Model-based self-supervised learning for quantitative assessment of myocardial oxygen extraction fraction and myocardial blood volume

Purpose: To develop a model-driven, self-supervised deep learning network for end-to-end simultaneous mapping of myocardial oxygen extraction fraction (mOEF) and myocardial blood volume (MBV). Methods: An asymmetrical spin echo–prepared sequence was used to acquire mOEF and MBV images. By integrating a physical model into the training process, a self-supervised learning (SSL) pattern can be regulated. A loss function consisted of the mean squared error, plus cosine similarity was used to improve the performance of network predictions for estimating mOEF and MBV simultaneously. The SSL network was trained and evaluated using simulated data with ground truths and human data in vivo from 10 healthy subjects and 10 patients with myocardial infarction. Results: In the simulation study, the SSL method demonstrated the ability of generating relatively accurate mOEF, MBV, and ΔB maps simultaneously. In the in vivo study, healthy volunteers had an average mOEF of 0.6–0.7 and MBV of 0.11–0.13, comparable to literature-reported values. In the myocardial infarction regions, the average mOEF and MBV in 5 tested patients reduced to 0.45 ± 0.09 and 0.09 ± 0.02, which were significantly lower (p < 0.001) than those in normal regions (0.67 ± 0.04 and 0.13 ± 0.01, respectively). Conclusion: This work has demonstrated the initial feasibility of generating mOEF and MBV maps simultaneously by a model-driven, self-supervised learning method.