High-precision brain tumor segmentation with switchable normalization in faster R-CNN architecture.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Brain tumor segmentation from multi-modal MRI is complex due to tumor heterogeneity and varying imaging protocols.
• Accurate segmentation is crucial for diagnosis, treatment planning, and monitoring.

Purpose of the Study:

• To propose a novel Switchable Normalization-based Faster R-CNN (SNFRC) framework for automated brain tumor segmentation.
• To enhance feature consistency and accurately identify irregular tumor regions in multi-modal MRI scans.

Main Methods:

• Developed a SNFRC framework utilizing a region proposal network (RPN) with switchable normalization (SN).
• Employed a detection-based segmentation strategy for explicit tumor localization and pixel-wise mask generation.
• Introduced a composite loss function incorporating Dice loss, L2 loss, and Kullback-Leibler divergence for joint optimization.

Main Results:

• The SNFRC framework achieved Dice scores above 93% and reduced Hausdorff distance by approximately 1.5 pixels.
• Switchable normalization demonstrated significant performance improvements (6-7%) over regular normalization.
• The model exhibited faster and more stable training convergence compared to the baseline Faster R-CNN.

Conclusions:

• The proposed SNFRC framework offers a reliable, efficient, and computationally effective approach for automated brain tumor segmentation in multi-modal MRI.
• The method shows promise for real-time applications, though further research on cross-dataset generalization and clinical validation is warranted.