Explainable rotation-invariant self-supervised representation learning.

Area of Science:

• Medical Imaging AI
• Computer Vision
• Machine Learning

Background:

• Real-world medical imaging often contains rotational noise, degrading AI model performance.
• Developing rotation-invariant models is crucial for reliable medical AI applications.

Purpose of the Study:

• To present a novel framework, RISC (rotation invariant self-supervised vision framework), for robust detection and classification of rotated medical images.
• To address rotational corruption in medical AI by incorporating self-supervised learning for rotation invariance.

Main Methods:

• Proposed a representation learning approach using self-supervised learning for rotation invariance.
• Developed the RISC framework to correct rotational corruptions in medical images.
• Utilized GradCAM for explainability of the self-supervised pretext task and classification outcomes.

Main Results:

• Achieved state-of-the-art rotation-invariant classification results on benchmark datasets.
• RISC demonstrated accuracy improvements of 22% on OrganAMNIST, 17% on PneumoniaMNIST, and 2% on RetinaMNIST compared to rotation-affected benchmarks.
• Provided explainability for the model's performance on rotated medical imagery.

Conclusions:

• The RISC framework effectively enhances the robustness of AI models against rotational variations in medical images.
• Self-supervised learning for rotation invariance is a viable strategy for improving medical AI diagnostic accuracy.
• Explainability methods confirm the effectiveness of RISC in handling rotationally corrupted data.