When faced with inverse problems involving large datasets but lacking ground-truth data, we propose an unsupervised deep learning framework to tackle the imaging challenge in Computed Tomography (CT). Our approach harnesses the inherent similarities among iterative reconstruction, Deep Image Prior (DIP), and unrolled optimization schemes, resulting in a training framework for amortized reconstruction.
After training on a dataset devoid of any image-domain ground truth, reconstruction of unseen scans simplifies to a single network forward pass.
We evaluated our method on the two-dimensional 2DeteCT dataset, demonstrating that our reconstructions are competitive with or superior to filtered back-projection, maximum-likelihood reconstruction, and a supervised network of identical architecture in a controlled, geometry-matched benchmark.
Compared to a per-image DIP baseline, our method achieves similar quality while eliminating the costly per-instance optimization, resulting in a speed-up of about four orders of magnitude. This positions our approach as a promising candidate for time-critical imaging applications.
Future work will focus on multi-dataset adaptability, counter-measures against over-smoothing, advanced uncertainty quantification, and further medical imaging inverse problems.
Blogger's Review: This research highlights the potential of unsupervised deep learning in medical imaging, particularly in scenarios lacking labeled data. By optimizing computational processes, the method not only enhances reconstruction speed but also maintains image quality, showing great promise for urgent medical imaging applications. Future research directions, especially regarding uncertainty handling and diverse dataset adaptation, are worth monitoring.