WP-008 - Deep Learning for High-resolution Myocardial Blood Flow Mapping: Rapid Automated Post-Processing with Robustness to Noise and Motion

Background: The application of machine learning to CMR analysis has experienced rapid growth owing to advancements involving deep neural networks (DNNs). In particular, new approaches using convolutional neural networks (CNNs) have been developed for automated analysis of cardiac function and T1 mapping [1,2]. Fully quantitative analysis of CMR perfusion datasets [3], however, has not yet been tackled by such approaches. Building on our recent CNN-based work for segmentation of perfusion images [4], we present a deep-learning approach for pixel-wise quantification of myocardial blood flow (MBF) without the need for motion correction. Furthermore, our approach replaces the noise-sensitive deconvolution process with a novel multi-stage DNN architecture with the goal of improving robustness to noise. We present initial results comparing our end-to-end DNN post-processing to conventional deconvolution-based quantification in the setting of high-resolution MBF mapping.

Methods: Stress/rest perfusion images from 76 volunteer patients with suspected ischemia were analyzed. All subjects underwent high-resolution free-breathing vasodilator-stress CMR (saturation-recovery FLASH with 1.7 x 1.7 x 8 mm³ at 3T). Conventional analysis was performed using Fermi-constrained deconvolution of the gadolinium (Gd) concentration time-curves. Fig 1 describes our DNN architecture, composed of a cascade of convolutional and fully-connected networks (CNNs/FCNs). The localizer/segmentation networks in Fig 1(A) were trained using ≈27,000 images with data augmentation (affine motion, zoom in/out, etc.). The feature-selectors and the multi-stage FCNs were trained independently to provide a progressively-refined and regularized MBF “prediction” while avoiding overfitting [5,6]. Enhanced robustness to noise for pixel-wise MBF mapping was achieved by meticulous data-augmentation of Gd time-curves.

Results: Fig 2 shows a representative stress study in a patient with known two-vessel disease. Fig 2(A) shows the deconvolution-based MBF maps with declining accuracy from low to high resolution (left to right) due to noise (arrows). However, the DNN-based MBF maps in Fig 2(B) demonstrate resilience to noise and remain consistent from low to high resolution. Fig 3 summarizes the variability of mean stress MBF for a given slice depending on the choice of resolution (ideally the variability should be zero) showing the superior robustness of our approach.

Conclusion: Our deep-learning approach enables, for the first time, comprehensive post-processing of free-breathing perfusion CMR images by generating high-resolution pixel-wise MBF maps without the need for motion correction. In addition to unprecedented computational speed up (< 2 seconds/study), our results suggest that the proposed multi-stage architecture has the potential to outperform conventional deconvolution-based quantification in the setting of high-resolution MBF mapping. Future work includes testing/validation of the developed framework in large multi-center datasets.