Quantification of myocardial oxygen extraction fraction on noncontrast MRI enabled by deep learning.

Purpose: To develop a new deep learning enabled cardiovascular magnetic resonance (CMR) approach for noncontrast quantification of myocardial oxygen extraction fraction (mOEF) and myocardial blood volume (MBV) in vivo.

Materials and methods: An asymmetric spin-echo prepared CMR sequence was created in a 3 T MRI clinical system. A UNet-based fully connected neural network was developed based on a theoretical model of CMR signals to calculate mOEF and MBV. Twenty healthy volunteers (20-30 years old, 11 females) underwent CMR scans at 3 short-axial slices (16 myocardial segments) on 2 different days. The reproducibility was assessed by the coefficient of variation. Ten patients with chronic myocardial infarction were examined to evaluate the feasibility of this CMR method to detect abnormality of mOEF and MBV.

Results: Among the volunteers, the average global mOEF and MBV on both days was 0.58 ± 0.07 and 9.5% ± 1.5%, respectively, which agreed well with data measured by other imaging modalities. The coefficient of variation of mOEF was 8.4%, 4.5%, and 2.6%, on a basis of segment, slice, and participant, respectively. No significant difference in mOEF was shown among 3 slices or among different myocardial segments. Female participants showed significantly higher segmental mOEF than male participants (P < .001). Regional mOEF decrease 40% in CMR-confirmed myocardial infarction core, compared to normal myocardial regions.

Conclusion: The new deep learning-enabled CMR approach allows noncontrast quantification of mOEF and MBV with good to excellent reproducibility. This technique could provide an objective contrast-free means to assess and serially measure hypoxia-relief effects of therapeutic interventional strategies to save viable myocardial tissues.