A Deep Spatial Context Guided Framework for Infant Brain Subcortical Segmentation.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence

Background:

• Accurate subcortical segmentation in infant brain MRI is vital for understanding early development and diagnosing diseases.
• Challenges include dynamic intensity variations, low tissue contrast, and small anatomical structures in infant brains.

Purpose of the Study:

• To develop an accurate and robust deep convolutional neural network (CNN) framework for subcortical segmentation in infant brain MR images.
• To leverage spatial context information and a coarse-to-fine strategy to overcome segmentation challenges.

Main Methods:

• A spatial context-guided, coarse-to-fine CNN framework was proposed, utilizing multi-modal MR images (T1w, T2w, T1w/T2w).
• A Signed Distance Map learning UNet (SDM-UNet) was employed at the coarse stage to predict SDMs, incorporating spatial and shape information, with a Correntropy loss for robustness.
• A Multi-Source and Multi-Path UNet (M2-UNet) was used at the fine stage, integrating predicted SDMs and multi-modal images for refined segmentation.

Main Results:

• The proposed method demonstrated superior performance in both qualitative and quantitative evaluations compared to four state-of-the-art methods.
• Consistent improvements in segmentation accuracy were observed on an infant brain MR image dataset.

Conclusions:

• The developed coarse-to-fine CNN framework effectively addresses the challenges of infant brain subcortical segmentation.
• This approach enhances the accuracy of subcortical segmentation, supporting research in early brain development and disease diagnosis.