Related Experiment Video
Updated: Jun 23, 2026
14:08
Automated Midline Shift and Intracranial Pressure Estimation based on Brain CT Images
Published on: April 13, 2013
42.6K
Incorporating a-priori information in deep learning models for quantitative susceptibility mapping via adaptive
Simon Graf1,2, Walter A Wohlgemuth1,2, Andreas Deistung1,2
1University Clinic and Polyclinic for Radiology, University Hospital Halle (Saale), Halle, Germany.
Frontiers in Neuroscience
|March 26, 2024
Summary
Deep learning for quantitative susceptibility mapping (QSM) is improved by incorporating imaging parameters. Adaptive convolution enhances QSM generalizability across various acquisition settings, leading to better tissue characterization.
Area of Science:
- Medical Imaging
- Neuroimaging
- Artificial Intelligence
Background:
- Quantitative susceptibility mapping (QSM) is crucial for brain tissue characterization using MRI phase images.
- Current deep learning models for QSM often overlook the impact of varying acquisition parameters like resolution and orientation.
- Limited training data covering diverse acquisition parameters hinders the generalizability of existing QSM deep learning approaches.
Purpose of the Study:
- To develop a novel deep learning approach, adaptive convolution, that integrates prior imaging parameter information into QSM.
- To enhance the generalizability of QSM deep learning models across different acquisition parameters.
- To improve the accuracy and robustness of susceptibility map computation.
Main Methods:
- Proposed an adaptive convolution method within a 3D U-Net architecture to learn the relationship between acquisition parameters and MRI phase images.
- Utilized a-priori information of imaging parameters to adaptively select convolution weights based on data attributes.
- Employed pre-training on synthetic data followed by transfer learning on clinical brain data.
Main Results:
- The adaptive convolution 3D U-Net demonstrated significant generalizability across varying acquisition parameters on both synthetic and in-vivo data.
- The proposed method outperformed standard models lacking adaptive convolution or transfer learning.
- Experiments confirmed the positive impact of incorporating side information (acquisition parameters) on model performance and susceptibility map accuracy.
Conclusions:
- Adaptive convolution effectively integrates imaging parameter information into deep learning for QSM, enhancing model generalizability.
- Transfer learning from synthetic to clinical data significantly improves QSM computation.
- The developed approach offers a more robust and adaptable solution for quantitative susceptibility mapping in diverse clinical settings.
