A cascaded nested network for 3T brain MR image segmentation guided by 7T labeling.

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Accurate brain segmentation is vital for analyzing brain anatomy in magnetic resonance (MR) imaging.
• Higher tissue contrast in 7T MR images aids in precise delineation for training segmentation models.

Purpose of the Study:

• To develop a cascaded nested network (CaNes-Net) for segmenting 3T brain MR images.
• To leverage 7T MR image labels for training the segmentation model and address 3T/7T image misalignment.

Main Methods:

• Proposed a cascaded nested network (CaNes-Net) comprising iterated nested network (Nes-Net) modules.
• Utilized geodesic distance maps for refining segmentation and a correlation coefficient map to mitigate 3T/7T image misalignment.
• Trained the model using tissue labels from 7T images for segmenting 3T brain MR images.

Main Results:

• CaNes-Net demonstrated substantial improvements in segmentation accuracy compared to SPM, FSL, and other deep learning models.
• The proposed method effectively reduced segmentation errors stemming from 3T/7T image misalignment.
• Evaluated on 18 adult subjects and the ADNI dataset, confirming robust performance.

Conclusions:

• CaNes-Net offers a significant advancement in brain MR image segmentation accuracy.
• The approach effectively addresses challenges associated with multi-modal MR image registration for deep learning.
• This method holds promise for enhanced visualization and quantification of brain anatomy.