Automatic calcium scoring in low-dose chest CT using deep neural networks with dilated convolutions (1711.00349v2)

Abstract: Heavy smokers undergoing screening with low-dose chest CT are affected by cardiovascular disease as much as by lung cancer. Low-dose chest CT scans acquired in screening enable quantification of atherosclerotic calcifications and thus enable identification of subjects at increased cardiovascular risk. This paper presents a method for automatic detection of coronary artery, thoracic aorta and cardiac valve calcifications in low-dose chest CT using two consecutive convolutional neural networks. The first network identifies and labels potential calcifications according to their anatomical location and the second network identifies true calcifications among the detected candidates. This method was trained and evaluated on a set of 1744 CT scans from the National Lung Screening Trial. To determine whether any reconstruction or only images reconstructed with soft tissue filters can be used for calcification detection, we evaluated the method on soft and medium/sharp filter reconstructions separately. On soft filter reconstructions, the method achieved F1 scores of 0.89, 0.89, 0.67, and 0.55 for coronary artery, thoracic aorta, aortic valve and mitral valve calcifications, respectively. On sharp filter reconstructions, the F1 scores were 0.84, 0.81, 0.64, and 0.66, respectively. Linearly weighted kappa coefficients for risk category assignment based on per subject coronary artery calcium were 0.91 and 0.90 for soft and sharp filter reconstructions, respectively. These results demonstrate that the presented method enables reliable automatic cardiovascular risk assessment in all low-dose chest CT scans acquired for lung cancer screening.

Automatic Calcium Scoring in Low-Dose Chest CT Using Deep Neural Networks with Dilated Convolutions

The paper presents a novel method for the automatic detection of calcifications in the coronary arteries, thoracic aorta, and cardiac valves in low-dose chest CT. By employing deep learning techniques, specifically convolutional neural networks (CNNs) with dilated convolutions, the authors address the challenge of detecting cardiovascular disease risks in heavy smokers undergoing lung cancer screenings. The innovative use of a two-stage network architecture allows for the precise identification and spatial labeling of calcifications based on their anatomical locations within chest CT images.

Methodology

The authors utilized a two-stage CNN approach. The first network is trained to identify potential calcifications and classify them by anatomical location, using dilated convolutions to expand the receptive field. This design eliminates the need for explicit segmentation or localization of anatomical structures within the image, a common requirement of prior methods. The second network refines these findings by distinguishing true calcifications from false positives, relying on local image information rather than broader spatial context.

The dataset, consisting of 1744 CT scans from the National Lung Screening Trial, provided a robust platform for evaluating the method. Scans were selected to encompass a wide variety of imaging parameters and conditions, ensuring the generalizability of the method. Notably, the method was tested on both soft and sharp image reconstructions to ensure its applicability across different acquisition conditions.

Results

The method demonstrated high reliability in detecting calcifications. The F$_1$ scores in the detection of coronary artery calcifications were \num{0.89} for soft reconstructions and \num{0.84} for sharp reconstructions, showcasing the system's adaptability to various scan qualities. For thoracic aorta and cardiac valve calcifications, the results were promising though varied more with reconstruction sharpness, with F$_1$ scores ranging from \num{0.67} to \num{0.66}.

A crucial outcome was the assignment of cardiovascular risk categories based directly on CAC scores, where the method achieved high kappa coefficients (\num{0.91} for soft and \num{0.90} for sharp reconstructions). These scores are indicative of a method capable of integrating into routine screening workflows to assess cardiovascular risks without additional imaging.

Discussion and Implications

The achievement of automated cardiovascular risk assessment using low-dose CT scans presents significant implications. It suggests a pathway for enhancing lung cancer screening programs by concurrently evaluating an individual's cardiovascular risk, thereby addressing the broader health implications of heavy smoking. The robustness of this method in diverse imaging settings positions it as a potential supplementary tool in clinical practice, aiding radiologists in efficiently managing cases with high calcification burdens.

Future research could aim to further improve the network by unifying the two stages into a single network to reduce complexity and computational demand. Moreover, continuing to expand the dataset used for training and validation, encompassing more varied population samples, could augment the method's robustness further. Additionally, investigating the clinical impact of incorporating calcium scoring into routine screenings will help uncover substantive benefits within public health strategies.

In conclusion, this paper contributes a sophisticated methodology for enhancing the capabilities of lung cancer screening programs through the automatic assessment of cardiovascular risk. It represents a substantial step towards the efficient utilization of existing CT imaging for broader health diagnostics in high-risk populations.