SCMR/ISMRM Co-Provided Workshop
SCMR 22nd Annual Scientific Sessions
Behzad Sharif
Assistant Professor of Bioengineering & Medicine
Cedars-Sinai Medical Center & UCLA
Zulma Sandoval, PhD
Postdoctoral Scientist
Cedars-Sinai Medical Center
John Van Dyke, BSc
M.S. Student
Cedars-Sinai Medical Center
Rohan Dharmakumar, PhD
Professor
Biomedical Imaging Research Institute, Cedars-Sinai Medical Center
Background: With recent technical advances, fully-quantitative stress perfusion CMR is being adopted as a potentially superior modality for detection of ischemia by providing a validated quantitative tool for assessment of presence/severity of vasodilator-induced hypoperfusion. It can be argued that the most significant technical impediment for wider clinical adoption of fully-quantitative perfusion CMR is the lack of a fully-automatic post-processing workflow across all scanner platforms. On select platforms, retrospective motion correction with non-rigid registration is available which enables a faster workflow for manual analysis – although the accuracy of motion-correction varies significantly depending on the technique. We present an initial proof-of-concept based on a deep-learning approach for quantification of myocardial blood flow (MBF) that eliminates the need for motion correction, hence enabling a rapid and platform-independent post-processing framework. This is achieved by optimizing a cascade of deep convolutional neural networks (CNNs) to learn the common spatio-temporal features in a first-perfusion image series and use it to jointly detect the myocardial contours across all dynamic frames in the dataset.
Methods: Stress/rest perfusion images from 62 volunteer patients with suspected/known ischemia and 10 healthy volunteers were analyzed (free-breathing 3-slice protocol). Mean MBF for each slice was quantified using manual segmentation and Fermi deconvolution of the gadolinium time-curves. The proposed deep-learning network (Fig. 1) is composed of a cascade of two CNNs each with an optimized U-net architecture, and was trained using 70 of the available 72 perfusion studies (≈ 12,000 images). For two patients (not among the training dataset), the agreement between automatic vs. manual segmentation was assessed (Dice score) and mean per-slice MBF for the two approaches (3 stress and 3 rest MBF quantified for each patient) were compared using Pearson correlation.
Results: Fig. 2 shows automatic vs. manual segmentation results for a representative patient (computation time ≈ 0.1 sec/patient) demonstrating accurate endocardial/epicardial contouring across different contrast enhancement phases during free breathing. Comparison of segmentation results showed good agreement between automatic vs. manual approaches (Dice for myocardium: 0.80). Fig. 3 compares the MBF quantification results for automatic vs. manual processing (R2 = 0.97, p < 0.001).
Conclusion: By leveraging the power of deep neural networks for learning the common spatio-temporal features among thousands of CMR perfusion images, the presented results demonstrate the potential of an optimized architecture of CNNs for automatic MBF quantification in free-breathing stress perfusion CMR with strong agreement compared to manual processing. Future work involves refinement of the CNN architecture by utilizing anatomical constraints (e.g. from the cine images) to enable 16-segment quantification of MBF.