目的 探讨基于卷积神经网络的ResNet50模型辅助肝细粒棘球蚴病（HCE）超声分型诊断的应用价值。方法 回顾性收集2020年1月至2024年6月新疆医科大学第一附属医院等5家医院（分别为A~E组，其中A组为省级三级甲等医院，B~D组为地州级三级甲等医院，E组为县级二级甲等医院；另将4家基层医院合并为F组，5家医院合并为G组）经术后病理或多种影像学长期随访确诊为HCE患者的肝脏内病灶超声图像，分别由1名低年资超声医师和1名高年资超声医师按照CE1（单囊型）、CE2（多子囊型）、CE3（内囊塌陷型）、CE4（坏死实变型）和CE5（钙化型）等5种类型进行分型诊断，再上传至基于ResNet50模型的HCE智能分型诊断超声影像云平台进行分型诊断。采用卡方检验比较各组间低、高年资超声医师和ResNet50模型HCE分型诊断准确率，3组比较有统计学意义时，进一步采用Bonferroni校正进行两两比较；采用kappa分析评价诊断一致性。结果 共纳入952例HCE患者的1 596张HCE超声图像，各组低、高年资超声医师和ResNet50模型诊断准确率分别为：A组88.3%（594/673）、95.7%（644/673）、89.2%（600/673），B组72.7%（157/216）、86.6%（187/216）、89.8%（194/216），C组77.6%（232/299）、87.6%（262/299）、91.3%（273/299），D组74.5%（164/220）、85.5%（188/220）、90.5%（199/220），E组69.1%（130/188）、84.0%（158/188）、88.3%（166/188），F组75.1%（693/923）、86.1%（795/923）、90.1%（832/923），G组80.6%（1 287/1 596）、90.2%（1 439/1 596）、89.7%（1 432/1 596）。A组ResNet50模型诊断总体准确率低于高年资超声医师（χ² = 27.140，P < 0.05）；B、C、D与E组的ResNet50模型总体诊断准确率均高于低年资超声医师（χ² = 25.381、24.307、21.121、24.213，均P < 0.05）；F组ResNet50模型总体诊断准确率高于低、高年资超声医师（χ² = 82.654，P < 0.05）；G组ResNet50模型总体诊断准确率高于低年资超声医师（χ² = 80.749，P < 0.05）；以上经Bonferroni校正后两两比较均P < 0.025。各组中ResNet50模型与高年资超声医师诊断一致性最高（kappa值均 ≥ 0.836），低年资与高年资超声医师间诊断一致性稍低（kappa值为0.754~0.863），ResNet50模型与低年资超声医师间诊断一致性最低（kappa值为0.674~0.855）。ResNet50模型中CE1最易被误诊为CE3，占CE1误诊的68.6%（24/35）；CE2易被误诊为 CE4、CE3，分别占CE2 误诊的37.9%（11/29）、31.0%（9/29）；CE3最易被误诊为CE2，占CE3误诊的65.6%（21/32）；CE4与CE5相互容易被误诊（61.9%，26/42；57.9%，22/38）。结论 基于ResNet50架构的卷积神经网络模型可有效解析HCE分型的超声特征，为基层医疗机构医师提供辅助诊断支持，提升分型判读的准确性。

Objective To examine the value of the convolution neural network-based ResNet50 model in assisting ultrasound classification of hepatic cystic echinococcosis (HCE). Methods Ultrasound images of intrahepatic lesions were retrospectively collected from HCE patients that were definitely diagnosed with postoperative pathology or long-term follow-up of multiple imaging tools admitted to five medical institutions including First Affiliated Hospital of Xinjiang Medical University (groups A to E, among them, group A is a provincial-level tertiary grade A hospital, groups B to D are prefecture level tertiary grade A hospitals, and group E is a county-level secondary Grade A hospital; and four grassroots hospitals were merged into group F and 5 hospitals merged into group G) during the period from January 2020 through June 2024. All images were classified by a junior sonographer and a senior sonographer according to five types of CE1 (univesicular cysts), CE2 (multiple daughter cysts), CE3 (cyst with detachment of laminated membrane), CE4 (heterogeneous or hyperechoic degenerative contents), and CE5 (calcified cysts), and uploaded to the ResNet50 model-based intelligent classification diagnosis ultrasound imaging cloud platform for classification diagnosis of HCE. The diagnostic accuracy of HCE classification by junior sonographers, senior sonographers and ResNet50 model was compared among groups using chi-square test, and pairwise comparisons were conducted with Bonferroni correction. In addition, the diagnostic consistency was evaluated with kappa analysis. Results A total of 1 596 ultrasound images of HCE were included from 952 HCE patients. The diagnostic accuracy of HCE classification by junior sonographers, senior sonographers and ResNet50 model were 88.3% (594/673), 95.7% (644/673) and 89.2% (600/673) in Group A, 72.7% (157/216), 86.6% (187/216), and 89.8% (194/216) in Group B, 77.6% (232/299), 87.6% (262/299) and 91.3% (273/299) in Group C, 74.5% (164/220), 85.5% (188/220) and 90.5% (199/220) in Group D, 69.1% (130/188), 84.0% (158/188) and 88.3% (166/188) in Group E, 75.1% (693/923), 86.1% (795/923) and 90.1% (832/923) in Group F and 80.6% (1 287/1 596), 90.2% (1 439/1 596) and 89.7% (1 432/1 596) in Group G. The overall diagnostic accuracy of HCE classification was lower by the ResNet50 model diagnosis than by senior sonographers in Group A (χ² = 27.140, P < 0.05), and the overall diagnostic accuracy was higher by the ResNet50 model than by junior sonographers in groups B, C, D, and E (χ² = 25.381, 24.307, 21.121 and 24.213; all P < 0.05). The overall diagnostic accuracy of HCE classification was higher by the ResNet50 model than by junior and senior sonographers in Group F (χ² = 82.654, P < 0.05), and the overall diagnostic accuracy was higher by the ResNet50 model than by junior sonographers in Group G (χ² = 80.749, P < 0.05). Moreover, all pairwise comparisons after Bonferroni correction were statistically significant with P < 0.025. The ResNet50 model demonstrated the highest diagnostic consistency with senior sonographers for HCE classification across all groups (kappa ≥ 0.836), followed by between junior and senior sonographers (0.754 ≤ kappa ≤ 0.863), and the lowest diagnostic consistency was seen between the ResNet50 model and junior sonographers (0.674 ≤ kappa ≤ 0.855). In the ResNet50 model, CE1 was most likely to be misdiagnosed as CE3 (68.6%, 24/35), CE2 misdiagnosed as CE4 (37.9%, 11/29) or CE3 (31.0%, 9/29), CE3 misdiagnosed as CE2 (65.6%, 21/32), and CE4 and CE5 were likely to be mutually misdiagnosed (61.9%, 26/42; 57.9%, 22/38). Conclusion The convolutional neural network-based ResNet50 model is effective to decipher the ultrasound features of HCE classification, which provides auxiliary diagnostic supports to doctors in primary medical institutions, and improves the accuracy of classification interpretation.