Unsupervised knowledge-transfer for learned image reconstruction

Area of Science:

• Medical imaging
• Deep learning
• Computational imaging

Background:

• Deep learning image reconstruction shows promise but requires large paired datasets, often unavailable in medical imaging.
• Existing methods struggle with data distribution shifts, impacting reconstruction quality and reliability.

Purpose of the Study:

• To develop a novel unsupervised knowledge-transfer paradigm for learned image reconstruction within a Bayesian framework.
• To enable accurate reconstruction from limited or varied medical imaging data while providing uncertainty quantification.

Main Methods:

• A two-phase training approach: initial training on simulated data, followed by unsupervised fine-tuning on realistic data.
• Utilizing a Bayesian framework to incorporate prior knowledge and generate predictive uncertainty maps.
• Experimental validation on low-dose and sparse-view computed tomography datasets.

Main Results:

• The proposed unsupervised method achieves competitive performance against state-of-the-art supervised and unsupervised techniques.
• Significant improvements in visual and quantitative metrics (PSNR, SSIM) were observed, especially for data with different distributions than the training set.
• The framework successfully provides predictive uncertainty information for reconstructed images.

Conclusions:

• The developed unsupervised Bayesian framework effectively addresses the data scarcity issue in deep learning-based medical image reconstruction.
• This approach offers robust and accurate reconstructions, even with domain shifts, and quantifies uncertainty, enhancing clinical utility.