mResU-Net: multi-scale residual U-Net-based brain tumor segmentation from multimodal MRI

Area of Science:

• Medical Image Processing
• Artificial Intelligence in Medicine
• Neuroimaging Analysis

Background:

• Accurate brain tumor segmentation is crucial for diagnosis and treatment planning.
• Existing methods face challenges in precisely delineating tumor subregions.
• Deep learning offers potential for improving segmentation accuracy in medical imaging.

Purpose of the Study:

• To develop an advanced end-to-end deep learning model for precise brain tumor segmentation.
• To enhance the accuracy of segmenting tumor core (TC), whole tumor (WT), and enhanced tumor (ET) regions.
• To address limitations in current brain tumor segmentation techniques.

Main Methods:

• Proposed a novel multi-scale deep residual convolutional neural network (mResU-Net).
• Utilized U-Net architecture with skip connections to bridge encoder-decoder semantic gaps.
• Incorporated residual structures to mitigate vanishing gradients and multi-scale convolution kernels for improved target detection.
• Integrated channel attention modules to further boost segmentation accuracy.

Main Results:

• Achieved high average Dice scores on the BraTS 2021 dataset: 0.9289 (TC), 0.9277 (WT), and 0.8965 (ET).
• Demonstrated significant improvements in segmenting brain tumor subregions compared to existing methods.
• The mResU-Net model showed robust performance in complex segmentation tasks.

Conclusions:

• The proposed mResU-Net model offers a significant advancement in automated brain tumor segmentation.
• The integration of multi-scale kernels and attention mechanisms enhances segmentation precision.
• mResU-Net shows great potential for clinical application in neuro-oncology.