Enhanced U-Net for Infant Brain MRI Segmentation: A (2+1)D Convolutional Approach.

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Infant brain MRI segmentation is crucial but challenging due to evolving tissue contrasts.
• Gray matter (GM) and white matter (WM) intensity convergence complicates accurate segmentation.
• Automating segmentation of cerebrospinal fluid (CSF), GM, and WM in infants is vital for developmental studies.

Purpose of the Study:

• To develop an enhanced U-net model for precise automatic segmentation of infant brain tissues.
• To improve the accuracy of segmenting CSF, GM, and WM in infant brain MRIs.
• To evaluate the model's performance on the iSeg-2017 dataset.

Main Methods:

• Utilized a U-net architecture with (2+1)D convolutional layers and skip connections.
• Applied intensity normalization via histogram alignment for MRI data standardization.
• Trained and evaluated the model on T1-weighted and T2-weighted MRI data from ten infant subjects using cross-validation.

Main Results:

• Achieved an average segmentation accuracy of 92.2%, a 0.7% improvement over previous methods.
• Demonstrated high performance metrics including sensitivity, precision, and Dice similarity scores.
• Identified a slight bias in misclassifying GM and WM, indicating areas for future refinement.

Conclusions:

• The U-net architecture is highly effective for infant brain tissue segmentation from MRI.
• Future research will focus on attention mechanisms and dual-network processing to further enhance accuracy.
• The developed model shows promise for advancing pediatric neuroimaging analysis.