720 Views
SCMR/ISMRM Co-Provided Workshop
SCMR 22nd Annual Scientific Sessions
Ganesh Adluru, PhD, FSCMR
Assistant Professor
UCAIR, University of Utah
Ye Tian, MSc
Graduate Research Assistant
UCAIR, University of Utah
Jason Mendes, PhD
Research Associate
UCAIR, University of Utah
Edward DiBella, PhD
Professor
UCAIR, University of Utah
Brent Wilson, MD, PhD
Director, Cardiovascular Center
University of Utah
Background: Myocardial T1 mapping is a promising technique for characterizing myocardial tissue for various cardiomyopathies (1-4). Native T1 mapping is used to identify necrosis after acute myocardial infarction (1, 5). Native T1 mapping at adenosine stress and at rest was shown to be promising in differentiating patients with and without obstructive CAD as well as in detecting microvascular disease (6-8). MOLLI (2) is a widely used technique and with variations for rapid T1 mapping (9, 10). Conventional MOLLI uses 17 heart beat breath-held acquisition (2) while a shortened MOLLI (shMOLLI) uses 9 heartbeats (9). A 11 beat MOLLI acquisition scheme has also been proposed (10). In this scheme, an inversion pulse is followed by the acquisition of 5 images; a 3 beat wait period for recovery is then followed by acquisition of 3 images. Here we explore the idea of using deep learning to obtain a T1 map using 5 images acquired from 5 beats after one inversion pulse.
Methods: A total of 116 short-axis and 4-chamber long axis slices from 35 patients were acquired pre-contrast at rest using the 11 beat MOLLI acquisition on a Siemens Prisma 3T scanner. The scan parameters were TR=2.2 msec, TE=1.12 msec, flip angle=35○. Five images that were acquired after the first inversion pulse along with their respective inversion times were input to the deep learning network. The network was trained to output the T1 map obtained using 11 beats. All the images were motion compensated prior to input to the network. We used a multi-layer perceptron network with 10 hidden layers. Each layer consisted of 64 neurons followed by a non-linear rectified linear unit. Adam optimization (11) was used with a learning rate of 0.0001. The network was trained with batch sizes of 60K voxels for 80 epochs using an L2 loss function.
Results: Figure 1 shows the root-mean-squared error (RMSE) between the ‘True’ and estimated T1 maps as a function of epochs. The error is monotonically decreasing. Figure 2 shows the T1 maps obtained using 5 beat data (that was not used in training) for a short-axis slice and a 4 chamber slice. Corresponding T1 maps obtained with conventional fitting (2) using the 5 beat data are also shown. T1 maps from deep learning match closely with the ‘True’ T1 maps. Figure 3 shows correlation and Bland-Altman plots for T1 values obtained from 27 myocardial segments from 3 patients that were not used in training. Short axis slices were segmented into 6 regions and the long axis slices were segmented into 5 regions. Mean percentage error in T1 values over all of the segments was ~1.2%. Although not shown here, mean percentage error increased to ~2.8% when the network was trained using 4 beat data instead of 5.
Conclusion: Deep learning is promising for rapid myocardial T1 mapping. The 5 beat acquisition can reduce the burden of changing heart-rate between image acquisitions and reduce the breath-holding duration which can be helpful especially at stress.