1348985 - Deep Learning Estimation of Three-Dimensional Left Atrial Shape from Two-Chamber and Four-Chamber Cardiac Long Axis Views

We developed a deep learning neural network to infer 3D LA shape, volume and surface area from two-chamber (2CH) and four-chamber (4CH) CMR image segmentations, through label-to-label mapping from incomplete input data [4, 5]. The pipeline is shown in Figure 1. A sparse 3D volume label map was generated from the 2CH and 4CH segmentations, by labelling only voxels on the image plane. A 3D UNet was trained to convert the sparse label map into a dense 3D label map with all voxels labelled.

We trained and tested the network using simulated 2CH and 4CH segmentations and accurate ground truth obtained from 3D coronary computed tomography angiography (CCTA) segmentations [5], from the SCOT-HEART study (n=1700, separated into 1400, 100 and 200 cases for training, validating and testing respectively). We also evaluated the method using an independent external cohort of patients with paired CMR and CCTA acquisitions (mean 2.5 months apart), in patients with suspected angina attributable to coronary artery disease, using the CMR 2CH and 4CH as input and the CCTA as the ground truth (n=20). 3D shape errors where quantified using Dice scores, and errors in volume and surface area were calculated between CCTA and the network, as well as the bi-plane area-length method. CCTA ground truth for volume was estimated using voxel summation, while surface area was estimated from a marching cubes surface representation computed on the voxel data.

Results: Demographics for the CCTA training/testing cohort and the external CMR testing cohort are shown in Table 1. The network achieved a mean Dice score value of 93.7%, for the CCTA test cases and 87.4% for the CMR test cases, showing excellent reconstruction of 3D shape. Table 2 shows volume and surface area errors relative to the CCTA ground truth. The network achieved much lower bias (mean error) and variation (standard deviation of error) than the area-length method. Although errors in the external CMR dataset were higher than for the CCTA test cases, this is partly due to the difference in loading and time interval between the CMR and CCTA acquisitions.

Conclusion:

The network showed higher accuracy and robustness than the area-length method for both simulated and real CMR segmentations. Also, it can be applied to any imaging modality since it relies only on segmentations. In the future, the combination with automatic CMR segmentation will enable fully automated real-time volumetric LA evaluation from 2CH and 4CH CMR views.