Focus Session
SCMR 22nd Annual Scientific Sessions
Vivek Jani, MSc
Graduate student
The Johns Hopkins University
Gisela Teixidó-Turà, MD
Faculty
Hospital Vall d'Hebron
Nadjia Kachenoura, PhD
PhD
Sorbonne Université, INSERM, CNRS
Kevin Bouaou, PhD
PhD student
Sorbonne Université, INSERM, CNRS
Alain De Cesare, PhD
PhD
Sorbonne Université, INSERM, CNRS
Colin Wu, PhD
Mathematical Statistician
National Institutes of Health
David Bluemke, MD, PhD
Professor
University of Wisconsin School of Medicine and Public Health
.jpg "Joao Lima, MD photo")
Joao Lima, MD
Professor of Medicine, Radiology and Epidemiology
Johns Hopkins University
Bharath Ambale-Venkatesh, PhD
Instructor
The Johns Hopkins University
Background: Direct measurement of central aortic stiffness by means of aortic distensibility has been shown to be an early marker of subclinical vascular alteration and associated with incident cardiovascular disease (CVD) and mortality. However, current methods to measure aortic area from magnetic resonance imaging (MRI) require manual intervention to identify the ascending and descending aorta. We propose application of 2D U-Net, a well-established convolutional neural network (CNN) architecture, for segmentation of the ascending and descending aorta, and compare its performance with user-dependent semi-automated analysis.
Methods: Images were obtained as part of the Multi-Ethnic Study of Atherosclerosis (MESA) in 2,600 participants between 2000-2002 on 1.5T MRI scanners. Each patient had an axial 2-dimensional phase-contrast (PC) MRI performed just above the level of pulmonary artery bifurcation to capture both the ascending (AA) and descending (DA) aortic cross-sections.. Imaging parameters were: repetition time of 10 ms, echo time of 1.9 ms, field of view of 34 cm, slice thickness of 8 mm, matrix of 256 × 224, temporal resolution of 20 ms, through plane encoding velocity of 150 cm/s, and bandwidth of 32 kHz. Time-resolved semi-automated segmentation of the AA and DA lumen was performed using the previously validated ARTFUN software (INSERM, Paris, France). Such segmentation along with the corresponding magnitude PC-MRI images from all phases of the cardiac cycle (15-20) were used to train the U-net segmentation algorithm with 4 contracting (each with a convolution unit, rectified linear unit, and a max pooling unit) and 4 expanding (up-convolution and concatenation) sequences. The CNN was trained on 24256 training images and then tested on 8085 validation images. The CNN was trained with a categorical cross entropy loss function and an Adam optimizer over 250 iterations with learning rate of 1 x 10-5. Two different classification schemes, namely binary segmentation (separating the ascending and descending aorta as one class from background), and multi-class segmentation (identification of the ascending and descending aorta as two separate classes), were assessed. Train and test set accuracy were determined utilizing a dice coefficient, comparing the similarity between manually segmented images and segmentation predictions from DL.
Results: The U-net algorithms (Figure 1) performed with high accuracy for segmentation of the AA and DA cross-sections with a mean Dice-coefficient of 98.8% (min: 86.4%, max: 99.9%) using the binary segmentation method, and a mean Dice coefficient of 98.7% (min: 72.6%, max: 99.9%) using the multi-segmentation method. Examples of semi-automated vs. DL-derived segmentations are shown in Figure 2 for binary segmentation, and Figure 3 for multi-segmentation.
Conclusion: U-net algorithm is shown to be accurate for aortic cross-sectional area segmentation from cine PC-MRI images.