The Evaluation Of Preoperative Imaging Planning Of Pelvic Tumor Based On Deep Learning

Objective:To explore the feasibility and effect of applying Fully Convolutional Networks(FCN)in the automatic segmentation on images of pelvic tumors,and to assess the accuracy of performing surgical planning during pelvic tumor resection through image registration method.Methods:A retrospective study was performed on the cases of pelvic or femoral head tumors admitted from March 2012 to June 2018 in our hospital.Preoperative CT and MRI imaging findings were summarized.Manual segmentations performed by Senior radiologists for diagnosis of musculoskeletal system were taken as gold standard.The densely connected fully convolutional network model with hole convolution and transfer learning were chosen to train with a 10-fold cross-validation with Dice coefficient and Hausdorff distance as evaluation indicators.Meanwhile,the rigid registration was performed between preoperative and postoperative prosthetic and pelvic bone models to e accuracy of performing surgical planning during pelvic tumor resection.The angle and distances between the preoperative planning plane and surgical osteotomy plane were choosen in this evaluation part.Results:90 patients were included in this study,43.3% of them(39 patients)were with the performance of high signal of hip synovium in T1 WI enhanced sequence,34 patients with malignant tumors;64.4% of them(58 patients)were with the performance of muscle interstitial edema around the humerus in MRI images,49 patients with malignant tumors;34.4% of them(31 patients)were with the performance of soft tissue mass in the pelvic tumors,all of were malignant tumors.The Dice Similarity Coefficient(DSC)between the AI segmentation results and gold standard was 0.85±0.08(0.621～0.963),lower than that of radiologists(0.92±0.03,0.843～0.964).The Hausdorff distance was 5.34±1.71(2.189～8.762)millimeter,lower than that of radiologists(5.80±3.07 mm,2.065～14.645).The AI can automatically identify pelvic tumor boundaries accurately.The angle between the actual osteotomy plane and the planned osteotomy plane is 6 degrees on average,and the distance between them is 5 milimeter on average.Conclusions:It is of great value to determined tumor boundaries on preoperative CT and MRI images.The application of full convolutional neural network to automatically segment of pelvic tumors’ boundaries was feasible.It is meaningful by the verification method of registration for overcoming the effects of postoperative image metal artifacts to assess the accuracy of performing surgical planning during pelvic tumor resection.