A geometric approach to robust medical image segmentation.

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Deep learning segmentation models are vital for clinical practice but struggle with variations in medical imaging data, such as magnetic resonance imaging (MRI) scans.
• Differences in MRI acquisition protocols lead to diverse image characteristics, impacting model performance and reliability.

Purpose of the Study:

• To develop a robust deep learning segmentation method for magnetic resonance imaging (MRI).
• To enhance model generalisation across different imaging domains by leveraging anatomical shape information.

Main Methods:

• Utilised multiple MRI sequences to learn texture-invariant and shape-equivariant features.
• Constructed a shape dictionary using vector quantisation and explored shape equivariance with varying group orders.
• Achieved shape equivariance via contrastive learning or by imposing constraints on convolutional kernels.

Main Results:

• Demonstrated that increased group order in shape equivariance improves model robustness.
• Achieved state-of-the-art performance in single domain generalisation for prostate and cardiac MRI segmentation.
• Developed a method that effectively captures anatomical shape variations for improved segmentation.

Conclusions:

• The proposed geometric approach enhances the robustness of deep learning segmentation models in clinical MRI.
• Harnessing anatomical shape variation is key to improving model generalisation and reliability across diverse imaging domains.
• The method offers a promising direction for reliable automated segmentation in medical imaging applications.