Chenchao Xu

Cerebral cortex segmentation from three-dimensional structural Magnetic Resonance (MR) brain images plays an important role in measuring loss of cortical tissues for disorders such as Alzheimer's disease (AD). U-Net, a type of deep convolutional neural networks architecture, is a widely-used approach for biomedical image segmentation in recent years.

In this thesis, I implemented 2D/3D U-Nets on MR images from 20 patients with labeled cerebral tissues and regions. A two-stage pipeline was designed for this task. In stage one, U-Net aims to generate a mask of grey matter to filter out other tissues in brain MRI images. In stage two, a similar U-Net architecture is used to label cerebral cortex sub-regions from images which only contains grey matter.
Both 2D U-Net and 3D U-Net do not work for labeling gyri/sulci, and only achieve approximate $55%$ Dice overlap for labeling cortex regions. In contrast, the cerebral cortex segmentation package in FreeSurfer achieves over $90%$ Dice overlap for labeling gyri/sulci by using a graphical-based probabilistic estimation method with prior information.

I believe that the main reason of bad performance of 2D/3D U-Net is the loss of global position information of pixels/voxels by cutting original MR images into small parts. The U-net architecture has weakness of handling high resolution 3D images with imbalanced number of classes. In the feature work, researchers could create hybrid methods to combine deep neural networks architectures with prior information to label cerebral cortical sub-regions.