Deep Learning-based Fully Automated Scan Range Detection in Chest CT Imaging

• Main Authors: Salimi, Y., Akhavanallaf, A., Shiri, I., Mansouri, Z., Saberimanesh, A., Sanaat, A., Pakbin, M., Askari, D., Sandoughdaran, S., Sharifipour, E., Arabi, H., Zaidi, H.
• Format: Conference Proceeding
• Language: English
• Subjects: chest imaging; Computed tomography; CT dosimetry; Deep learning; Lung; Neural networks; Prediction algorithms; Radiation effects; scan range; Three-dimensional displays
• ISSN: 2577-0829
• DOI: 10.1109/NSS/MIC44867.2021.9875808

This study aimed to develop an automated scan range selection for minimal patient irradiation in CT examinations within the chest region. A total number of 20,820 chest CT images acquired for various indications were collected, the 3D lung masks were generated using a Deep Neural Network (DNN) developed by our group. Consequently, 2D projected localizer images and masks were computed in lateral (lat) and anterior-posterior (AP) directions. We developed a deep learning algorithm to predict the lung mask from 2D localizer images. Thereby, the scan range is automatically determined without the technologist's intervention. Lastly, the impact of over-ranging on patients' effective dose was investigated through personalized dosimetry of the given cohort. A significant over-scanning range (31±24 mm) was observed in the clinical setting for more than 95% of cases. The average Dice coefficient for 2D lung segmentation was 0.96 and 0.97 for AP and lateral projections, respectively. The proposed approach resulted in errors of 0.08±1.46 and -1.5±4.1 mm in the superior and inferior directions, respectively. The effective dose (ED) was reduced by 21% in the unseen external dataset when using the proposed automated scan range selection.