On the Role of Adversarial Perturbations in Unsupervised Deep Learning for Enhancing Image Quality of Degraded Medical-Images

Unsupervised learning enables deep-neural-network (DNN) training without high-quality data. Typical DNN methods for unsupervised image quality enhancement either (i) focus only on spatially uncorrelated noise, (ii) demand a training set comprising multiple noisy acquisitions per scene, or (iii) apply only to specific data; these aspects limit their applicability to medical images in clinical applications. Moreover, most such existing methods lack robustness to out-of-distribution (OOD) images observed during deployment. We propose a novel unsupervised DNN learning formulation to enhance a variety of medical images. Our unsupervised DNN training optimizes sets of multistage adversarial perturbations to (i) auto-generate multiple degraded inputs and latent-space encodings for each training image, (ii) adapt to the textural statistics of real-world degradations across medical imaging domains, and (iii) gain robustness to OOD images. This formulation also includes uncertainty modeling and estimation. Results on three publicly available datasets (CT, PET, MRI) show the benefits of our method in image-quality enhancement on in-distribution and OOD images.