Cascaded convolutional networks for unsupervised brain tissue segmentation and bias field estimation

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence

Background:

• Accurate brain tissue segmentation is vital for neuroimaging studies.
• Supervised deep learning (DL) methods are limited by the need for manually labeled datasets.
• Leveraging large unlabeled neuroimaging datasets is crucial for advancing segmentation techniques.

Purpose of the Study:

• To develop an unsupervised deep learning model for joint brain tissue segmentation and bias field estimation.
• To overcome the limitations of supervised learning in brain MRI segmentation.
• To utilize abundant unlabeled brain imaging data effectively.

Main Methods:

• A novel unsupervised deep learning model employing cascaded convolutional networks.
• Recursive bias field estimation and correction using cascaded modules.
• A segmentation module utilizing a Gaussian mixture model for intensity statistics and model fitting error as the unsupervised loss function.

Main Results:

• The unsupervised DL model achieved competitive performance in both bias field correction and brain tissue segmentation.
• Quantitative evaluations on HCP-Aging and HCP-Development datasets demonstrated the model's efficacy.
• The method showed comparable results to state-of-the-art bias field correction and unsupervised segmentation techniques.

Conclusions:

• The proposed unsupervised DL approach offers a viable solution for brain tissue segmentation without manual labels.
• This method effectively addresses bias field correction and segmentation simultaneously in neuroimaging.
• The study highlights the potential of unsupervised learning for large-scale brain imaging data analysis.