Self-supervised pre-training based hybrid network for deep gray matter nuclei segmentation.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Medicine
• Neuroimaging

Background:

• Accurate segmentation of deep gray matter nuclei is vital for diagnosing and treating neurological disorders.
• Supervised learning methods require extensive labeled datasets, which are difficult and time-consuming to acquire for medical imaging.
• Convolutional Neural Networks (CNNs) have limitations in capturing long-range dependencies crucial for complex segmentation tasks.

Purpose of the Study:

• To develop an improved method for deep gray matter nuclei segmentation using a novel hybrid network.
• To leverage self-supervised learning to reduce the reliance on large labeled datasets.
• To enhance segmentation performance by combining local and global feature extraction.

Main Methods:

• Proposed a CNN-Transformer hybrid network (CTNet) integrating 3D CNN for local features and Vision Transformer (ViT) for global context.
• Implemented a self-supervised learning (SSL) strategy involving rotation prediction and masked feature reconstruction for pre-training CTNet on unlabeled data.
• Evaluated the method on 3T and 7T human brain MRI datasets.

Main Results:

• The proposed CTNet demonstrated superior segmentation performance compared to existing models.
• The self-supervised pre-training strategy outperformed other advanced SSL methods.
• Even with only one labeled sample, the pre-trained CTNet showed an 8.4% improvement in Dice Similarity Coefficient (DSC) over a randomly initialized network.

Conclusions:

• The developed CTNet, pre-trained using a novel SSL approach, offers a robust and efficient solution for deep gray matter nuclei segmentation.
• This method significantly enhances segmentation accuracy, particularly in low-data regimes, addressing a key challenge in medical image analysis.
• The hybrid architecture effectively captures both local and global information, leading to state-of-the-art performance in neuroimaging segmentation.