Supervised Domain Adaptation for Automatic Sub-cortical Brain Structure Segmentation with Minimal User Interaction.

Area of Science:

• Medical Imaging
• Neuroimaging
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) are used for sub-cortical brain structure segmentation in MRI.
• Existing CNN methods face reproducibility issues due to domain shifts (protocol, scanner, intensity variations).
• Retraining CNNs for each new domain limits clinical applicability.

Purpose of the Study:

• To address the domain shift problem in MRI segmentation using transfer learning.
• To reduce the number of required training images and improve training speed.
• To enhance the clinical applicability of CNN-based segmentation methods.

Main Methods:

• Employed a transfer learning strategy to leverage knowledge from pretrained networks.
• Reduced the number of trainable parameters in the CNN to accelerate training.
• Tested the method on MICCAI 2012 and IBSR datasets, comparing it with the FIRST approach.

Main Results:

• Achieved segmentation results comparable to fully trained CNNs.
• Required significantly fewer training images from the target domain.
• Outperformed the FIRST approach with minimal training data (1 image for MICCAI 2012, 3 for IBSR).

Conclusions:

• Transfer learning effectively solves the domain shift problem in MRI segmentation.
• The proposed method enhances efficiency and reduces data requirements for CNN-based segmentation.
• This approach holds promise for improving the clinical utility of automated brain MRI analysis.