Fully Automated, Machine Learning-based Assessment of Myocardial Extracellular Volume vs. Clinical Routine in a large CMR-cohort

Background: Cardiovascular Magnetic Resonance (CMR) imaging provides an indispensable tool in assessing myocardial tissue composition and function. With extracellular volume (ECV) quantification it offers a non-invasive means to assess myocardial fibrosis, interstitial expansion, and overall tissue health. Conventional ECV measurement involves a multi-step process of motion correction and manually contoured pre- and post-contrast mapping. We tested a fully automated, machine learning-based ECV assessment in a large routine clinical cohort.

Methods: We retrospectively analyzed 868 patients that had visited our center between 2015 and 2020, and which had T1 mapping obtained and a point of care or laboratory hematocrit within 24 hours of scanning available. Native and enhanced T1 images were acquired using the modified look locker (MOLLI) sequence in a 3T (Ingenia, Philips Healthcare) and a 1.5T scanner (Achieva, Philips Healthcare). Regions of interest were manually drawn in the basal septum and the blood pool according to the guidelines. ECV was calculated using the established formula.

Development of the automated assessment followed an iterative process. Applying an expert-in-the-loop approach, U-nets were trained for segmentation of the relevant structures such as endo- and epicardial borders. Papillary muscles were removed from the blood pool mask using Otsu‘s thresholding method. Landmark extraction of right ventricular insertion points allowed generation of a septum mask, which was subsequently shrunk by one voxel to avoid overlap with the blood pool and thus, reduce the influence of partial volume effects.

Agreement between the two methods was assessed by Bland-Altman analysis and for performance of classification a contingency table was provided.

Results: The studied patient population was evenly distributed among the two scanners. ECV was slightly higher in 3T (p=0.013). As shown in Figure 1, automated ECV values of the entire myocardium (mean 1.08, limits of agreement (LoA) 7.20 - 5.05) and the septal myocardium (mean 1.60, LoA 10.88 - 7.68) showed good agreement with clinical ECV. With automated segmentation of the septum, 794 patients were classified correctly (631 as negative with ECV ≤30% and 162 as positive with ECV > 30%), while 70 patients would have been false positives (false positive rate = 8.06%), only 5 would have been classified as false negatives (false negative rate = 0.58%). Outliers were appreciated and inferior image quality or artifacts were identified as primary limitation leading to inaccurate segmentation.

Conclusion: The proposed machine learning-based approach provides a fully automatic and fast assessment of global and septal ECV in CMR with good agreement to manually drawn, conventional ECV measurement. 

Figure 1: Bland-Altman plots show the agreement of the different automated segmentations (2A Shrunken Myocardium; 2B Shrunken Septum) with the  manually measured ECV in clinical routine. 2C displays the contingency table indicating patient classification using fully automated ECV (ECV is considered elevated and here positive, if ECV is >30%).