MF‐ResUnet: A 3D Liver Image Segmentation Method Based on Multi‐Scale Feature Fusion

ABSTRACT Background Due to the variable shapes of the liver parenchyma, minimal voxel intensity differences with adjacent organs, and discontinuous liver boundaries, automatic liver segmentation from computerised tomography images poses significant challenges. Methods In this study, we propose a 3D liver segmentation method based on multiscale feature fusion. This network employs SE channel attention to recalibrate liver features. Additionally, it utilises an AMF module for multiscale feature fusion to obtain rich spatial information. Furthermore, we introduce the NGAB module to address the deteriorating effects of dilated convolutions as the dilation rate increases, contributing to enhanced feature representation and improving accuracy in liver segmentation. Results Experimental results on the publicly available LiTS2017 dataset and 3DIRCADb dataset show that our proposed framework achieves a DSC of 0.977 and 0.967 in liver segmentation, respectively. Conclusions The proposed method can adequately capture multiscale characteristics, showing promising prospects for automatic liver segmentation.