论文部分内容阅读
目的:探讨基于深度学习构建胆道闭锁(biliary atresia，BA)超声人工智能(artificial intelligence，AI)诊断模型的可行性。方法:前瞻性收集2018年9月至2020年10月湖南省儿童医院诊治的177例BA患儿(BA组)共计531张胆囊超声初始影像及195例非BA患儿(非BA组)共计585张胆囊超声初始影像，各组按2∶1分为训练集与测试集。使用训练集训练深度神经网络模型Mask R-CNN后，采用测试集分别以患儿和图像为单位对该模型进行测试，评价模型对胆囊的检测率及诊断准确率。另将测试集图像分别以患儿、图像为单位进行随机编号，分别邀请4名超声医师进行图片判读，计算诊断准确率。对模型诊断准确率与超声医师诊断准确率进行比较。结果:在胆囊器官的自动检测方面：模型在BA组与非BA组的检测率均达到100%，但在总计372张测试集图像中有17张出现虚警，虚警率4.57%(17/372)。在诊断方面：以患儿为单位时，模型在测试集中总的诊断准确率为95.97%，高于外院超声医师及本院中级职称超声医师(均n P<0.005)，略高于本院高级职称超声医师(91.94%)，但差异无统计学意义(n P＝0.183)。以图片为单位时，模型在测试集中总的诊断准确率为97.04%，均高于外院超声医师及本院中级职称超声医师(均n P<0.001)，略高于本院高级职称超声医师(94.09%)，但差异无统计学意义(n P＝0.05)。n 结论:基于Mask R-CNN的AI模型可较准确地检测胆囊器官，对BA的诊断准确率较高，该模型切实可行，值得进一步研究。“,”Objective:To explore the feasibility of artificial intelligence ultrasound to diagnose of biliary atresia (BA) based on deep learning.Methods:A total of 531 gallbladder ultrasound images in 177 cases of BA patients (BA group) and 585 gallbladder ultrasound images in 195 cases of Non-BA patients (Non-BA group) were collected in Hunan Children′s Hospital from September 2018 to October 2020. For the BA and Non-BA groups, all images were divided into training set and test set according to the ratio of 2∶1. The Mask R-CNN model was trained by training samples, and then the model was tested, according to patient and image as a unit respectively, to evaluate the gallbladder organ detection rate and the diagnostic accuracy of BA. In addition, the images of the test set were randomly numbered.Four sonographers were invited to interpret the images to calculate the diagnostic accuracy individually. Last, the diagnostic accuracy of the Mask R-CNN model was compared with that of sonographers.Results:In terms of the automatic detection of gallbladder organs, the detection rate in both BA and Non-BA group reached 100%, but there were 17 false alarms in 372 test images, with a false alarm rate of 4.57%. In terms of the diagnostic rate of gallbladders, when taking patient as a unit, the total diagnostic accuracy of the model in the test set was 95.97%, which was higher than that of the sonographers in other hospitals and the sonographer with intermediate professional title in our hospital, and the difference was statistically significant (n P<0.005). It was slightly higher than that of sonographer with senior professional title in our hospital (91.94%), but the difference was not statistically significant (n P=0.183). When taking picture as a unit, the total diagnostic accuracy of the model was 97.04%, which was higher than that of the sonographers in other hospitals and the sonographer with intermediate professional title in our hospital, and the difference was statistically significant (n P<0.001). It was slightly higher than that of sonographer with senior professional title in our hospital (94.09%), but the difference was not statistically significant (n P=0.05).n Conclusions:The artificial intelligence technology based on Mask R-CNN can automatically and accurately detect gallbladder organs and diagnose BA, which is worthy of further study.