Research On The Application Value Of Enhanced CT Based On Deep Learning Algorithm In The Differential Diagnosis Of Pancreatic Cystic Lesion

Part Ⅰ Construction of an automatic pancreatic segmentation model based on deep learning algorithm on dual-phase enhanced CTPurpose:To construct a deep learning model for automatic pancreas segmentation on dualphase enhanced CT.To evaluate the segmentation performance using subjective and objective evaluation systems,respectively,and to explore the factors affecting pancreas segmentation on dual-phase enhanced CT.Materials and Methods:A total of 218 cases with dual-phase enhanced pancreatic CT from January to November 2019 were retrospectively collected.The data were randomly divided into the training,validation,and testing set(139,20,and 59 cases,in order).A two-stage global-local progressive fusion network was used for constructing the segmentation model,the testing set was used to evaluate the model’s performance subjectively and objectively.Both the Likert 5-point scale for subjective evaluation and the Dice similarity coefficient(DSC)for objective evaluation were based on the overlapping of pancreas between the model’s segmentation and doctors’pre-annotation.Results:Subjective scores for the arterial phase were significantly higher than that of the venous phase for the critical regions of the pancreas at the duodenum,duodenal jejunal flexure(DJF),and left adrenal gland(LAG),and splenic artery(SPA).Subjective scores of venous phase were significantly higher than that of the arterial phase for the critical regions of the pancreas at the portal vein,superior mesenteric vein,and splenic vein.DSC for the venous phase was slightly higher than that of the arterial phase,without a significant difference(DSC:0.932 versus 0.921,P=0.952).When the fat gap between the pancreas and the surrounding organs occurred,there were no significant differences in subjective scores on dual-phase enhanced CT.Except for the spleen,the density differences between the critical regions of the pancreas and other surrounding organs were statistically significant on dual-phase enhanced CT.Conclusion:The automatic pancreatic segmentation model got a stable performance on dual-phase enhanced CT.Subjective evaluation was helpful to improve the segmentation ability for the critical regions of the pancreas.The arterial phase could add complementary value to the venous phase for pancreatic critical regions’ segmentation at the duodenum,DJF,LAG,and SPA.Density differences between the pancreas and surrounding organs could affect the pancreatic segmentation performance.Part Ⅱ Construction of an automatic detection model for pancreatic cystic lesions based on deep learning algorithm on dual-phase enhanced CTPurpose:To construct a deep learning model for automatic detection(DLM-AD)of pancreatic cystic lesions(PCLs)on dual-phase enhanced CT,and compare the model’s detection performance to the senior and junior radiologists.Materials and Methods:A total of 758 cases with dual-phase enhanced pancreatic CT from January 2018 to February 2020 were retrospectively collected,including normal pancreas and PCLs.The data were randomly divided into a training set(557 cases),a validation set(36 cases),and a testing set(165 cases).A two-stage global-local progressive fusion network was used for the construction of DLM-AD,and the testing set was used to evaluate the PCL detection effect.The lesions were divided into different subgroups according to the location,size,presence or absence of mural nodules or solid components and pancreatic duct dilatation,and whether they were communicated to the main pancreatic duct.Receiver operating characteristic curve and area under the curve(AUC)were used to calculate the dectection performance of the model.To compare the accuracy,sensitivity,and specificity of the model with that of the senior and junior radiologists for overall lesion detection.Compare the results between DLM-AD and radiologists for detection of each subgroup of lesions.Results:The AUC of DLM-AD for PCLs’ overall detection was 0.959,and the accuracy,sensitivity,and specificity were 0.933,0.910,and 0.961,respectively.The accuracy and specificity of the DLM-AD in detecting lesions were without significant difference to that of the senior radiologist.However,the sensitivity of the DLM-AD was significantly lower than that of the senior radiologist(P=0.016).The accuracy,sensitivity and specificity of the DLM-AD were all significantly higher than that of the junior radiologist(P=0.043,0.031,<0.001).The correct detection rate of the model was significantly lower than that of the senior radiologist in the types of lesions with a maximum diameter<3 cm,without pancreatic duct dilatation and without communication with the pancreatic duct(P=0.016,0.016,0.016).For the two types of PCLs,including a maximum diameter of<3cm and without pancreatic duct dilatation,the detection rate of the junior radiologist was significantly increased with the aid of the model(P=0.016,0.031).The diagnostic time of the junior radiologist was significantly shortened(P<0.001).Conclusion:The DLM-AD achieved a stable and senior radiologist-level performance in detecting PCLs.With the assistance of the model,the detection rate of PCLs for the junior radiologist significantly improved,especially for lesions<3cm.In addition,the model could significantly improve the detection speed of the junior radiologist.Part Ⅲ Construction of a differentiation model for benign and malignant pancreatic cystic lesions based on deep learning algorithm on dual-phase enhanced CTPurpose:To construct a deep learning model for differentiating benign and malignant(DLM-DBM)pancreatic cystic lesions(PCLs)on dual-phase enhanced CT,and compare the diagnostic performance to traditional radiomics model and doctors with different seniority.Materials and Methods:Dual-phase enhanced CT images of 363 patients with 368 PCLs obtained from two hospitals were retrospectively assessed.Based on the examination date,our hospital’s data were designated to a training and validation set of 266 PCLs and an internal testing set of 52 PCLs.The external hospital’s data was used as an independent external testing set of 50 PCLs.The DLM-DBM was developed using 3D specially designed densely connected convolutional networks for PCL differentiation.Radiomic features were extracted to build a traditional radiomics model for differentiating benign and malignant(TRM-DBM)PCLs.Clinical and radiological characteristics were compared between the benign and malignant PCLs.Receiver operating characteristic curve and area under the curve(AUC)were used to calculate the diagnostic performance of DLM-DBM and TRM-DBM.Compare the diagnostic accuracy of DLM-DBM,TRMDBM,and a senior radiologist,a junior radiologist,and a surgeon.Results:The AUC of the DLM-DBM is higher than that of the TRM-DBM for both the internal and external test sets(internal testing set:0.933 versus 0.879;external testing set:0.911 versus 0.768).The accuracy for differential diagnosis was 0.904 with DLM-DBM,which was the highest in the internal testing set.Accuracy differences between the DLMDBM and senior radiologist were not significant both in the internal and external testing set.With the assistance of the DLM-DBM,the accuracy and specificity of a junior radiologist were significantly improved(all P<0.05),and all doctors’ diagnostic time was significantly shortened(all P<0.05).Conclusion:The DLM-DBM achieved senior radiologist-level performance in differentiating benign and malignant PCLs which could improve the junior radiologist’s interpretation and speed of doctors with different seniority.