Unsupervised knowledge-transfer for learned image reconstruction*

Deep learning-based image reconstruction approaches have demonstrated impressive empirical performance in many imaging modalities. These usually require a large amount of high-quality paired training data, which is often not available medical imaging. To circumvent this issue we develop novel unsupervised knowledge-transfer paradigm for learned within Bayesian framework. The proposed approach learns network two phases. first phase trains with set ordered pairs comprising ground truth images ellipses and the corresponding simulated measurement data. second fine-tunes pretrained to more realistic data without supervision. By construction, framework capable delivering predictive uncertainty information over reconstructed image. We present extensive experimental results on low-dose sparse-view computed tomography showing that competitive several state-of-the-art supervised techniques. Moreover, test distributed differently from can significantly improve quality only visually, but also quantitatively terms PSNR SSIM, when compared methods trained synthetic dataset only.