FULLY CONVOLUTIONAL NETWORKS FOR MULTI-MODALITY ISOINTENSE INFANT BRAIN IMAGE SEGMENTATION.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Developmental Neuroscience

Background:

• Infant brain tissue segmentation (white matter, gray matter, cerebrospinal fluid) is vital for understanding early brain development.
• The isointense phase (6-8 months) presents segmentation challenges due to low contrast between white and gray matter in T1 and T2 MRI scans.
• Existing methods often fail to fully leverage multi-modal information, limiting segmentation performance.

Purpose of the Study:

• To develop an improved method for segmenting infant brain MR images during the challenging isointense phase.
• To effectively integrate multi-modal information (T1, T2, FA) for enhanced segmentation accuracy.
• To address the limitations of existing methods that do not fully exploit multi-modality.

Main Methods:

• Proposed a novel approach using fully convolutional networks (FCNs) for isointense phase brain MR image segmentation.
• Trained separate FCNs for each modality (T1, T2, FA) and fused their high-level features for final segmentation.
• Implemented a convolution-pooling stream to process multi-modal information separately before high-level feature fusion.

Main Results:

• The proposed FCN-based multi-modal fusion method significantly outperformed existing segmentation techniques.
• Achieved higher accuracy in segmenting infant brain tissues during the isointense phase.
• Demonstrated the effectiveness of the proposed strategy for integrating multi-modality images.

Conclusions:

• The developed FCN approach offers a superior method for infant brain tissue segmentation in the isointense phase.
• Effective fusion of multi-modal MRI data significantly enhances segmentation accuracy.
• This work provides a promising direction for improving the analysis of early brain development using neuroimaging.