental results show that the Dice box-plot of the proposed method exhibits a trend of fewer outliers, larger median, shorter box length, smaller upper and lower interval, and shorter two lines outside the box, effectively improving the segmentation of atrophic lesion images of high myopia.

Key words: deep learning, multi-scale deep supervision, high myopia, image segmentation