Beyond CNNs: Exploiting Further Inherent Symmetries in Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Computer-Aided Diagnosis
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) excel in medical image segmentation but overlook inherent image symmetries (rotations, reflections).
• This limitation hinders precise tumor segmentation and accurate computer-aided diagnosis.
• Existing CNNs primarily leverage translation invariance, failing to exploit other crucial symmetries.

Purpose of the Study:

• To develop a novel group equivariant segmentation framework that encodes inherent image symmetries.
• To improve the precision of medical image representations by learning from symmetries.
• To enhance the performance of automatic tumor and lesion segmentation.

Main Methods:

• Proposed a group equivariant segmentation framework incorporating kernel-based equivariant operations for orientation-specific learning.
• Designed distinctive group layers with layer-wise symmetry constraints for global equivariance.
• Developed a group equivariant Res-UNet (GER-UNet) model.

Main Results:

• GER-UNet significantly outperformed regular CNN-based counterparts and state-of-the-art methods in hepatic tumor segmentation, COVID-19 lung infection segmentation, and retinal vessel detection.
• The framework demonstrated improved segmentation accuracy on real-world clinical data.
• GER-UNet showed potential in reducing sample complexity and filter redundancy.

Conclusions:

• The proposed group equivariant segmentation framework effectively leverages image symmetries for enhanced medical image analysis.
• GER-UNet represents a significant advancement over traditional CNNs for segmentation tasks.
• This approach offers potential for broader applications in medical imaging, including organ delineation across modalities.