Deep learning to estimate cardiac magnetic resonance–derived left ventricular mass

Figure 1. Overview of left ventricular (LV) mass algorithms. Depicted is an overview of the 3 approaches to cardiac magnetic resonance–derived LV mass estimation compared in the current study. The top model utilizes deep learning–based regression trained by manually labeled LV mass. The middle model performs deep learning–based segmentation informed by InlineVF contours. The bottom model utilizes the InlineVF automated contours alone. For the deep learning segmentation and InlineVF models, LV segmentations were converted to LV mass by summing pixel volume and multiplying by the density of LV myocardium (1.05 g/mL, see text).

Figure 2. Distributions of cardiac magnetic resonance (CMR)-derived left ventricular (LV) mass obtained using each estimation method. Depicted are density plots showing the distribution of CMR-derived LV mass (x-axis) using mean-centered InlineVF (left panel), the deep learning regression model (middle panel), and the deep learning segmentation model (right panel). Results are shown for the full sample with available CMR imaging (N=38,464).

Figure 3. Correlation between manually labeled left ventricular (LV) mass and derived left ventricular mass estimated using each model. Depicted are plots illustrating the correlation between manually labeled LV mass (y-axis) and cardiac magnetic resonance–derived LV mass using InlineVF (left panel), deep learning regression (middle panel), and deep learning segmentation (right panel). Results are shown among individuals within the test set independent of model training. Estimates for InlineVF and the segmentation model are displayed after centering the distribution upon the observed sex-stratified mean manually labeled LV mass (see text).

Figure 4. Bland-Altman plots comparing manually labeled left ventricular (LV) mass and derived LV mass using each model. Depicted are Bland-Altman plots16 showing agreement between manually labeled LV mass and LV mass estimated using InlineVF (top), ML4Hreg (middle), and ML4Hseg (bottom). Estimates using InlineVF and ML4Hseg are depicted after mean centering (see text). In each plot, each point represents a paired observation (ie, the manually labeled LV mass estimate and the model predicted LV mass estimate). The x-axis depicts increasing mean of the paired observations. The y-axis depicts the difference between the paired observations, with negative values representing pairs in which manually labeled LV mass was larger than model-predicted LV mass (underestimation using the model). The colored horizontal line shows the overall mean difference within each sample, and the hashed horizontal lines show the upper and lower bounds of the mean difference (defined as ±1.96 standard deviations of the difference). The corresponding bounds (a surrogate for level of agreement) and the proportion of observations within those bounds are depicted on each plot. A total of 13 (InlineVF), 1 (ML4Hreg), and…