Machine learning based quantification of extrapulmonary diseases in chest CT

Abstract

In several countries, lung cancer screening programs are being implemented, in which heavy cigarette smokers are regularly invited for a CT scan of their lungs (chest CT). This scan can help detect lung cancer in an early stage, when treatment is more effective. However, chest CT scans also allow for quantification of various other pulmonary and extrapulmonary diseases. Cardiovascular disease risk is related to the amount of arterial calcification. Manual quantification of the amount of arterial calcification in CT scans is often a tedious task, especially in chest CT scans, which are not optimized for cardiac imaging and therefore frequently contain cardiac motion artifacts. We therefore developed a method for automatic calcification quantification in chest CT scans using convolutional neural networks. A network using dilated convolutions, a special type of convolution that enables the network to include information from a large area of the image, searches the CT image for potential calcifications and determines their anatomical location, i.e., labels them according to the calcified artery or valve. A second network with conventional convolutions, and hence focused on a smaller area of the image, subsequently analyzes the detected calcification candidates to discard false-positive detections of the first network. This method enabled a retrospective case-control study with data from a large lung cancer screening trial. We investigated sex differences in the prevalence and severity of arterial calcification, and whether there are differences in the association of calcium scores and cardiovascular mortality between male and female heavy smokers. The study confirmed that calcification of the coronary arteries is much more common and severe in men, but revealed no such prominent difference for calcification of the thoracic aorta. Osteoporosis risk is related to characteristics of the vertebrae, such as the bone density. First steps toward analysis of the vertebrae are their detection and segmentation in the image, and their anatomical identification, for which we developed an automatic method based on an iteratively applied convolutional neural network. This method searches the image in a sliding-window fashion until it finds the spine, then traverses along the spine to segment the vertebrae one after the other. The vertebrae are additionally anatomically labeled and examined for completeness, since in many types of scans several vertebrae are only partially visible. Using a global maximum likelihood model, the anatomical labels of all vertebrae in a scan are used to find the most plausible anatomical labeling. Markers for osteoporosis and osteoporotic fracture risk are often measurements of only part of the vertebra, the vertebral body. Therefore, we also developed a method for partitioning of segmentation masks into two substructures and applied this to partitioning of vertebra segmentations into the vertebral body and the posterior elements. The method uses thin-plate splines to model the boundary between the two structures. A convolutional neural network analyzes the segmentation mask and predicts the location of control points defining the thin-plate spline surface.