Robust Multiple Sclerosis Lesion Inpainting with Edge Prior
Huahong Zhang1, Rohit Bakshi2, Francesca Bagnato3
1Vanderbilt University, Nashville, TN 37235, USA.
Machine Learning in Medical Imaging. MLMI (Workshop)
|December 24, 2021
Summary
This study introduces a novel deep learning method for brain MRI inpainting in multiple sclerosis (MS) patients. The approach uses edge information to realistically fill in lesion areas, improving downstream analysis and segmentation accuracy.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuroscience
Background:
- Inpainting brain lesions in multiple sclerosis (MS) MRIs is crucial for accurate analysis.
- Existing methods often disregard lesion area information, potentially impacting results.
- Accurate preprocessing is vital for tasks like tissue segmentation and surface reconstruction.
Purpose of the Study:
- To develop a novel deep learning approach for brain MRI inpainting.
- To leverage lesion edge information as a prior for improved inpainting.
- To enhance the accuracy of subsequent neuroimaging analyses.
Main Methods:
- A deep learning network utilizing T1-weighted MRI, lesion masks, and edge maps.
- Incorporation of edge information around lesions to guide the inpainting process.
- Preservation of local tissue structure (white matter/grey matter) using edge priors.
Main Results:
- Qualitative results demonstrate realistic and shape-consistent lesion inpainting.
- Quantitative evaluation shows superior performance compared to state-of-the-art methods.
- Improved FreeSurfer segmentation accuracy using the proposed inpainting technique.
Conclusions:
- The proposed edge-guided deep learning method effectively inpaints brain lesions in MS MRIs.
- The approach achieves state-of-the-art performance and enhances downstream analysis accuracy.
- The method exhibits robustness to imprecise lesion masks, increasing practical utility.
