Detecting lumbar lesions in 99mTc‐MDP SPECT by deep learning: Comparison with physicians. Issue 8 (11th July 2021)

• Record Type: Journal Article
• Authors: Petibon, Yoann; Fahey, Frederic; Cao, Xinhua; Levin, Zakhar; Sexton‐Stallone, Briana; Falone, Anthony; Zukotynski, Katherine; Kwatra, Neha; Lim, Ruth; Bar‐Sever, Zvi; Chemli, Yanis; Treves, S. Ted; Fakhri, Georges El; Ouyang, Jinsong
• Is Part Of: Medical physics. Volume 48:Issue 8(2021)
• Page Start: 4249
• Page End: 4261
• Publication Date: 2021-07-11
• Subjects: bone SPECT -- CNN -- deep learning -- localization ROC; Medical physics -- Periodicals; Medical physics; Geneeskunde; Natuurkunde; Toepassingen; Biophysics; Periodicals; Periodicals; Electronic journals; 610.153
• DOI: 10.1002/mp.15033
• Languages: English
• ISSNs: 0094-2405

Abstract

Abstract: Purpose: 99m Tc‐MDP single‐photon emission computed tomography (SPECT) is an established tool for diagnosing lumbar stress, a common cause of low back pain (LBP) in pediatric patients. However, detection of small stress lesions is complicated by the low quality of SPECT, leading to significant interreader variability. The study objectives were to develop an approach based on a deep convolutional neural network (CNN) for detecting lumbar lesions in 99m Tc‐MDP scans and to compare its performance to that of physicians in a localization receiver operating characteristic (LROC) study. Methods: Sixty‐five lesion‐absent (LA) 99m Tc‐MDP studies performed in pediatric patients for evaluating LBP were retrospectively identified. Projections for an artificial focal lesion were acquired separately by imaging a 99m Tc capillary tube at multiple distances from the collimator. An approach was developed to automatically insert lesions into LA scans to obtain realistic lesion‐present (LP) 99m Tc‐MDP images while ensuring knowledge of the ground truth. A deep CNN was trained using 2.5D views extracted in LP and LA 99m Tc‐MDP image sets. During testing, the CNN was applied in a sliding‐window fashion to compute a 3D "heatmap" reporting the probability of a lesion being present at each lumbar location. The algorithm was evaluated using cross‐validation on a 99m Tc‐MDP test dataset which was also studied by five physicians in a LROC study. LP images in the test set were obtained by incorporating lesions at sites selected by a physician based on clinical likelihood of injury in this population. Results: The deep learning (DL) system slightly outperformed human observers, achieving an area under the LROC curve (AUCLROC ) of 0.830 (95% confidence interval [CI]: [0.758, 0.924]) compared with 0.785 (95% CI: [0.738, 0.830]) for physicians. The AUCLROC for the DL system was higher than that of two readers (difference in AUCLROC [ΔAUCLROC ] = 0.049 and 0.053) who participated to the study and slightly lower than that of two other readers (ΔAUCLROC = −0.006 and −0.012). Another reader outperformed DL by a more substantial margin (ΔAUCLROC = −0.053). Conclusion: The DL system provides comparable or superior performance than physicians in localizing small 99m Tc‐MDP positive lumbar lesions.