SCMR/ISMRM Co-Provided Workshop
SCMR 22nd Annual Scientific Sessions
Cardiovascular magnetic resonance (CMR) tissue tagging is the non-invasive gold standard for strain analysis1. However, fully automated strain analysis for CMR tagging is difficult, and can lead to significant post-processing time2–4.
We developed a fully automated machine learning pipeline to estimate the left ventricular (LV) circumferential strain (Ecc) and radial strain (ERR) from three short-axis CMR tagging slices (base, middle, and apex). The method employs two different neural networks, which perform the following tasks: 1) localisation of the left ventricle using a convolutional neural network (CNN), and 2) detection and tracking of myocardial landmarks through all the cardiac frames using a combination of a recurrent neural network (RNN) and a CNN (RNNCNN). Finally, strain calculations were obtained based on the displacement of the landmark points in every cine frame.
Figure 1 shows the end-to-end machine learning pipeline for strain estimation. In the landmark detection and tracking pipeline, we used weighted loss function that simultaneously minimizes mean squared distance, radial strain and circumferential strain errors, in each frame.
4,508 cases (12,409 image sequences) were obtained from the UK Biobank5. The data were randomly divided into 90% (training and validation), and 10% test data. Each case consisted of 2-3 short-axis slices with ~20 time frames. The ground truth was manual analysis from five expert readers by using Cardiac Image Modeller software (CIM v6.0, Auckland)6. Landmarks were defined in 168 locations (7 in radial direction with 24 points in each circumference) equally spaced inside the myocardium. Accuracy was calculated based on 1) ERR and ECC errors between the predicted strain and ground truth, 2) mean distance error of all landmarks within a frame.
The predicted strains at end-systole (ES) for the test data were within 1% of the manual ground truth on average (Table 1), except for basal ERR which was underestimated by 2.4%. The predicted strain intraclass correlation (ICC) also showed comparable agreement with the manual inter-observer ICC for both ECC and ERR. Mean error of the landmark positions at end-diastole (ED) was 3.0 ± 1.6 mm and at ES was 3.0 ± 1.2 mm. Figure 2 shows an example of the landmarks detected by the network at ED and ES and strain estimation for the whole sequence.
Combined neural network pipelines enabled fully automated estimation of radial and circumferential strains from CMR tagging images. To our knowledge, our method is the first to perform strain analysis based on machine learning from CMR tagging images to track as many as 168 landmarks by combining spatial and temporal features.