Adaptive cascaded transformer U-Net for MRI brain tumor segmentation

Area of Science:

• Medical imaging analysis
• Artificial intelligence in oncology
• Neuroscience

Background:

• Brain tumor segmentation on MRI is crucial for cancer diagnosis and treatment.
• Cascaded U-Net models show promise but struggle with diverse tumor characteristics.
• Conventional convolutions are limited in addressing global and local variations in brain tumors.

Purpose of the Study:

• To propose a novel Adaptive Cascaded Transformer U-Net (ACTransU-Net) for enhanced MRI brain tumor segmentation.
• To improve the adaptive capture of global and local tumor information.
• To overcome limitations of conventional convolutions in segmenting varied brain tumors.

Main Methods:

• Developed a two-stage cascaded 3D U-Net architecture for coarse-to-fine segmentation.
• Integrated omni-dimensional dynamic convolution modules to enhance local detail representation.
• Incorporated 3D Swin-Transformer modules to capture long-range dependencies and global context.

Main Results:

• ACTransU-Net demonstrated effectiveness on BraTS 2020 and BraTS 2021 datasets.
• Achieved competitive performance with state-of-the-art methods.
• Reported average Dice Similarity Coefficient (DSC) of 84.96% and 91.37%, and Hausdorff Distance 95th percentile (HD95) of 10.81 mm and 7.31 mm.

Conclusions:

• ACTransU-Net adaptively captures global and local brain tumor features for improved segmentation.
• The method aids physicians in accurate diagnosis and has potential for segmenting other lesions.
• The developed ACTransU-Net offers a significant advancement in medical image analysis for oncology.