Compositionally Equivariant Representation Learning.

Area of Science:

• Computer Vision
• Medical Imaging Analysis
• Machine Learning

Background:

• Deep learning models require extensive labeled data for effective training in medical image segmentation.
• Human visual systems demonstrate rapid, generalizable learning from minimal supervision due to compositional pattern recognition.
• Current medical imaging models lack robust representation of compositional structures found in human anatomy.

Purpose of the Study:

• To develop interpretable and generalizable representations for medical image segmentation by leveraging compositionality.
• To propose that generative factors in medical images possess compositional equivariance, which good representations should approximate.
• To explore enforcing compositional equivariance in un-, weakly-, and semi-supervised learning settings.

Main Methods:

• Utilized compositionality to learn representations for medical image segmentation.
• Modeled compositional representations using learnable von-Mises-Fisher (vMF) kernels.
• Investigated design and learning biases to enforce compositional equivariance.

Main Results:

• Achieved state-of-the-art performance on semi-supervised domain-generalized medical image segmentation.
• Demonstrated superior results compared to several strong baselines.
• Validated the effectiveness of compositional equivariance in representation learning.

Conclusions:

• Compositional equivariance is a key property for learning effective representations in medical image segmentation.
• The proposed method enhances model generalization and interpretability, particularly in low-data regimes.
• Future work can build upon this framework for improved medical AI applications.