IDAA-NET: An Image Domain Adaptive Alignment Network for Unsupervised Liver Vessel Segmentation from CTA Images
- Resource Type
- Conference
- Authors
- Geng, Haixiao; Fan, Jingfan; Yuan, Yujia; Ai, Danni; Duan, Feng; Yang, Jian
- Source
- 2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) Bioinformatics and Biomedicine (BIBM), 2023 IEEE International Conference on. :1925-1928 Dec, 2023
- Subject
- Bioengineering
Computing and Processing
Engineering Profession
Robotics and Control Systems
Signal Processing and Analysis
Training
Image segmentation
Adaptive systems
Three-dimensional displays
Liver diseases
Supervised learning
Predictive models
CTA image
MRA image
Vessel segmentation
Unsupervised learning
Adversarial learning
- Language
- ISSN
- 2156-1133
Accurate segmentation of liver vessel from CTA image is important for the diagnosis and treatment of liver diseases. The quality of labeled data directly affects the prediction results of the segmentation model. Compared with CTA image, MRA image has clearer 3D vasculature. Therefore, in order to reduce the reliance of the labeled CTA image which may contain ambiguous vessel contours, we propose a novel unsupervised liver vessel segmentation method based on image domain adaptive alignment network (IDAA-Net) by using labeled MRA and unlabeled CTA images. The IDAA-Net mainly contains three modules: 1) A spatial alignment module (SAM) is introduced to convert MRA image slice to synthetic CTA image slice for achieving spatial alignment of the different modality data in the feature and image levels; 2) An artifact removal module (ARM) is designed to eliminate background artifacts of synthetic CTA from SAM by using the liver label in MRA; 3) An adversarial segmentation module (ASM) is proposed to obtain the optimal segmentation by jointly adversarial learning and supervised learning between the predicted segmentation and the ground-truth label of MRA image. Experiments on the public and private datasets show that our method achieves comparable performance with state-of-the-art supervised method and outperforms the existing unsupervised segmentation methods.