Robust brain ROI segmentation by deformation regression and deformable shape model

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate segmentation of brain regions of interest (ROIs) is crucial for neurological studies.
• Conventional deformable models often struggle with local errors and initialization sensitivity.
• Existing multi-atlas methods can produce fragmented segmentations.

Purpose of the Study:

• To develop a robust and efficient learning-based deformable model for segmenting ROIs in structural MR brain images.
• To improve segmentation accuracy and computational efficiency compared to state-of-the-art methods.
• To overcome limitations of conventional deformable models and multi-atlas approaches.

Main Methods:

• A novel learning-based deformable model guided by an image-based regressor predicting voxel displacement vectors towards the ROI boundary.
• A joint classification and regression random forest for multi-task learning of the regressor and ROI classifier.
• Incorporation of a prior shape model to prevent isolated segmentations and high-level context features for improved prediction consistency.

Main Results:

• The proposed method demonstrated superior segmentation accuracy and computational efficiency across three public brain MR datasets.
• Outperformed state-of-the-art multi-atlas-based and other learning-based segmentation methods.
• The model showed robustness against local prediction errors and initialization variations.

Conclusions:

• The learning-based deformable model offers a significant advancement in automated brain ROI segmentation from MR images.
• The method provides a more accurate, efficient, and robust alternative to existing segmentation techniques.
• Future work could explore further refinements in feature extraction and model architecture for enhanced performance.