Feature interaction network based on hierarchical decoupled convolution for 3D medical image segmentation.

Area of Science:

• Medical imaging
• Artificial intelligence
• Neuroscience

Background:

• Manual segmentation of multimodal brain tumors is time-consuming and challenging.
• Accurate segmentation is crucial for clinical treatment decisions and surgical planning.
• Deep learning faces challenges in medical image segmentation due to tumor diversity and limited computational resources.

Purpose of the Study:

• To develop an automatic and accurate method for segmenting multimodal brain tumors.
• To improve the performance of neural network segmentation using a novel feature fusion module and attention mechanism.
• To address the category imbalance problem in medical image segmentation.

Main Methods:

• Proposed a feature fusion module based on a hierarchical decoupling convolution network and an attention mechanism.
• Replaced U-shaped network skip connections with the feature fusion module.
• Introduced a global attention mechanism to integrate encoder features and explore context information.

Main Results:

• Achieved Dice similarity coefficient (DSC) of 0.775 (enhance tumor), 0.900 (whole tumor), and 0.827 (tumor core) on the BraTS 2019 dataset.
• Achieved DSC of 0.800 (enhance tumor), 0.902 (whole tumor), and 0.841 (tumor core) on the BraTS 2018 dataset.
• Demonstrated the method's generality and effectiveness for brain tumor image studies.

Conclusions:

• The proposed method offers a powerful tool for brain tumor image analysis.
• The feature fusion module and attention mechanism effectively improve segmentation accuracy.
• The approach provides a general solution for complex medical image segmentation tasks.