Abstract: In this paper, an improved multi-task learning method is proposed. The main structure of proposed network contains a feature extraction network and a dual path network for optic disc segmentation and location respectively. And through end-to-end training and testing, the multi-task purpose of automatic optic disc segmentation and location can be achieved. In the decoding phase of the feature extraction network, a dense layer is used to extract the context features of fundus images. The optic disc segmentation task relies on the decoding stage to gradually restore original image resolution and obtain the entire optic disc outline, whereas the optic disc localization task is to obtain accurate optic disc center coordinates by extracting context features of fundus images with atrous space pyramid module and pyramid pooling module. Optic disc segmentation and center localization of 350 fundus images are demonstrated. Experimental results show that the Dice coefficient between the automatically segmented optic disc results and manually marked optic disc areas is 0.965, and the average distance between the automatic localization and the manually marked optic disc center is 0.191 mm (34.7 pixels).

Key words: optic disc segmentation, optic disc localization, dense block, spatial pyramid pooling, atrous spatial pyramid pooling, multi-task learning