Leveraging SO(3)-steerable convolutions for pose-robust semantic segmentation in 3D medical data.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) are widely used in medical image analysis.
• Standard CNNs lack rotational equivariance, limiting performance and requiring extensive data augmentation.
• Existing segmentation networks often rely on standard convolutional kernels, despite the benefits of rotationally-equivariant layers.

Purpose of the Study:

• To introduce a novel family of segmentation networks utilizing equivariant voxel convolutions based on spherical harmonics.
• To enhance robustness to unseen data poses and reduce the need for rotation-based data augmentation.
• To improve segmentation performance in medical imaging tasks, specifically for MRI brain tumor and healthy brain structure segmentation.

Main Methods:

• Development of a new family of segmentation networks employing equivariant voxel convolutions.
• Utilizing spherical harmonics to achieve SO(3)-steerable kernels for improved parameter sharing and translational equivariance.
• Implementation of these networks for MRI brain tumor and healthy brain structure segmentation tasks.

Main Results:

• Demonstrated improved segmentation performance compared to standard convolutional networks.
• Showcased enhanced robustness to reduced amounts of training data.
• Achieved improved parameter efficiency and reduced need for rotation-based data augmentation.

Conclusions:

• Equivariant voxel convolutions offer significant advantages for medical image segmentation.
• These networks provide increased robustness, better sample efficiency, and smaller network sizes.
• The proposed networks represent a promising advancement for medical image analysis, particularly in challenging segmentation tasks.