Multi-Modal Segmentation of 3D Brain Scans Using Neural Networks.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Anatomical segmentation of brain scans is crucial for diagnostics and neuroradiology research.
• Conventionally, segmentation relies on T1-weighted MRI due to its soft-tissue contrast.
• Investigating alternative imaging modalities is essential for broader applicability.

Purpose of the Study:

• To compare automated brain segmentation performance across different MRI contrasts and CT scans.
• To evaluate the anatomical soft-tissue information in various imaging modalities.
• To assess the utility of learning-based segmentation beyond T1-weighted MRI.

Main Methods:

• Training convolutional neural networks (CNNs) on a large database of 853 MRI/CT brain scans.
• Benchmarking CNN performance on T1-weighted MRI, FLAIR MRI, DWI MRI, and CT scans.
• Assessing segmentation accuracy for 27 anatomical substructures relative to FreeSurfer labels.

Main Results:

• Average Dice scores achieved were: 86.7% (T1-weighted MRI), 81.9% (FLAIR MRI), 80.8% (DWI MRI), and 80.7% (CT).
• The segmentation pipeline incorporates dropout sampling for quality control of input scans and segmentations.
• Full 3D volume segmentation (<1 second on GPU) is highly efficient.

Conclusions:

• Learning-based automated brain segmentation is feasible and effective across multiple MRI contrasts and CT scans.
• The study highlights the potential of CNNs for robust anatomical segmentation in diverse neuroimaging datasets.
• The developed pipeline offers efficient and quality-controlled segmentation for neuroradiology research.