TY - JOUR T1 - Deep learning-based detection of bone lesions in <sup>99m</sup>Tc-MDP SPECT: comparison with human observers<strong/> JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 510 LP - 510 VL - 61 IS - supplement 1 AU - Yoann Petibon AU - Frederic Fahey AU - Xinhua Cao AU - Zakhar Levin AU - Briana Sexton-Stallone AU - Anthony Falone AU - Katherine Zukotynski AU - Neha Kwatra AU - Ruth Lim AU - Zvi Bar-Sever AU - S. Ted Treves AU - Georges El Fakhri AU - Jinsong Ouyang Y1 - 2020/05/01 UR - http://jnm.snmjournals.org/content/61/supplement_1/510.abstract N2 - 510Objectives: Low back pain (LBP) is a common symptom in the pediatric population, with an especially high incidence among young athletes. 99mTc-MDP SPECT is well-established in the diagnosis of lumbar stress injury, a major cause of LBP in young patients. However, the low quality of 99mTc-MDP images hampers lesion detectability and leads to a high interobserver variability, meaning that diagnosis greatly varies depending upon the reviewer and their level of expertise and experience. A computer-aided diagnostic system could help improve lesion detection and reduce variability by providing lesion localization assistance to physicians, ultimately leading to improved quality of care. We developed an approach based on a deep convolutional neural network (CNN) to detect lumbar lesions in 99mTc-MDP SPECT and compared its performance to that of nuclear medicine physicians using ROC and localization ROC (LROC) analyses. Methods: The strategy used to carry out CNN training and observer studies relied on the incorporation of artificial lesions into clinical lesion-absent (LA) 99mTc-MDP scans. This has several advantages: (1) a small number of clinical datasets can be used to build an almost unlimited number of labeled sets for CNN training; (2) it ensures perfect knowledge of the presence and location of each lesion, which is critical for CNN training and ROC/LROC studies. Sixty-five LA 99mTc-MDP spinal scans (Siemens Symbia SPECT) performed in adolescent athletes (mean age, 14.9 yrs) referred to nuclear medicine clinic with a diagnosis of LBP were retrospectively identified. An artificial lesion was modeled as a point source blurred to the system resolution determined with a 99mTc-filled capillary tube imaged in air at multiple distances from the collimator on the same camera. To generate LP 99mTc-MDP data for CNN training, lesions were incorporated at random locations in the lumbar spine of each patient with varying contrast values. This was accomplished by adding sphere sinograms to patient sinograms after application of a scaling factor (to obtain desired contrast) and Poisson deviates, followed by FBP reconstruction. More than 1.6M spinal 2.5D patches (Roth et al 2016) were extracted in LP and LA 99mTc-MDP images to train a deep CNN classifier. Evaluation of a 99mTc-MDP image during testing was performed by applying the trained CNN in a sliding-window fashion, yielding a ‘heatmap’ representing the probability of a lesion being present in each pixel. The CNN approach was evaluated by 3-fold cross-validation on a test set comprised of 50 LA and 50 LP scans (no overlap with training), which was also studied by physicians in a ROC/LROC study. The task was lumbar stress detection in pediatric 99mTc-MDP SPECT. To create LP data for the test set, artificial lesions were incorporated in LA scans at spinal sites (L3-5 on the right or left) selected by a physician as prone to exhibit stress injury in this population. Five physicians reviewed all test images, providing a confidence rating of a lesion being either present or absent and the coordinates of the most likely lesion location. The trained CNN was used to compute heatmaps for all test scans, which were in turn used to obtain ROC/LROC data. Results: All ROC/LROC curves were fitted using the standard binormal model. The mean area under the ROC curve was 0.893 for physicians and 0.905 for the CNN. For LROC, the mean area under the curve was 0.811 for the CNN and 0.787 for physicians. Conclusions: The developed deep learning approach offers similar performance to that of physicians in detecting and localizing 99mTc-MDP positive lumbar lesions. The use of this technology as a clinical decision support tool appears to have significant potential. ER -