Cardiac Computed Tomography Angiography Plane Prediction and Comprehensive LV Segmentation

Abstract

The use of cardiac computed tomography angiography (CCTA) has dramatically increased over the past decade, with an increasingly recognized role for functional assessment; however, reformatting these datasets into standard cardiac planes and performing quantitative analysis remains time consuming and disruptive to clinical workflows. Here, we propose a fully automated, volumetric, end-to-end trained network for simultaneous detection of standard cardiac planes and comprehensive left ventricular (LV) segmentation in the predicted short axis coordinate system. The architecture consists of a coarse segmentation module, a transformation module, and a fine segmentation module. The coarse segmentation module provides an initial segmentation of the full field of view (FOV) axial images at low resolution. The transformation module predicts the rotations corresponding to the standard cardiac planes (short axis, SAX; two chamber, 2CH; three chamber, 3CH; and four chamber, 4CH) and reformats the source volume into the predicted SAX coordinate system at high resolution. Finally, the fine segmentation module segments the narrow FOV, high resolution SAX volume. The dataset consisted of 313 CCTA studies partitioned into training, validation, and testing in an 80:10:10 split. Architectural decisions are justified using ablation experiments. On the test set, the proposed architecture achieved accurate plane predictions (mean angle errors of $9.1\pm6.2^\circ$, $9.5\pm5.4^\circ$, $9.0\pm5.9^\circ$, and $8.8\pm5.9^\circ$ for the SAX, 2CH, 3CH, and 4CH planes, respectively) and high quality segmentations (Dice scores of $0.955\pm0.008$, $0.928\pm0.016$, and $0.808\pm0.029$ for the bloodpool, myocardium, and trabeculations, respectively). This fully automated pipeline has the potential to replace current manual workflows, expediting the availability of standard cardiac planes and quantitative analysis for clinical interpretation.