Accurate fully automated assessment of left ventricle, left atrium, and left atrial appendage function from computed tomography using deep learning.

Aims: Assessment of cardiac function is essential for diagnosis and treatment planning in cardiovascular disease. Volume of cardiac regions and the derived measures of stroke volume (SV) and ejection fraction (EF) are most accurately calculated from imaging. This study aims to develop a fully automatic deep learning approach for calculation of cardiac function from computed tomography (CT).

Methods and results: Time-resolved CT data sets from 39 patients were used to train segmentation models for the left side of the heart including the left ventricle (LV), left atrium (LA), and left atrial appendage (LAA). We compared nnU-Net, 3D TransUNet, and UNETR. Dice Similarity Scores (DSS) were similar between nnU-Net (average DSS = 0.91) and 3D TransUNet (DSS = 0.89) while UNETR performed less well (DSS = 0.69). Intra-class correlation analysis showed nnU-Net and 3D TransUNet both accurately estimated LVSV (ICC_nnU-Net = 0.95; ICC_3DTransUNet = 0.94), LVEF (ICC_nnU-Net = 1.00; ICC_3DTransUNet = 1.00), LASV (ICC_nnU-Net = 0.91; ICC_3DTransUNet = 0.80), LAEF (ICC_nnU-Net = 0.95; ICC_3DTransUNet = 0.81), and LAASV (ICC_nnU-Net = 0.79; ICC_3DTransUNet = 0.81). Only nnU-Net significantly predicted LAAEF (ICC_nnU-Net = 0.68). UNETR was not able to accurately estimate cardiac function. Time to convergence during training and time needed for inference were both faster for 3D TransUNet than for nnU-Net.

Conclusion: nnU-Net outperformed two different vision transformer architectures for the segmentation and calculation of function parameters for the LV, LA, and LAA. Fully automatic calculation of cardiac function parameters from CT using deep learning is fast and reliable.