Related Experiment Video
Updated: Jun 5, 2026
10:25
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
Deep learning-based cross-modal MR-CT registration for brain metastases radiotherapy with multi-scale feature
Chengjian Xiao1, Weixiang Lin1, Feilong Tian1
1Department of Radiation Oncology, Ganzhou Cancer Hospital, Ganzhou, 341000, People's Republic of China.
Scientific Reports
|June 3, 2026
Summary
This study introduces MSFRTransMorph, a Transformer-based framework for accurate brain MRI-CT image registration, improving radiotherapy planning. While enhancing alignment accuracy, it noted a trade-off with deformation stability.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiotherapy
Background:
- Accurate multimodal deformable registration between MR and CT images is crucial for radiotherapy target delineation in brain metastases.
- Modality differences and complex deformations pose challenges to robust MR-CT image alignment.
Purpose of the Study:
- To propose an enhanced Transformer-based registration framework, MSFRTransMorph, for improved cross-modal spatial correspondence modeling.
- To strengthen hierarchical feature interactions and fine-grained deformation modeling using a multi-scale feature refinement (MSFR) module.
Main Methods:
- Developed MSFRTransMorph, integrating a multi-scale feature refinement module into the TransMorph architecture.
- Utilized brainstem segmentation for anatomical guidance during training, excluding tumor annotations.
- Trained and evaluated the model on 141 patient MR-CT image pairs.
Main Results:
- MSFRTransMorph demonstrated superior registration accuracy for the brainstem and gross tumor volume (GTV), achieving a GTV DSC of 74.25% and HD95 of 16.92 mm.
- Increased volumetric overlap correlated with a higher proportion of local deformation folding, indicating reduced topological regularity.
- The multi-scale refinement mechanism improved cross-modal feature representation and volumetric alignment accuracy.
Conclusions:
- The proposed MSFRTransMorph enhances MR-CT registration accuracy for brain metastases radiotherapy.
- A trade-off exists between registration precision and deformation stability, necessitating explicit balancing in Transformer-based frameworks.