Deep learning-based segmentation and strain analysis of left heart chambers from long-axis CMR images

Abstract

Aims: Feature tracking (FT) is increasingly used on dynamic cardiac magnetic resonance (CMR) images for myocardial strain evaluation but often requires manual initialization, which is tedious and source of variability, especially on the challenging long-axis (LAX) images. Accordingly, we designed a pipeline combining deep learning (DL) with FT for left ventricular (LV) and left atrial (LA) longitudinal myocardial strain estimation.

Methods and results: We studied a multivendor database of 684 individuals divided into: training = 845, tuning = 281, and testing = 116 LAX-CMR cine 2- and/or 4-chamber views. Images were centre cropped. Then, a 2D- and 3D-ResUnet, which considers time as the third dimension, were designed for LV/LA segmentation and used to (i) estimate LV and LA strains (Full 2D-/3D-DL) and (ii) initialize an FT algorithm and further derive LV and LA strains (FT-initialized by 2D-/3D-DL). Left ventricular and LA contours and strain peaks were compared against reference standard (RS) measures performed by an expert using a semiautomated software. Intraclass-correlation-coefficient (ICC) was used to study reproducibility. 3D-DL outperformed 2D-DL segmentation (Dice-scores: 0.94 ± 0.02 vs. 0.90 ± 0.09, P = 0.002) and was stable across vendors, field strengths and imaging views. The added value of combining DL with FT was revealed by higher correlations and lower Bland-Altman biases against RS for FT initialized by 3D-DL strains (r ≥ 0.91, |mean-bias|≤0.65%) than for full 3D-DL strains (r ≤ 0.80, |mean-bias|<3.07%). Semiautomated human vs. FT initialized by 3D-DL (ICC ≥ 0.76) and inter-human strain reproducibility was equivalent.

Conclusion: Generalizable DL-based LV and LA segmentation on LAX-CMR images was proposed. Its combination with FT resulted in fully automated and reliable LV and LA strain measures, reaching human reproducibility.

Keywords: CMR; deep learning; feature tracking; left atrium; longitudinal strain.

Figure 1

Analysis pipeline for the estimation of LV and LA strain from CMR LAX images. Step 1: data preparation included resizing of all images to 512 × 512 pixels, when needed. Step 2: DL segmentation included dual network used to centre and crop images and then to segment left heart structures. Step 3: resulting DL contours were used on a single reference time to initialize FT, which was then achieved through all remaining phases of the cardiac cycle. Step 4: LV and LA strain waveforms were derived.

Figure 2

Contour extraction from LV and LA masks. From left to right: predicted labels; landmark definition, landmarks correspond to mitral anulus extremities, LV apex and LA roof, as well as to the point located at 2/3 of the LV long axis length from base to apex and to the LA center of mass; radial detection and equidistant distribution of contour points, as the transition between the heart label (one) and the background (zero); superimposition of contour points on the native CMR image.

Figure 3

Examples of left heart chamber DL-based contours and corresponding LV and LA time-resolved strain waveforms for a healthy individual, and a patient with abnormal cardiac chambers.