MCFA-UNet: Multiscale Cascaded Feature Attention U-Net for Liver Segmentation

Abstract

Objectives

Accurate automatic liver segmentation has important value for subsequent tumor segmentation, diagnosis, and treatment. In this paper, a Multiscale Cascaded Feature Attention U-Net (MCFA-UNet) neural network model was proposed to solve the problem of edge detail feature loss caused by insufficient feature extraction in existing segmentation methods.

Material and methods

MCFA-UNet is a 3D segmentation network based on U-Net encoding and decoding structure. First, this paper proposes a multiscale feature cascaded attention (MCFA) module, which extracts multiscale feature information through multiple continuous convolution paths, and uses double attention to realize multiscale feature information fusion of different paths. Second, the attention-gate mechanism is used to fuse different levels of feature information, which reduces the semantic difference between coding and decoding paths. Finally, the deep supervision learning method was employed to optimize the network segmentation effect through the feature information of each hidden layer in the decoding path.

Results

MCFA-UNet was evaluated on LiTS and 3DIRCADb datasets. The Dice scores of 0.955 and 0.981 are obtained respectively. Compared with the baseline network, the segmentation accuracy is improved by 5% and 3.5%.

Conclusion

Experimental results show that MCFA-UNet has more accurate segmentation performance than baseline model and other advanced methods.