Unsupervised Deep Learning for Bayesian Brain MRI Segmentation.

Area of Science:

• Medical Imaging
• Neuroimaging Analysis
• Artificial Intelligence in Medicine

Background:

• Probabilistic atlas priors are crucial for robust brain MRI segmentation but are computationally intensive.
• Deep learning offers computationally efficient segmentation but typically requires extensive manual annotations.
• Training deep learning models for new MRI datasets necessitates costly manual labeling or suboptimal adaptation strategies.

Purpose of the Study:

• To develop a deep learning-based brain MRI segmentation method that does not require manually segmented images for training.
• To enable efficient adaptation of segmentation models to new MRI datasets with varying contrasts.
• To reduce the computational cost and labor associated with brain MRI segmentation.

Main Methods:

• A hybrid approach combining conventional probabilistic atlas-based segmentation with deep learning was proposed.
• The method trains a segmentation model without relying on manually annotated images.
• Experiments were conducted on thousands of brain MRI scans with diverse contrasts.

Main Results:

• The proposed method successfully trained a deep learning segmentation model without manual annotations.
• The approach demonstrated good accuracy in segmenting brain MRI scans across different contrasts.
• The model achieved a test time of approximately 15 seconds on a GPU.

Conclusions:

• A novel deep learning strategy effectively integrates probabilistic atlases for MRI segmentation, eliminating the need for manual labels.
• This method offers an efficient and accurate solution for segmenting brain MRI scans, adaptable to new datasets and contrasts.
• The approach significantly reduces the cost and time associated with creating and adapting neuroimage analysis pipelines.