Prototypical few-shot segmentation for cross-institution male pelvic structures with spatial registration

Area of Science:

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

Background:

• Few-shot learning enables efficient medical image analysis by utilizing limited labeled data for new class segmentation.
• Adapting models to new clinical structures often requires substantial data and expert annotations, posing a significant challenge.

Purpose of the Study:

• To develop a fully 3D prototypical few-shot segmentation algorithm for effective adaptation to unseen anatomical structures using minimal labeled data.
• To address spatial variability between institutions and imperfect data alignment in few-shot learning for medical imaging.

Main Methods:

• Integration of a novel spatial registration mechanism within prototypical learning, including a segmentation head and spatial alignment module.
• Proposal of a support mask conditioning module to leverage available annotations in support images for improved training.
• Extensive experiments on segmenting eight pelvic anatomical structures from 589 T2-weighted MR images across seven institutes.

Main Results:

• Demonstration of the efficacy of the 3D formulation, spatial registration, and support mask conditioning modules, both individually and collectively.
• Significant improvement in few-shot segmentation performance compared to existing 2D alternatives.
• Robust performance regardless of whether support data originated from the same or different institutes.

Conclusions:

• The proposed 3D prototypical few-shot segmentation algorithm effectively adapts to new clinical structures with limited data.
• Spatial registration and support mask conditioning are crucial for enhancing robustness and accuracy in cross-institutional few-shot segmentation.
• The developed method offers a statistically significant advancement over previous 2D approaches for medical image segmentation.