Motivation: Supervised deep learning methods require application-specific training datasets and perform poorly with out-of-distribution data. Scan-specific methods do not require training data, but need careful hyperparameter tuning. Goal(s): To propose an automatically hyperparameter-optimized, scan-specific deep learning method that reconstructs various highly accelerated MRI acquisitions. Approach: We developed a self-supervised, bilevel-optimized, implicit neural representation (INR) network. It splits undersampled data into training and validation sets and applies Bayesian optimization for hyperparameter tuning. A multiresolution trainable parametric encoder reconstructs accelerated MRI scans. Results: Our method achieved performance comparable to oracle-optimized reconstructions, demonstrating the benefits of automatic hyperparameter optimization, and outperformed existing model-based and self-supervised methods. Impact: By automatically optimizing hyperparameters for scan-specific deep learning, our method reconstructs accelerated MRI scans across diverse protocols with superior image quality. It avoids reliance on training data and complicated task-dependent tuning, enhancing the clinical applicability of deep learning in MRI.
Yu et al. (Tue,) studied this question.