Segmentation Ability of Pulmonary Nodules Using Deep Learning in Dual-Energy Subtraction Images

In recent years, computer-aided detection (CAD) systems for chest X-ray (CXR) images have been promoted to decrease the number of deaths from lung cancer and to overcome the shortage of radiologists. Dual-energy subtraction (DES) images of pulmonary nodules have been reported to provide higher detection rates than CXR, but DES has the disadvantage of increasing the radiation dose. Therefore, in this study, DES images were obtained using an entrance surface dose half of that used in CXR by incorporating a noise reduction process. The automatic segmentation of pulmonary nodules, one of the important steps of CAD, was performed using deep learning techniques. Segmented images of the pulmonary nodules were created for use with U-net, one of the deep learning techniques, and the detection accuracy of the pulmonary nodules was compared for CXR and DES images. The data for the U-net were prepared from 348 CXR images and from the same number of DES images obtained using artificial nodules placed on a commercially available anthropomorphic lung phantom. Since noise reduction was important for improving pulmonary nodule segmentation in DES images, the DES images were pre-processed with a bilateral filter, a non-local means filter, and an anti-correlated noise reduction (ACNR) algorithm. In the artificial nodule segmentation, the recall and the precision were higher in the DES images with noise reduction processes than in the CXR images. These results suggest the usefulness of CAD in DES images taken at an entrance surface dose of half the CXR dose.