Deep learning enhanced ultra-fast SPECT/CT bone scan: quantitative assessment and clinical performance

Abstract

Abstract Background To evaluate clinical performance of deep learning enhanced ultra-fast SPECT/CT bone scan. Methods One hundred and two patients were enrolled in this retrospective study. The probable malignant tumor sites continuously underwent a 20min SPECT/CT and a 3min SPECT scan. A deep learning model was applied to generate algorithm-enhanced images (3min-DL SPECT). Two reviewers evaluated general image quality, 99mTc-MDP distribution, artifacts, and diagnostic confidence independently. The sensitivity, specificity, accuracy, and inter-observer agreement were calculated. Linear regression was analyzed for lesion SUVmax between 3min-DL and 20min SPECT. Peak signal-to-noise ratio (PSNR), image similarity (SSIM) were evaluated. Results The general image quality, 99mTc-MDP distribution, artefact, and diagnostic confidence of 3min-DL images were significantly superior to those of 20min images (P < 0.0001). The sensitivity, specificity and accuracy of 20min and 3min-DL SPECT/CT had no difference by both reviewers (0.903 vs 0.806, 0.873 vs 0.873, 0.882 vs 0.853; 0.867 vs 0.806, 0.944 vs 0.936, 0.912 vs 0.920, P > 0.05). The diagnosis results of 20min and 3min-DL images showed a high inter-observer agreement (Kappa = 0.822, 0.732). PSNR and SSIM of 3min-DL images were significantly higher than 3min images (51.44 vs 38.44, 0.863 vs 0.752, P < 0.05). A strong linear relationship was found between the SUVmax of 3min-DL and 20min images (r = 0.987; P < 0.0001). Conclusion An ultra-fast SPECT/CT with 1/7 scan time could be enhanced by deep learning method to have competitive image quality and equivalent diagnostic value to those of standard acquisition.