3D-MRI brain glioma intelligent segmentation based on improved 3D U-net network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-Oncology

Background:

• Accurate glioma segmentation in 3D-MRI is crucial for clinical decision-making.
• Existing methods face challenges in handling complex tumor structures and variations.

Purpose of the Study:

• To develop and validate an intelligent deep learning-based method for enhanced 3D-MRI glioma segmentation.
• To provide guidance for medical diagnosis, grading, and treatment strategy selection.

Main Methods:

• Utilized the BraTS2023 dataset with preprocessing including 3D clipping, resampling, artifact elimination, and normalization.
• Introduced a space pyramid pool module for multi-scale feature perception and a multi-scale fusion attention mechanism for focused segmentation.
• Employed a combined Dice and Focal loss function to address class imbalance and improve learning of challenging voxels.

Main Results:

• The enhanced 3D U-Net model achieved stable training loss and demonstrated superior segmentation performance.
• Achieved high Dice Similarity Coefficient (DSC), Recall, and Precision scores across Whole Tumor (WT), Core Tumor (TC), and Enhanced Tumor (ET) segments.
• Specific scores include: WT (DSC 0.9168, Recall 0.9426, Precision 0.9375), TC (DSC 0.8954, Recall 0.9014, Precision 0.9369), and ET (DSC 0.8674, Recall 0.9045, Precision 0.9011).

Conclusions:

• The proposed method significantly enhances glioma segmentation accuracy and generalization.
• The high performance indices in WT, TC, and ET segments offer a reliable foundation for clinical diagnosis and treatment planning.
• This advancement improves the accuracy and reliability of glioma diagnoses.