Cascade Residual Multiscale Convolution and Mamba-Structured UNet for Advanced Brain Tumor Image Segmentation

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Computational Neuroscience

Background:

• Precise brain tumor segmentation is vital for diagnosis and treatment planning.
• Convolutional Neural Networks (CNNs) and transformer models have limitations in processing sequential data and computational efficiency for large datasets.
• Existing methods struggle to balance accuracy with computational demands in medical image segmentation.

Purpose of the Study:

• To introduce MambaBTS, a novel model for enhanced brain tumor segmentation in MRI.
• To synergize the strengths of CNNs and transformers, inspired by the Mamba architecture.
• To improve computational efficiency and segmentation accuracy compared to existing models.

Main Methods:

• Developed MambaBTS, integrating CNNs, transformers, and Mamba-inspired architecture with cascade residual multi-scale convolutional kernels.
• Employed a mixed loss function combining dice loss and cross-entropy for refined segmentation.
• Evaluated performance on the MICCAI BraTS 2019 dataset for brain tumor segmentation.

Main Results:

• MambaBTS achieved high dice coefficients: 0.8450 (Whole Tumor), 0.8606 (Tumor Core), and 0.7796 (Enhancing Tumor).
• The model demonstrated superior accuracy, computational efficiency, and parameter efficiency over existing methods.
• Reduced computational complexity and enhanced the receptive field for improved segmentation.

Conclusions:

• MambaBTS offers a balanced, efficient, and effective solution for brain tumor segmentation in MRI.
• The model overcomes limitations of traditional CNNs and transformers in medical imaging.
• Results indicate significant potential for improving clinical diagnostics and treatment planning.