Automated segmentation of an intensity calibration phantom in clinical CT images using a convolutional neural network

Abstract: Purpose: To apply a convolutional neural network (CNN) to develop a system that segments intensity calibration phantom regions in computed tomography (CT) images, and to test the system in a large cohort to evaluate its robustness. Methods: A total of 1040 cases (520 cases each from two institutions), in which an intensity calibration phantom (B-MAS200, Kyoto Kagaku, Kyoto, Japan) was used, were included herein. A training dataset was created by manually segmenting the regions of the phantom for 40 cases (20 cases each). Segmentation accuracy of the CNN model was assessed with the Dice coefficient and the average symmetric surface distance (ASD) through the 4-fold cross validation. Further, absolute differences of radiodensity values (in Hounsfield units: HU) were compared between manually segmented regions and automatically segmented regions. The system was tested on the remaining 1000 cases. For each institution, linear regression was applied to calculate coefficients for the correlation between radiodensity and the densities of the phantom. Results: After training, the median Dice coefficient was 0.977, and the median ASD was 0.116 mm. When segmented regions were compared between manual segmentation and automated segmentation, the median absolute difference was 0.114 HU. For the test cases, the median correlation coefficient was 0.9998 for one institution and was 0.9999 for the other, with a minimum value of 0.9863. Conclusions: The CNN model successfully segmented the calibration phantom's regions in the CT images with excellent accuracy, and the automated method was found to be at least equivalent to the conventional manual method. Future study should integrate the system by automatically segmenting the region of interest in bones such that the bone mineral density can be fully automatically quantified from CT images.