Abstract
3034
Introduction: Bone scan has been performed by visual evaluation; however quantitative evaluation has become possible because of bone scan index (BSI). BSI is an imaging biomarker which evaluates bone metastasis burden in bone scan. BSI is applied to prognosis prediction and treatment response. BSI is measured using segmented bone and extracted bone metastasis. It is important to perform skeletal segmentation for measuring BSI. Skeletal segmentation methods have developed an atlas-based method and a deep learning-based method. In this study, we compared the accuracy of two skeletal segmentation methods.
Methods: We conducted 383 prostate cancer patients. These patients were divided into two groups: 175 patients were injected with Technetium-99m methylene diphosphonate (99mTc-MDP; FUJIFILM Toyama Chemical) and 208 patients were injected with 99mTc hydroxy methylene diphosphonate (99mTc-HMDP; Nihon Medi-Physics). Bone scan was performed from 1.9 to 4.8 h after an intravenous injection from 725.9 to 1097.4 MBq. Whole-body images were acquired with Symbia Intevo 16 (Siemens Healthineers), E.CAM (Canon Medical Systems), and Infinia 3 (GE Healthcare). All gamma cameras were equipped with low-energy high-resolution parallel-hole collimator. The matrix size was 256 × 1024. The energy peak was centered at 140 keV with a 15 % window. Scan speed of whole-body images was 15 to 20 cm per minute according to uptake time. Skeletal segmentations were performed by BONENAVI (FUJIFILM Toyama Chemical) which used an atlas-based method for patients injected 99mTc-MDP and by VSBONE BSI (Nihon Medi-Physics) which used a deep learning-based method for patients injected 99mTc-HMDP, respectively. Three technologists classified skeletal segmentation using correct (almost consistent with bone morphology) or not correct (disagreement with bone morphology, such as distortion, overestimation or underestimation) by following these regions: skull, cervical vertebrae, thoracic vertebrae, lumbar vertebrae, pelvis, sacrum, left and right humerus, left and right rib, sternum, left and right clavicle, left and right scapula, left and right femur, and outside of regions. Segmentation error was defined when more than two technologists classified “not correct” at least for one region.
Results: Segmentation error showed 58 cases (33.1%) in BONENAVI and 49 cases (23.6%) in VSBONE BSI. Segmentation errors were more common in the pelvis and the axial skeletons such as skull and rib for BONENAVI, while segmentation errors were more common in the long tube bones such as humerus, femur, and clavicle for VSBONE BSI (Figure 1). The skeletal segmentations were also performed outside of regions in VSBONE BSI (Figure 2).
Conclusions: We compared two skeletal segmentation accuracy for an atlas-based method and a deep learning-based method. Segmentation errors were specific in each skeletal segmentation method. It suggested that each skeletal segmentation method should be used with understanding of its characteristics.