PSACNN: Pulse sequence adaptive fast whole brain segmentation.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence

Background:

• Supervised learning methods, particularly convolutional neural networks (CNNs), are widely used for brain segmentation.
• Training these models requires large, manually annotated datasets, which are difficult to obtain.
• Existing CNNs struggle to generalize across different magnetic resonance imaging (MRI) acquisition protocols, limiting their use in multi-center studies.

Purpose of the Study:

• To develop a CNN-based brain segmentation algorithm that is accurate, fast, and resilient to variations in MRI acquisition protocols.
• To address the challenge of limited manually annotated data and protocol heterogeneity in neuroimaging datasets.

Main Methods:

• Proposed a CNN-based segmentation algorithm utilizing approximate forward models of pulse sequences to generate synthetic training data.
• Augmented the training dataset by sampling over a wide variety of pulse sequences to create an image contrast-invariant model.
• Trained a single CNN capable of segmenting T1-weighted and T2-weighted MRI images using only T1-weighted training data.

Main Results:

• Achieved highly accurate brain segmentations with state-of-the-art performance (overall Dice overlap = 0.94).
• Demonstrated a fast run time of approximately 45 seconds for segmentation.
• Showcased consistent segmentation performance across a wide range of MRI acquisition protocols.

Conclusions:

• The proposed CNN-based approach effectively overcomes the limitations of traditional supervised learning for brain segmentation.
• The method's resilience to varying acquisition protocols makes it suitable for multi-center neuroimaging studies and clinical applications.
• This technique enables accurate and efficient brain segmentation without the need for extensive manual annotation or protocol standardization.