Deep Learning Reconstruction does not influence T1 values on cardiac T1 mapping (RF_SA_438)

Cardiac T1 mapping in MRI is a valuable non-invasive technique for assessing myocardial tissue. The accuracy and precision of T1 values are crucial for diagnosis, but these values can be influenced by factors like sequences, field strength, and scanners(1). Deep learning reconstruction (DLR) methods have emerged as a way to enhance image quality, including on quantitative imaging(2), but their effect on cardiac T1 mapping values remain to be assessed.

Methods: This prospective study involved 91 patients (average age 52 years; 36 females) who underwent cardiac MRI with T1 mapping using the MOLLI 5(3s)3 sequence, both with and without DLR. The MRI scans were conducted on a 1.5T SIGNA Artist scanner (GE HealthCare) with software version 30.0 using the AIR Recon DL algorithm. The acquisitions were performed at three slice levels (basal, mid-ventricular, and apical short-axis) The acquired T1 maps were then analyzed segmenting each slice into 100 centroids and merging them into the 16 standard AHA segments. The mean and standard deviation of T1 values were calculated for each segment.
In addition to the in vivo patient data, a T1 phantom(3) with known T1 values was also imaged to assess the accuracy of the T1 measurements with and without DLR.

Results:

The in vitro analysis revealed that both standard reconstruction and DLR underestimated the true T1 values, with mean differences of -62.0±119.7ms and -70.0±129.9ms, respectively. However, they demonstrated excellent agreement with each other (R²=1, CoV=0.84%, ICC=1[1 1], Mean Difference= -7.9±27.4ms). In the in vivo analysis, the mean T1 value for the population was 1033.0±47.9ms without DLR and 1034.1±52.1ms with DLR, showing no significant difference. The agreement between global T1 mapping values with and without DLR was excellent (ICC=0.92[0.88, 0.94], CoV=1.39%, Mean Difference=1.1±40.0ms). Regional T1 map values also showed excellent agreement (ICC=0.82[0.81, 0.84], CoV=2.68%, Mean Difference=0.7±79.3ms), with only two out of 16 segments showing statistically significant differences, but with a CoV of less than 3.7%. The most striking finding was the significant decrease in the standard deviation of local T1 values with DLR, indicating a decrease in noise. The mean SD relative decrease across all 16 segments was 24%.

Conclusion:

The application of deep learning reconstruction to MOLLI T1 mapping maintains the T1 values while significantly decreasing noise (standard deviation within ROI). The study supports that DLR can be safely incorporated into clinical practice without the need for new reference T1 values.