Abstract: Fundus images often contain a large number of small blood vessels with significant noise interference and blurred boundaries, making segmentation challenging. To address these characteristics, a fundus image segmentation method called MDF _Net&CD ( Multi-Directional Features neural Network and Connectivity Detection) is proposed, based on multidirectional features and connectivity detection. A deep neural network model, MDF_Net( Multi-Directional Features neural Network), is designed to take different directional feature vectors of pixels as input. MDF_Net is used for the initial segmentation of the fundus images. A connectivity detection algorithm is proposed to revise the preliminary segmentation results of MDF _ Net, according to the geometric characteristics of blood vessels. In the public fundus image dataset, MDF_Net&CD is compared with recent representative segmentation methods. The experimental results show that MDF_Net&CD can effectively capture the detailed characteristics of pixels, and has a good segmentation effect on irregular, severely noisy, and blurred boundaries of small blood vessels. The evaluation indices are balanced, and the sensitivity, F1 score, and accuracy are better than other methods participating in the comparison.

Key words: vessel image segmentation, multi-directional features, connectivity detection, deep neural network