MSFF-Net: Multi-Scale Feature Fusion Network for Gastrointestinal Vessel Segmentation

Purpose The accurate and automatic segmentation of gastrointestinal wall vessels can help to prevent endoscope tip related perforation. Methods based on deep learning and convolution neural network in many different kinds of medical image segmentation tasks have achieved remarkable performance, but in the gastrointestinal vessels segmentation task, because fold and vascular structure characteristics are very similar, blood vessels, fuzzy boundaries, flare and other interference factors, it is very easy to produce false segmentation and rupture of blood vessels. We therefore propose a new multi-scale future fusion network to tackle the aforementioned issues. Methods Our proposed segmentation network consists of encoding, decoding modules, attention module and future fusion module. Through the convolution operation of future fusion module, the output features of each encoder are effectively fused, and the multi-scale information is fully utilized. In addition, we further improve the loss function and enhance the ability of the network to distinguish folds and vessels and predict vascular connectivity by giving different weights to the front background. Results The proposed network is evaluated on our own gastrointestinal wall vessel data set. Experimental results show that compared with the existing advanced vascular segmentation networks, the proposed network has better segmentation performance in the gastrointestinal wall vascular dataset. Conclusion The proposed future fusion method and attention structure loss can better perform feature extraction and fusion according to the characteristics of the gastrointestinal wall vessels to achieve better results.