Deep learning segmentation and quantification of the left ventricle from the parasternal short axis view

Abstract Background Although parasternal short-axis (PSAX) images are routinely acquired during echocardiographic examinations, they remain under-utilized for quantitative analysis of the left ventricle (LV) morphology and function. This is even though PSAX offers a complete cross-sectional view of the LV myocardium, providing complementary information to apical views. The limited use of quantitative PSAX measurements is largely due to time-consuming manual workflows and inter-observer variability. Automated methods could help overcome these barriers and facilitate broader use of PSAX-based measurements in clinical practice. Purpose To develop and evaluate a fully automated deep learning (DL) method for quantification of LV morphology and function from PSAX images. Methods Two clinicians manually annotated the LV lumen and myocardium in PSAX images from 384 subjects for model development and 47 subjects for evaluation. A segmentation model based on nnU-Net was trained to delineate the LV lumen and myocardium. Model performance was assessed using Dice coefficients and the 95th percentile Hausdorff distance (HD95). Based on the segmentations, we calculated the cross-sectional lumen area and mean wall thickness (MWT) at end-diastole (ED) and end-systole (ES), LV fractional area change (LVFAC), and global circumferential strain (GCS). Agreement between DL-based and clinician-derived measurements was assessed and compared to inter-observer variability. We compared MWT at ED between subjects with hypertension (systolic blood pressure ≥ 160 mmHg or use of antihypertensive medication) and those without hypertension. Results The nnU-Net model achieved Dice coefficients of 0.94±0.03 and 0.86±0.07, and HD95s of 3.0±1.6 mm and 3.3±1.5 mm, for the lumen and myocardium, respectively. The nnU-Net outperformed inter-observer Dice (0.93±0.04 and 0.82±0.07) and HD95 (3.7±1.5 mm and 4.2±1.4 mm). Feasibility of the measurement pipeline was 90%. Agreement between DL-based measurements and observers was similar for GCS. For all other measurements, DL−observer agreement exceeded inter-observer agreement. For MWT ED the agreement with the mean of two observers (bias -0.4 mm and limits of agreement (LoA) -1.8 to 1.0 mm) was in line with the inter-observer comparison (bias 1.7 mm and LoA -0.2 to 3.5 mm). MWT at ED was significantly higher in subjects with hypertension than in those without (p < 0.001). Conclusion The proposed DL-based measurement pipeline for PSAX enables automatic measurements of LV cross-sectional area, wall thickness and subsequent functional measurements. Agreement with manual measurements was high, and the method was able to detect structural differences associated with hypertension. These results highlight the potential of DL to unlock the value of PSAX analysis in comprehensive LV assessment.Agreement analysis (bias and LoA) Boxplot of MWT ED vs. Hypertension