Adaptive Feature Recombination and Recalibration for Semantic Segmentation With Fully Convolutional Networks

Area of Science:

• Computer Vision
• Medical Image Analysis
• Deep Learning

Background:

• Fully convolutional networks (FCNs) excel in image semantic segmentation due to efficient voxel processing.
• Existing channel recalibration methods (e.g., Squeeze-and-Excitation) are not optimal for FCNs in segmentation tasks.

Purpose of the Study:

• To propose a novel spatially adaptive recalibration block (SegSE) for FCNs in semantic segmentation.
• To enhance feature map discriminative power by considering both cross-channel and spatial information.

Main Methods:

• Developed the SegSE block, integrating feature recombination and spatially adaptive channel recalibration.
• Applied the SegSE block to fully convolutional networks for semantic segmentation.

Main Results:

• The SegSE block significantly improved performance on a competitive baseline.
• The method generalized well across diverse medical imaging applications.

Conclusions:

• The SegSE block effectively enhances semantic segmentation in FCNs by incorporating spatial adaptivity.
• This approach achieves state-of-the-art or competitive results in brain tumor segmentation, stroke penumbra estimation, and ischemic stroke lesion outcome prediction.