Deep learning-based magnetic resonance image segmentation technique for application to glioma

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Brain glioma segmentation is crucial for diagnosis and treatment.
• Deep learning models offer potential but struggle with segmentation stability and integrating patient-specific data.
• Existing methods often overlook unique genomic and clinical information, impacting diagnostic accuracy.

Purpose of the Study:

• To develop a more stable and accurate brain glioma segmentation model.
• To integrate multimodal patient data, including genomic and basic information, into the segmentation process.
• To address the limitations of current deep learning approaches in handling sample imbalance and noise-like segmentation errors.

Main Methods:

• Utilized DeepLabv3+ architecture with RegNet as the image encoder.
• Developed an attribute encoder module for multimodal fusion of image and patient data.
• Implemented a weighted cross-entropy and Dice loss function, alongside a novel loss to suppress specific region sizes for enhanced stability.

Main Results:

• Achieved a Dice score of 94.36% and an Intersection over Union (IoU) score of 91.83% on the Lower-Grade Glioma Segmentation Dataset.
• Demonstrated superior performance compared to other popular segmentation models.
• The proposed model yielded higher-stability segmentation results, reducing errors in noise-like regions.

Conclusions:

• The novel model significantly enhances brain glioma segmentation accuracy and stability.
• Integrating multimodal patient data improves diagnostic and treatment planning capabilities.
• The proposed approach represents a substantial advancement in medical image analysis for neuro-oncology.