Aims
To develop an automated method for bloodpool segmentation and imaging plane re-slicing of cardiac CT via deep learning (DL) for clinical use in coronary artery disease (CAD) wall motion assessment and reproducible longitudinal imaging.
Methods and Results
100 patients who underwent clinically indicated cardiac CT scans with manually segmented left ventricle (LV) and left atrial (LA) chambers were used for training. For each patient, long-axis (LAX) and short-axis (SAX) planes were manually defined by an imaging expert. A DL model was trained to predict bloodpool segmentations and imaging planes. DL bloodpool segmentations showed close agreement with manual LV (median Dice: 0.91, Hausdorff distance: 6.18mm) and LA (Dice: 0.93, HD: 7.35mm) segmentations and strong correlation with manual EF (Pearson r: 0.95 LV, 0.92 LA). Predicted planes had low median location (6.96mm) and angular orientation (7.96 °) errors which were comparable to inter-reader differences (p > 0.71). 84 – 97% of DL-prescribed LAX planes correctly intersected American Heart Association (AHA) segments, which was comparable (p > 0.05) to manual slicing.
In a test cohort of 144 patients, we evaluated the ability of the DL approach to provide diagnostic imaging planes. Visual scoring by two blinded experts determined ≥94% of DL-predicted planes to be diagnostically adequate. Further, DL enabled visualization of LV wall motion abnormalities due to CAD and provided reproducible planes upon repeat imaging.
Conclusion
A volumetric, DL approach provides multiple chamber segmentations and can re-slice the imaging volume along standardized cardiac imaging planes for reproducible wall motion abnormality and functional assessment.
Nets (1), can perform semantic segmentation of medical images, including three-dimensional (3D) CT data (2). However, 3D image volumes are challenging to analyze given their large size and their associated large memory footprint, which limits the implementation of 3D architectures on commercially available graphics processing units (GPUs) to shallow designs (3,4). Although previous approaches have analyzed image volumes as a series of twodimensional sections (5-7), this process irrevocably disrupts the 3D content and/or information. Further, 3D approaches downsample and/or crop image volumes to keep 3D information intact (3,4). However, cropping reduces the extent of information available to the network and can lead to multiple inferences that may require postprediction combination, whereas downsampling is a lossy operation that decreases image fidelity, thereby limiting the accuracy of the resulting segmentation (8). Although alternative image representations, such as point clouds (9,10) and sparse tensors (11) exist, there is no obvious way to apply CNNs to these data structures.In this study, we explored an octree-based representation for 3D CT images that provides high data compression without sacrificing 3D content or spatial resolution. The approach maintains a grid structure that enables application of convolution-based neural network architectures. To adapt the framework to medical imaging, we introduced an intensity tolerance parameter in the octree subdivision algorithm to govern image compression. We found that across a range of compression levels, the octree-based representation preserved image and segmentation features better than spatial downsampling. Further, the octree representation enabled semantic segmentation with a 3D U-Net architecture at the native image resolution, which improved segmentation accuracy, especially at the object border. We demonstrated these findings in semantic segmentation of left ventricular (LV) and left atrial (LA) blood pools on clinical cardiac CT angiograms.
Materials and Methods
DatasetElectrocardiographically gated end-systolic and enddiastolic cardiac CT angiographic studies from 100 patients obtained between
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.