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

Abstract

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. MethodsA 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. ResultsAfter 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 4 other, with a minimum value of 0.9863. ConclusionsThe 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. The source code and the model used for segmenting the phantom are open and can be accessed via https://github.com/keisuke-

Development and validation of an open-source tool for opportunistic screening of osteoporosis from hip CT images

“…The accuracy of femoral segmentation was reported as a mean Dice coefficient 22 of 0.985 (standard deviation (SD) 0.0065) and a mean symmetric surface distance (MSD) 23 of 0.175 mm (SD 0.084), 21 whereas the accuracy of the phantom segmentation was reported as a median Dice coefficient of 0.977 (IQR 0.963 to 0.986) and a MSD of 0.116 mm (IQR 0.084-0.193). 20 Use of phantom segmentation enabled determination of a linear correlation equation between the mean Hounsfield units (HUs) and the known densities of the phantom (0, 50, 100, 150, and 200 mg/cm 3 in hydroxyapatite). Subsequently, using the correlation equation, the HUs of each voxel were converted into density (mg/cm 3 ).…”

“…On the basis of a previous report, deep learning was used to develop a system to measure the BMD of the proximal femur. 17 Specifically, models previously reported to segment the calibration phantom and the femur were used, 17,20 and a new model to detect the landmarks of the proximal femur was developed and used (Figure 1). Femur and phantom segmentation from CT images.…”

Bone mineral density estimation from a plain X-ray image by learning decomposition into projections of bone-segmented computed tomography

“…Depending on the availability of data, the I2I model can be trained in a paired (Isola et al, 2017) and unpaired (Zhu et al, 2017) manner while the paired-trained model usually learns faster and better. Our method appreciated the paired (Uemura et al, 2021), bone segmentation (Hiasa et al, 2020), 2D-3D registration to the X-ray image (Otake et al, 2012) and DRR generation by projecting QCT.…”

“…5 visualized the T-SNE of the collected multi-pose X-ray images of dataset B, which includes poses of standing, adduction, abduction, and supine, suggesting high potential for conducting multiple clinical validations. The calibration phantom (B-MAS200, Kyoto Kagaku, Kyoto, Japan) (Uemura et al, 2021), which is used to convert radiodensity [in Hounsfield units] to the ground truth QCT-BMD (in mg/cm 3 ), contains known densities of hydroxyapatite Ca 10 (PO 4 ) 6 (OH) 2 . All CT images used in this study were obtained using the OptimaCT660 scanner (GE Healthcare Japan, Tokyo, Japan), and all DXA images of the proximal femur were acquired for the operative side (Discovery A, Hologic Japan, Tokyo, Japan) to obtain the ground truth DXA-BMD.…”