Investigating the Feasibility of Using DenseNet to Improve Coronary Calcification Detection in CT

Interscan reproducibility of coronary artery calcium (CAC) scoring can be improved by using a smaller slice thickness but at the cost of higher image noise. This study aimed to investigate the feasibility of using densely connected convolutional network (DenseNet) to reduce the image noise in CAC scans reconstructed with slice thickness < 3 mm for improving coronary calcification detection in CT.

Phantom data acquired with QRM and CIRS phantoms were used for model training and testing, where the DenseNet model adopted in this work was a convolutional neural network (CNN) designed for super resolution recovery. After phantom study, the proposed method was evaluated in terms of its ability to improve calcification detection using patient data. The CNN input images (IMG_input) were CAC scans reconstructed with 0.5-, 1.0- and 1.5-mm slice thickness, while CNN label images were CAC scans reconstructed with 3-mm slice thickness (IMG^3mm). Region of interest (ROI) analysis was carried out on IMG^3mm, IMG_input and CNN output images (IMG_output). Two-sample t test was used to compare the difference in Hounsfield Unit (HU) values within ROI between IMG^3mm and IMG_output.

For the calcifications in QRM phantoms, no statistically significant difference was found when comparing the HU values of 400- and 800-HA calcifications identified on IMG^3mm to those on IMG_output with slice thickness of 0.5, 1.0 or 1.5 mm. On the other hand, statistically significant difference was found when comparing the HU values of 200-HA calcifications identified on IMG^3mm to those on IMG_output with a slice thickness of 0.5 and 1.0 mm. Meanwhile, no statistically significant difference was found when comparing the HU values of 200-HA calcifications identified on IMG^3mm to those on IMG_output with a slice thickness of 1.5 mm. As for the rod inserts in CIRS phantoms simulating 9 different tissue types in human body, there was no statistically significant difference between IMG^3mm and IMG_output with slice thickness of 1.5 mm, and all the p values were larger than 0.10. With regards to patient study, more calcification pixels were detected on IMG_output with a slice thickness of 1.5 mm than on IMG^3mm, so calcifications were more clear on the denoised images.

According to our results, the CNN-based denoising method could reduce statistical noise in IMG_input with a slice thickness of 1.5 mm without causing significant texture change or variation in HU values. The proposed method could improve cardiovascular risk prediction by detecting small and soft calcifications that are barely identified on 3-mm slice images used in conventional CAC scans.