Related Experiment Video
Updated: Jan 8, 2026
07:13
Author Spotlight: An Efficient and Robust Software for Automated Fusion of Multiple Preclinical Imaging Modalities
Published on: October 27, 2023
1.6K
BIASNet: A bidirectional feature alignment and semantics-guided network for weakly-supervised medical image
Housheng Xie1, Xiaoru Gao1, Guoyan Zheng1
1Institute of Medical Robotics, School of Biomedical Engineering, Shanghai Jiao Tong University, No.800 Dongchuan Road, Shanghai, 200240, China.
Medical Image Analysis
|December 18, 2025
Summary
This study introduces BIASNet, a novel weakly-supervised method for medical image registration that leverages anatomical labels for improved accuracy. BIASNet outperforms existing methods by effectively aligning multi-scale features and enforcing anatomical consistency.
Area of Science:
- Medical Imaging
- Computer Vision
- Machine Learning
Background:
- Medical image registration is crucial for clinical applications but unsupervised deep learning methods show limitations.
- Weakly-supervised methods offer potential but haven't fully utilized segmentation label information.
- Advancements in universal segmentation models enable easier extraction of anatomical labels.
Purpose of the Study:
- To develop a novel weakly-supervised medical image registration method (BIASNet) that effectively utilizes anatomical and structural prior information from segmentation labels.
- To improve the accuracy and robustness of medical image registration by integrating semantics-wise and intensity-wise feature alignment.
Main Methods:
- Proposed BIdirectional feature Alignment and Semantics-guided Network (BIASNet) for weakly-supervised image registration.
- Employed a dual-attribute learning scheme with a novel BIdirectional Alignment and Fusion (BIAF) module to extract semantics-wise and intensity-wise features.
- Integrated a semantics-guided progressive registration framework and anatomical region deformation consistency learning.
Main Results:
- BIASNet demonstrated superior performance compared to state-of-the-art deformable registration methods on three challenging datasets.
- The proposed method effectively leverages multi-scale features and anatomical priors for accurate deformation field estimation.
- The anatomical region deformation consistency learning regularized deformations in target anatomical regions.
Conclusions:
- BIASNet represents a significant advancement in weakly-supervised medical image registration by effectively incorporating segmentation label information.
- The method achieves state-of-the-art results, offering a more accurate and robust solution for clinical applications.
- Publicly available source code facilitates further research and development in medical image registration.
