Two-Branch network for brain tumor segmentation using attention mechanism and super-resolution reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate brain tumor segmentation is crucial for MRI diagnosis and treatment monitoring.
• Challenges include inconsistent lesion degrees, structural variations, low contrast, and blur in brain MR images.
• Current deep learning algorithms struggle with precise segmentation due to these complexities.

Purpose of the Study:

• To develop a novel end-to-end brain tumor segmentation algorithm.
• To improve the accuracy and robustness of deep learning models for brain tumor segmentation.
• To integrate super-resolution image reconstruction with an improved 3D U-Net network.

Main Methods:

• Proposed a novel end-to-end framework integrating an improved 3D U-Net and super-resolution image reconstruction.
• Embedded a coordinate attention module before upsampling in the backbone network to enhance feature capture.
• Trained and evaluated the algorithm on BraTS datasets, comparing it with other deep learning methods using Dice similarity scores.

Main Results:

• Achieved a Dice similarity score of 89.61% for enhancing tumors on the BraTS2021 dataset.
• Obtained Dice scores of 88.30% for tumor cores and 91.05% for whole tumors.
• Demonstrated superior performance compared to the baseline 3D U-Net method, indicating robustness across varied brain tumor MR images.

Conclusions:

• The proposed integrated framework effectively addresses challenges in brain tumor segmentation.
• The coordinate attention mechanism enhances the capture of both local texture and global location features.
• The algorithm exhibits robust and high-performance segmentation for brain tumor MR images with considerable structural variations.