Myocardial motion tracking and strain calculation using Deep Learning networks
Edward Ferdian, Dr Avan Suinesiaputra and Professor Alistair Young, Department of Anatomy and Medical Imaging, Faculty of Medical and Health Sciences
Description
Cardiac Magnetic Resonance (CMR) is a non-invasive imaging technique to analyse patient heart. CMR is currently the gold standard among other imaging techniques due to its superb contrast, no radiation exposure and its versality to produce different images in a single examination session. CMR tagging is one of the image protocols that artificially creates stripe or grid markers on the acquired images. These markers follow the deformation of the heart, allowing high resolution of cardiac motion tracking and analysis. However, full analysis of cardiac motion from CMR tagging is still time consuming and requires significant post-processing time.
In this project, we developed a machine learning algorithm to fully automatically detect the location of landmark points inside the myocardium from CMR tagging images when the heart is at full relaxation (end-diastole) and track the motion of these points in subsequent image frames throughout the whole cardiac cycle. We built two deep learning neural networks for these tasks. The first network was trained to locate a bounding box covering the myocardium and it was based on 7 layers of convolutional neural networks (CNN). The output of this network is fed into the second network, which was trained to detect landmark points and follow their movements. The second network was based on RNNCNN; a combination of CNN and Recurrent Neural Network (RNN) with Long Short-Term Memory (LSTM) units.
Training the networks
Deep learning networks require massive size of data and high computational power during training. We trained the whole pipeline using 4,508 CMR cases, obtained from the UK Biobank study. In total, there were 12,409 image sequences, each consisted of 20 image frames, and the input for the network was 256×256 image pixels. The whole network consists of 39.4 millions of parameters to optimise. With the help and support from the Center for e-Research, we trained our networks on the Nectar Research Cloud with GPU processing capability (NVIDIA Tesla K40). The training process required approximately 12 hours using TensorFlow ver. 1.5.0 library for deep learning.
Results
We calculated myocardial strain from the landmark points predicted by the networks and compared with expert analyses. The predicted strains were within 1% of the manual ground truth on average, except for basal radial strain (2.4%). The predicted strains also showed comparable intraclass correlation coefficient with the manual ground truth. This fully automated analysis of CMR tagging, including strain calculation, could detect landmarks and predict strains up to 920 frames per second (~4.5 minutes for the whole 12,409 image sequences).
In the future, we are aiming to develop an integrated machine learning analysis for cardiac MRI, including segmentation, breath-hold mis-registration, diagnosis and prognosis of cardiovascular disease, and 4D Flow analysis, without any manual interventions. Without the access of GPU-powered Nectar cloud, this research would not be practicable.
