SCMR/ISMRM Co-Provided Workshop
SCMR 22nd Annual Scientific Sessions
Pedro Ferreira, PhD
Physicist
The Royal Brompton Hospital
Andrew Scott
Physicist
The Royal Brompton Hospital
Zohya Khalique, MD
Academic clinical lecturer
Royal Brompton Hospital
Guang Yang, PhD, MSc
Senior Research Fellow
Imperial College London
Sonia Nielles-Vallespin, PhD, MSc, BSc
Principal Physicist
Royal Brompton Hospital, Imperial College of London
Dudley Pennell, MD, FSCMR
Director CMR Unit
Royal Brompton Hospital
David Firmin, PhD
Professor of Biomedical Imaging
Royal Brompton Hospital
Background:
Diffusion tensor cardiac MR (DT-CMR) is an emerging technique that provides information on myocardial microstructure. Many different diffusion parameters can be extracted from the tensors including tensor orientation measures which have been shown to relate to the orientation of the local cardiomyocytes and their sheetlet structure [1].
Currently, DT-CMR post-processing is done retrospectively, offline. Multiple steps require manual input, including removing frames corrupted by motion artefacts, and thresholding and segmentation of the left ventricular (LV) myocardial wall. Myocardial segmentation is particularly important in order to enable visualization of the tensor orientation in relation to the local cardiac coordinates defined by the longitudinal, circumferential and radial orthogonal directions.
In this work we develop and validate a fully automated post-processing framework of in vivo DT-CMR data with deep learning. with the future aim of producing on-the-fly DT-CMR results at the scanner, providing real-time feedback on the scan quality.
Methods:
Two convolutional neural networks (CNN) were trained and tested using previously collected DT-CMR scans, acquired with a STEAM-EPIsequence. The data include a total of 348 healthy and 144 cardiomyopathy patient scans with a total of 26,675 diffusion images. All data had been previously examined for artifacts and segmented by an experienced user. These data were divided 60% for training and 40% for testing sets. MATLAB was used for the deep learning and diffusion tensor post-processing. Class balancing, and data augmentation were used to improve training.
Image classifier:A simplified version of the VGG16 CNN [2] (figure 1A) was used to detect and reject diffusion frames corrupted by motion artefacts.
Image segmentation:A SegNet CNN [3] (figure 1B) was trained to segment the heart and more importantly the LV myocardium. The encoder network is also based on the VGG16classification network.
After training and testing, the CNNs were integrated within the DTI post-processing pipeline and compared against an experienced user on the test data.
Results:
An accuracy of 94% was achieved when detecting corrupted frames. For the myocardial segmentation a global median Dice coefficient of 0.93 was achieved (figure 2).
Figure 3 shows Bland-Altman plots comparing deep learning against an experienced user for global values of secondary eigenvector orientation (E2A), fractional anisotropy and mean diffusivity in healthy and patient data.
Conclusion:
The fully automated DT-CMR analysis with deep learning performed effectively with high levels of accuracy when compared to an experienced user. Of note, the myocardial segmentation learned to correctly exclude papillary and right ventricular trabeculation from the LV myocardium, even in patient data with a more varied morphology. We are currently porting this post-processing pipeline to enable online post-processing and provide feedback to the scanner operator during acquisition.