Preoperative Diagnosis Of Bile Duct Tumor Thrombus Based On Object Detection Neural Network

BDTT,abbreviation of bile duct tumor thrombus,is mainly caused by the invasion of liver tumor into biliary system.BDTT can block the bile duct and cause cholestasis,which can lead to bile duct dilatation and obstructive jaundice.BDTT has a low clinical incidence in patients with hepatocellular carcinoma(HCC)and is relatively difficult to diagnose.Several clinical studies have shown that if BDTT cannot be removed from the patient’s body timely and effectively,it will increase the risk of postoperative recurrence of these HCC patients.Therefore,the early diagnosis of BDTT is of great clinical importance for determining the reasonable treatment plan and improving the postoperative prognosis of these patients.Currently,the preoperative diagnosis of BDTT is usually based on identifying intrahepatic biliary dilation on medical images(eg.,CT and MRI images).However,compared with tumors and other lesion,dilated bile ducts(DBDs)are not conspicuous enough on medical images and can be easily overlooked when doctors report imaging scan results,leading to a high misdiagnosis rate for BDTT in practice.The present study was to develop an artificial intelligence(AI)-assisted method for preoperative diagnosis of HCC patients with BDTT.Specifically,this study proposed to apply object detection neural networks to identify DBDs on CT images,thereby realizing the indirect diagnosis and evaluation of BDTT.The proposed AI-assisted method was applied to a clinical imaging dataset collected from four hospitals and demonstrated its excellent performance.Based on individual object detection neural network,the proposed method achieved an average true positive rate of 0.92 for identifying DBDs per patient.For patient-level diagnosis of BDTT,the method achieved a true positive rate of 1.00 with a F1 score of 0.94 and an AUC value of 0.95(95%CI:0.88,1.00),which is significantly better than the result achieved by traditional machine learning method random forest(F1 score:0.71,AUC value:0.71(95%CI:0.51,0.90)).Furthermore,this study adopted the model ensemble strategy to aggregate the detection results from multiple individual object detection neural networks.Model ensemble was shown to effectively reduce the number of false positive detection of DBDs and further improved the patient-level diagnosis result(F1 score:0.97,AUC:0.97(95%CI:0.89,1.00)).