MM-UNet: A multimodality brain tumor segmentation network in MRI images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Brain tumors represent a significant global health challenge with high mortality rates, particularly in children.
• Accurate brain tumor segmentation is crucial for diagnosis and treatment planning.
• Traditional manual segmentation is inefficient and subjective, while single-image modalities offer limited diagnostic information.

Purpose of the Study:

• To develop an advanced deep learning model for accurate and efficient brain tumor segmentation using multimodality imaging.
• To introduce the MM-UNet, a novel network architecture designed for fusing features from multiple imaging sources.

Main Methods:

• Developed a multimodality feature fusion network (MM-UNet) with a multi-encoder, single-decoder structure.
• Utilized hybrid attention blocks to enhance feature extraction and fusion across different imaging modalities.
• Evaluated the model's performance on the BraTS 2020 dataset for brain tumor segmentation.

Main Results:

• The MM-UNet achieved a mean Dice score of 79.2% and a mean Hausdorff distance of 8.466.
• Demonstrated consistent performance improvements compared to baseline models like U-Net, Attention U-Net, and ResUNet.
• Validated the effectiveness of multimodality feature fusion for enhanced brain tumor segmentation.

Conclusions:

• The proposed MM-UNet effectively segments brain tumors by integrating information from multiple imaging modalities.
• This advanced deep learning approach offers a significant improvement over existing methods for clinical applications.
• MM-UNet shows promise for improving diagnostic accuracy and treatment planning in neuro-oncology.