Multimodal Brain Tumor Segmentation Boosted by Monomodal Normal Brain Images

Area of Science:

• Medical imaging analysis
• Artificial intelligence in radiology
• Computational neuroscience

Background:

• Deep learning methods for brain tumor segmentation often overlook external information like normal brain appearance.
• Radiologists use normal brain appearance as a reference for identifying lesions.

Purpose of the Study:

• To propose a novel deep learning framework for brain tumor segmentation that incorporates normal brain images as a reference.
• To enhance tumor-related features by comparing tumorous and normal brain images in a learned feature space.

Main Methods:

• A novel deep framework for brain tumor segmentation is proposed, utilizing normal brain images as a reference.
• A feature alignment module (FAM) is introduced to address the challenge of comparing multimodal tumor images with monomodal normal images.
• The framework aligns feature distributions to enable effective comparison between normal and tumor regions.

Main Results:

• The proposed framework effectively improves brain tumor segmentation accuracy on both public (BraTS2022) and in-house datasets.
• Experimental results demonstrate superior performance compared to state-of-the-art segmentation methods.
• The method successfully highlights and enhances tumor-related features for more accurate segmentation.

Conclusions:

• The novel deep framework effectively leverages normal brain appearance to boost brain tumor segmentation performance.
• The feature alignment module is crucial for handling the multimodal vs. monomodal data discrepancy.
• This approach offers a promising advancement in automated brain tumor segmentation, outperforming existing methods.