3-D Fully Convolutional Networks for Multimodal Isointense Infant Brain Image Segmentation.

Area of Science:

• Medical Imaging
• Neuroscience
• Computer Vision

Background:

• Accurate infant brain segmentation is crucial for studying early development.
• The isointense phase (6-8 months) presents low tissue contrast in MR images, challenging segmentation.
• Existing methods often rely on single-modality, patch-based approaches.

Purpose of the Study:

• To develop a novel 3-D multimodal fully convolutional network (FCN) for improved infant brain MR image segmentation.
• To address the segmentation challenges in the isointense phase.
• To enhance segmentation accuracy and efficiency.

Main Methods:

• Proposed a 3-D multimodal FCN architecture, extending 2-D FCNs.
• Integrated coarse and dense feature maps to model small tissue regions.
• Introduced transformation and fusion modules for better feature map connection and integration.
• Utilized multimodal information for enhanced segmentation.

Main Results:

• The proposed 3-D multimodal FCN significantly outperformed baseline and state-of-the-art methods in segmentation accuracy.
• Achieved faster segmentation results compared to existing methods.
• Demonstrated that integrating feature maps, using batch normalization, and multimodal information boost performance.

Conclusions:

• The novel 3-D multimodal FCN is highly effective for segmenting isointense phase infant brain MR images.
• The proposed architectural components and multimodal integration contribute to superior performance.
• This framework offers a promising solution for robust and efficient infant brain analysis.