Mod-SE(2): a geometric deep learning framework for brain tumor classification and segmentation in MRI images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Accurate brain tumor classification and segmentation are crucial for patient diagnosis and treatment.
• Heterogeneous tumor morphology presents significant challenges for traditional Convolutional Neural Networks (CNNs).
• CNNs often lack rotational and translational invariance, limiting their performance on diverse MRI data.

Purpose of the Study:

• To introduce a novel geometric deep learning framework, Modified Special Euclidean (Mod-SE(2)), for enhanced brain tumor analysis.
• To improve spatial consistency and reduce data augmentation dependency in tumor classification and segmentation.
• To leverage geometric priors and symmetry-preserving group convolutions for robust medical image analysis.

Main Methods:

• Developed the Mod-SE(2) framework integrating geometric priors and symmetry-preserving group convolutions.
• Applied Mod-SE(2) to tumor classification (Mod-Cls-SE(2)) and segmentation (Mod-Seg-SE(2)) tasks.
• Evaluated performance on multiple MRI and medical image datasets, benchmarking against U-Net, NN U-Net, VGG16, VGG19, and ResNet.

Main Results:

• Mod-Cls-SE(2) achieved an average classification accuracy of 0.914, surpassing ResNet101 (0.682) and VGG16 (0.705).
• Mod-Seg-SE(2) attained a Dice coefficient of 0.9503 and IoU of 0.9616 on BraTS2020, outperforming U-Net (0.797) and NN U-Net (0.815).
• The model demonstrated reduced inference time and strong computational performance.

Conclusions:

• Mod-SE(2) enhances spatial consistency, efficiency, and interpretability in brain tumor analysis through its symmetry-aware design.
• The framework generalizes better across varying tumor shapes, outperforming traditional methods in key metrics.
• Mod-SE(2) supports precise boundary delineation for neurosurgical planning and other clinical applications.