Pulmonary vascular disease has a differential effect on pulmonary arteries and veins and automated separation of these on imaging reconstructions are crucial to accurate analysis of disease impact on structure as manual separation of these is very time-consuming process. We have developed a fully automatic approach based on Artificial Intelligence (AI) that classifies vessels from chest CT images into arteries and veins and provides their radius.
For the artery/vein (A/V) classification of pulmonary vasculature, the a priori probability for a vessel to be either an artery or a vein is extracted by a 3D convolutional neural network (CNN) on patches of 32x32x5 voxels, followed by a graph cuts (GC) optimization to ensure fully connected branches. A 2D CNN, trained on synthetic patches of 32x32 pixels, is then used to measure the A/V size (in mm). The A/V technique was validated by comparing the automatic approach against manually labeled vascular trees in 18 subjects with classification accuracy, specificity and sensitivity, and with an analysis by COPD and emphysema status. For the A/V vessel radius size, we computed the relative error (RE) on 200,000 synthetic patches and analyzed the intra-class correlation coefficient (ICC) in 50 subjects from the COPDGene Phase 2 study when varying kernel, FOV, dosage, and reconstruction.
The accuracy, sensitivity and specificity for the A/V classification were 93.6±4.9, 97.3±2.5 and 89.5±9.2. For all metrics, there were no statistically significant differences in performance by COPD status or emphysema diagnosis (Table 1). The RE obtained for vessel radii was 2.25%. The ICC for kernel, FOV, dosage, and reconstruction was 99.0%, 99.9%, 95.3%, and 96.4%.
CNNs enable artery-vein separation and highly accurate vessel sizing in CT scans without contrast. The method is robust and shows very good performance without being affected by disease status or scan parameters.