1351972 - Deep Learning Assisted CMR Myocardial Oxygen Extraction Fraction Measurements

Myocardial oxygen extraction fraction (mOEF) reflects the balance of oxygen supply and demand in the heart. A CMR 2D asymmetric-spin-echo (ASE) echo-planar imaging technique was previously reported to quantify mOEF in vivo [1]. However, the method suffers from image distortion and field inhomogeneity artifacts in myocardial wall adjacent to lung air. A new 2D ASE prepared sequence was developed to minimize distortion artifacts, but still prone to field inhomogeneity. In this project, we developed a deep-learning (DL) approach to reduce inhomogeneity artifacts and demonstrated much improved image quality of mOEF maps.

Methods:

The project was approved by local human study committee. Severn healthy volunteers were recruited to be scanned twice at different days. The mOEF maps at 3-5 short-axis directions were obtained during each scan. Each mOEF map was acquired with a breath-holding time of 16 RR intervals, using an ASE prepared balanced steady-state free precession (SSFP) pulse sequence. The spatial resolution was 1.7 x 1.7 mm².

A total of 48 mOEF maps from 5/7 volunteers were used for deep learning training. To reduce the field inhomogeneity around lateral free myocardial wall, mean mOEF and its standard deviation in the septum (the minimal field inhomogeneity area) were measured as the representative values for entire myocardial mOEF. In this way, a synthetic ground truth of OEF map was reconstructed. These simulated ground truth maps were then used as input to accomplish pixel-wise regression task using a UNet-based fully connected neural network (UFCN), as shown in Fig.1. The model received a multi-channel input which consists of ASE prepared images at different 180^º pulse shifts. The U-net structure was comprised of an encoder and a decoder respectively. The encoder was composed of 5 convolutional blocks each with 64, 128, 256, 512, and 1024 convolutional filters. The decoder consisted of 4 deconvolutional blocks with 512, 256, 128 and 64 output channels. A final 1×1 kernel size convolutional layer with the same filter number as input channel was used at the end to implement weighted summation in the former convolutional layer. The CMR data from rest of 2/7 volunteers were tested for the mOEF map and an initial testing of reproducibility.

 

Results:

The OEF maps showed dramatic improved image quality, compared with original OEF maps (Fig 2). The averaged global myocardial OEF was 0.62 ± 0.07, which agrees well with literature values of 0.64 ± 0.14 measured by positron emission tomography (PET) methods [2]. The initial reproducibility tests provided a coefficient of variance of 5% {0%, 8.4%].

Conclusion:

The UFCN postprocessing method dramatically reduced the calculation error of mOEF in the field inhomogeneity area of myocardium in healthy hearts. Future study may need much larger data sets in diseased hearts for the training sets to detect regional difference in mOEF.