Precision measurement of cardiac structure and function in cardiovascular magnetic resonance using machine learning

Abstract

Background: Measurement of cardiac structure and function from images (e.g. volumes, mass and derived parameters such as left ventricular (LV) ejection fraction [LVEF]) guides care for millions. This is best assessed using cardiovascular magnetic resonance (CMR), but image analysis is currently performed by individual clinicians, which introduces error. We sought to develop a machine learning algorithm for volumetric analysis of CMR images with demonstrably better precision than human analysis.

Methods: A fully automated machine learning algorithm was trained on 1923 scans (10 scanner models, 13 institutions, 9 clinical conditions, 60,000 contours) and used to segment the LV blood volume and myocardium. Performance was quantified by measuring precision on an independent multi-site validation dataset with multiple pathologies with n = 109 patients, scanned twice. This dataset was augmented with a further 1277 patients scanned as part of routine clinical care to allow qualitative assessment of generalization ability by identifying mis-segmentations. Machine learning algorithm ('machine') performance was compared to three clinicians ('human') and a commercial tool (cvi42, Circle Cardiovascular Imaging).

Findings: Machine analysis was quicker (20 s per patient) than human (13 min). Overall machine mis-segmentation rate was 1 in 479 images for the combined dataset, occurring mostly in rare pathologies not encountered in training. Without correcting these mis-segmentations, machine analysis had superior precision to three clinicians (e.g. scan-rescan coefficients of variation of human vs machine: LVEF 6.0% vs 4.2%, LV mass 4.8% vs. 3.6%; both P < 0.05), translating to a 46% reduction in required trial sample size using an LVEF endpoint.

Conclusion: We present a fully automated algorithm for measuring LV structure and global systolic function that betters human performance for speed and precision.

Keywords: Cardiac magnetic resonance; Cardiovascular imaging; Image processing; Machine learning; Ventricular function.

Fig. 1

Overview of study design. A training set of segmented images from 1932 patients with multiple diseases from multiple centres were used to train four convolutional neural networks (CNNs). CNN segmentations were combined to measure left ventricular (LV) cavity volumes, systolic function and myocardial mass. Machine segmentations were compared to clinical segmentations on an independent dataset to measure precision. EDV end diastolic volume, ESV end systolic volume, EF ejection fraction, LVM LV mass, MV mitral valve, SAx short axis

Fig. 2

Structure of the Unet used for short axis image segmentation. The model takes a grayscale CMR image with dimension 192 × 192 and creates a segmentation mask of the same dimension with 3 channels (one channel for each of: LV blood pool (white), myocardium (gray) and background (black)). The Unet used for long axis segmentations were the same, but image sizes and final layer were different—see Additional file 1: Table S1 for full details

Fig. 3

Spatial normalisation. The geometric relationship between the SAx, 2Ch and 4Ch planes are known—the three planes are overlaid in 3D in the left image. Spatial normalisation of each image is performed by transformation to a normalised reference frame as shown in the right image. 2Ch 2-chamber, 4Ch 4-chamber, SAx short-axis

Fig. 4

Mitral annular position encoding. The image on the left shows the lateral mitral annular point overlaid on the CMR image. The image on the right was created by measuring the distance to the mitral annular point from each pixel position and weighting with a Gaussian function; the position of the point is overload for illustration. The bottom image shows the CMR image and distance map overlaid. For clarity, only one of the two points is shown here. MV mitral valve

Fig. 5

Composition of training data. List of countries, cities, institutions, scanner brand, scanner models and conditions (disease or healthy) used in the training dataset. AFD Anderson-Fabry Disease, AS  aortic stenosis, HCM  hypertrophic cardiomyopathy

Fig. 6

Example segmentations by machine learning algorithm. Top row: a pair of diastole images from the scan:rescan dataset that has been segmented by the automated algorithm. Note that the LV metrics are not exactly the same due to intrinsic variability in how slices are prescribed. Bottom left: example of an error (1 in 479 error rate) where laminar thrombus had been mis-identified as myocardium since this had not been ‘seen’ in the training data before. Bottom right: a mis-segmentation due to a pericardial effusion

Fig. 7

Machine and human precision evaluated on 109 subjects. Intra-observer reliability and scan-rescan repeatability, expressed as coefficient of variations (%) with 95% confidence intervals in brackets. Note that the intra-observer reproducibility is zero for all LV metrics. *Denotes statistical significance; ** denotes highly significant difference. EDV end diastolic volume, ESV end systolic volume, EF ejection fraction, LVM LV mass