Deep learning-based detection of functionally significant stenosis in coronary CT angiography

Abstract

Patients with intermediate anatomical degree of coronary artery stenosis require determination of its functional significance. Currently, the reference standard for determining the functional significance of a stenosis is invasive measurement of the fractional flow reserve (FFR), which is associated with high cost and patient burden. To address these drawbacks, FFR can be predicted non-invasively from a coronary CT angiography (CCTA) scan. Hence, we propose a deep learning method for predicting the invasively measured FFR of an artery using a CCTA scan. The study includes CCTA scans of 569 patients from three hospitals. As reference for the functional significance of stenosis, FFR was measured in 514 arteries in 369 patients, and in the remaining 200 patients, obstructive coronary artery disease was ruled out by Coronary Artery Disease-Reporting and Data System (CAD-RADS) category 0 or 1. For prediction, the coronary tree is first extracted and used to reconstruct an MPR for the artery at hand. Thereafter, the coronary artery is characterized by its lumen, its attenuation and the area of the coronary artery calcium in each artery cross-section extracted from the MPR using a CNN. Additionally, characteristics indicating the presence of bifurcations and information indicating whether the artery is a main branch or a side-branch of a main artery are derived from the coronary artery tree. All characteristics are fed to a second network that predicts the FFR value and classifies the presence of functionally significant stenosis. The final result is obtained by merging the two predictions. Performance of our method is evaluated on held out test sets from multiple centers and vendors. The method achieves an area under the receiver operating characteristics curve (AUC) of 0.78, outperforming other works that do not require manual correction of the segmentation of the artery. This demonstrates that our method may reduce the number of patients that unnecessarily undergo invasive measurements.

Keywords: convolutional neural networks; coronary artery tree; coronary computed tomography angiography; fractional flow reserve; transformer.

Figure 1

Data included in the study.

Figure 2

Overview of our method for assessing the presence of a functionally significant stenosis in a coronary artery. From the CCTA scan, we extract a coronary artery centerline tree. For each artery we generate an MPR that is further analyzed to predict the lumen area, its average attenuation and the calcium area per centerline point. These characteristics, as well as characteristics indicating the presence of bifurcations and whether the artery is a main branch or to a side branch of a main artery, are fed to the classification network to determine the presence of a functionally significant stenosis in the artery.

Figure 3

Pre-processing steps.

Figure 4

Architecture of the network for extracting artery characteristics. Stacks of 3 cross-sectional artery slices are fed to a 2D CNN with 4 pooling layers interleaved with convolutions. The network is trained to predict the lumen area, its average attenuation and the calcium area for the central slice of the 3 input slices.

Figure 5

Architecture of the network used for stenosis assessment. The lumen area and its attenuation predicted by the characterization network are first pre-encoded and subsequently concatenated with the calcium area, and with additional characteristics indicating bifurcations and whether the analysis is performed in the main- or side-branch of the artery. The combined encodings are thereafter fed to the encoder. In the encoder, the features are first pooled and thereafter, convolutions and a transformer layer are applied. For final classification, two separate output heads are applied. In the regression head, two convolutional layers and the ReLU activation function are used. The resulting sequence is pooled along the artery dimension and subtracted from 1 to yield a single FFR value. In the classification head, the features are pooled to a fixed length of 5 (2.5 mm). Thereafter, two dense layers are used in combination with the sigmoid activation function to yield output probabilities for the presence of a functionally significant stenosis in the artery.

Figure 6

Transformation from predicted FFR values to pseudo probabilities.

Figure 7

Top: Scatter plots relating the invasively measured FFR with the predictions for each artery from the Test_Cath data set. The graph on the left-hand side corresponds to the merged output probability, in the graph in the middle the output probability from the classification head is shown and in the graph on the right-hand side the regressed FFR value is depicted. Points are colored in red according to their prediction uncertainty. Background colors indicate in which arteries the functional significance was assessed correctly (white) or incorrectly (gray). Whereas, for probabilities (left and middle), high values correspond to the positive class, for regression (right), low output values correspond to the positive class. Black lines show the linear fit to the data. Bottom: MPRs and predicted characteristics for two arteries (positions in scatter plots indicated by blue circles). The location of the annotated lesion is plotted in green. Whereas the merged probability assigned to the artery on the left corresponds to the correct class, for the artery on the right, output of the classification head was strongly negative (low probability for functionally significant stenosis), which when combined with the regressed FFR caused the merged probability to yield the incorrect class. Incorrect output of the classification head may be related to a visually minor step-and-shot artifact causing a low intensity section in the MPR on the right (indicated by arrow).