CAREER: Robust Machine-Learning Methods for Imaging

Numerous models and algorithms are used in various imaging applications - including medical imaging, industrial- and security-imaging systems, non-destructive testing, seismic imaging, astronomical imaging, among others - to effectively characterize both static and time-varying objects and phenomena. Developing effective machine-learning (ML) approaches to enable high image and signal quality from limited or corrupted measurements in imaging is a critical area of ongoing research.

However, a primary challenge is that many existing deep-learning-based methods often require large training sets with the desired characteristics already labeled; yet acquiring or generating such training sets can be time-consuming - or even impractical - in many applications. Moreover, existing methods also face instabilities in handling perturbations in the imaging measurements, such as unseen noise, new experimental settings, and diverse image features at test time.

This project aims to significantly advance robust learning schemes for image reconstruction and data correction in imaging, encompassing both supervised and unsupervised learning, while using only limited amounts of training data. The proposed approaches will systematically leverage sparsity and other forms of prior information to learn neural networks from limited training data such that the networks are robust to various training-test variations and sources of performance degradation.

The work will develop methods and theory as well as demonstrate performance on data from magnetic resonance imaging (MRI), X-ray computed tomography (CT), and electroencephalography (EEG). Concurrently, a robust educational program will be run with participation from under-represented groups that encompasses undergraduate research experiences with external collaborations, student research conferences, and workshops with follow-up training for high-school students and teachers, all aimed at increasing participation in ML-driven imaging.

This project will develop ML methods for image reconstruction and data correction in imaging that are robust to a variety of measurement artifacts, training-test variations, and distribution shifts while using limited training data. First, learning methods will be developed for robust reconstruction in the supervised setting by exploiting sparsity of underlying networks as well as diffusion-model priors, and the underlying trade-offs surrounding robustness and generalization will be analyzed.

Second, novel measures and insights on robustness of learned models in unsupervised settings will be developed so as to extend reconstruction methods from supervised to unsupervised scenarios, again with limited training data. In situations in which the there is no training set at all, or cases in which there are no labels for a training set (for example, as occurs in deep-image-prior methods), the investigators will develop network structures and regularization schemes to alleviate instabilities and overfitting to measurement artifacts, applying the resulting methods to MRI and CT reconstruction with limited or corrupted measurements.

The investigators will also explore the extension of the schemes for correcting EEG artifacts and improving task accuracy. Finally, the research work will study theoretical properties of developed algorithms as well as performance limits.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.