Mitigating Spurious Invariance in Contrastive Learning: A Probabilistic Self-Supervised Learning Framework for...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Manual annotation of medical images is costly and time-consuming.
• Self-supervised learning (SSL) shows promise for medical imaging tasks.
• Contrastive learning, a type of SSL, faces challenges with medical image features being suppressed during augmentation.

Purpose of the Study:

• To address the limitations of contrastive learning in medical imaging.
• To propose a novel probabilistic framework for SSL in medical imaging.
• To improve the representation of medical image features during self-supervised learning.

Main Methods:

• Developed a method to encode learned feature embeddings as probability distributions.
• Proposed a unified probabilistic framework integrating multiple SSL approaches.
• Evaluated the framework on medical image segmentation and classification tasks across multiple datasets.

Main Results:

• The probabilistic framework achieved superior performance compared to state-of-the-art SSL methods.
• Demonstrated enhanced robustness to stronger augmentations.
• Showcased label efficiency with limited fine-tuning labels.

Conclusions:

• The proposed probabilistic framework offers a more nuanced and effective approach to SSL in medical imaging.
• This method overcomes the limitations of traditional contrastive learning by preserving critical medical features.
• The framework shows significant potential for improving downstream medical imaging tasks with greater efficiency and robustness.