HorusEye: a self-supervised foundation model for generalizable X-ray tomography restoration.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• X-ray tomography is crucial in science and medicine.
• Image degradation, especially in low-dose scenarios, hinders analysis.
• Current restoration methods lack generalizability across different modalities and degradations.

Purpose of the Study:

• To develop a generalizable X-ray tomography image restoration method.
• To introduce HorusEye, a self-supervised foundation model for image restoration.
• To overcome limitations of existing task-specific restoration techniques.

Main Methods:

• Formulated image restoration as learning nonparametric acquisition degradation from data.
• Introduced HorusEye, a self-supervised foundation model using interslice contrastive pretraining.
• Trained HorusEye on over 100 million images without paired supervision.

Main Results:

• HorusEye demonstrated generalization across diverse modalities and restoration tasks.
• Achieved superior performance compared to task-specific methods.
• Showcased improved photon efficiency and recovery of high-frequency details.
• Clinical studies confirmed enhanced detectability of low-contrast anatomy and lesions.

Conclusions:

• HorusEye effectively restores X-ray tomography images by learning degradation directly from data.
• The model offers a general postprocessing solution applicable to various X-ray imaging scenarios.
• HorusEye improves diagnostic accuracy and downstream task performance in medical imaging.