Auto-Context Convolutional Neural Network (Auto-Net) for Brain Extraction in Magnetic Resonance Imaging

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Accurate brain extraction is critical for neuroimage analysis pipelines.
• Current methods often fail due to reliance on image registration and geometric assumptions.
• There is a need for robust, learning-based, and registration-free brain extraction tools.

Purpose of the Study:

• To develop an accurate, learning-based, geometry-independent, and registration-free brain extraction tool.
• To present a novel auto-context convolutional neural network (CNN) for brain segmentation.
• To evaluate the performance of the proposed method on benchmark datasets and challenging applications.

Main Methods:

• Utilized an auto-context convolutional neural network (CNN) learning intrinsic local and global image features through 2-D patches.
• Explored two architectures: a voxelwise approach with parallel 2-D pathways and a U-net based fully convolutional network.
• Iteratively used posterior probability maps as context information with original image patches for improved brain extraction.

Main Results:

• Achieved superior Dice overlap coefficients on LPBA40 (97.73%) and OASIS (97.62%) datasets compared to state-of-the-art methods.
• Demonstrated significant improvement with the auto-context algorithm.
• Outperformed other methods in extracting arbitrarily oriented fetal brains from MRI data (Dice coefficient: 95.97%).

Conclusions:

• The proposed auto-context CNN provides accurate and robust brain extraction, outperforming existing techniques.
• The method is geometry-independent and registration-free, offering advantages in various neuroimaging applications.
• This approach can mitigate challenges associated with image registration in segmentation tasks, particularly in complex cases like fetal brain MRI.