Objective: To explore the application of manual screening collaborated with the Artificial Intelligence TPS-Assisted Cytologic Screening System in urinary exfoliative cytology and its clinical values. Methods: A total of 3 033 urine exfoliated cytology samples were collected at the Henan People's Hospital, Capital Medical University, Beijing, China. Liquid-based thin-layer cytology was prepared. The slides were manually read under the microscope and digitally presented using a scanner. The intelligent identification and analysis were carried out using an artificial intelligence TPS assisted screening system. The Paris Report Classification System of Urinary Exfoliated Cytology 2022 was used as the evaluation standard. Atypical urothelial cells and even higher grade lesions were considered as positive when evaluating the recognition sensitivity, specificity, and diagnostic accuracy of artificial intelligence-assisted screening systems and human-machine collaborative cytologic screening methods in urine exfoliative cytology. Among the collected cases, there were also 1 100 pathological tissue controls. Results: The accuracy, sensitivity and specificity of the AI-assisted cytologic screening system were 77.18%, 90.79% and 69.49%; those of human-machine coordination method were 92.89%, 99.63% and 89.09%, respectively. Compared with the histopathological results, the accuracy, sensitivity and specificity of manual reading were 79.82%, 74.20% and 95.80%, respectively, while those of AI-assisted cytologic screening system were 93.45%, 93.73% and 92.66%, respectively. The accuracy, sensitivity and specificity of human-machine coordination method were 95.36%, 95.21% and 95.80%, respectively. Both cytological and histological controls showed that human-machine coordination review method had higher diagnostic accuracy and sensitivity, and lower false negative rates. Conclusions: The artificial intelligence TPS assisted cytologic screening system has achieved acceptable accuracy in urine exfoliation cytologic screening. The combination of manual screening and artificial intelligence TPS assisted screening system can effectively improve the sensitivity and accuracy of cytologic screening and reduce the risk of misdiagnosis.
目的： 探讨人工阅片协同人工智能TPS辅助阅片系统在尿液脱落细胞学中的应用价值。 方法： 收集河南省人民医院2015—2022年3 033例尿液脱落细胞学样本，液基薄层细胞学制片，显微镜下人工阅片，并将玻片利用扫描仪进行数字化呈现，采用人工智能TPS辅助阅片系统进行智能识别及分析，以2022版尿脱落细胞学巴黎报告分类系统为评判标准，将非典型尿路上皮细胞及以上级别作为阳性标准，评估人工智能辅助阅片系统以及人机协同阅片方式在尿脱落细胞学上的识别准确率、灵敏度及特异度；其中有组织病理对照者1 100例，以其为金标准，分析不同阅片方式识别准确率、灵敏度、特异度、假阴性率及假阳性率。 结果： 人工智能辅助阅片系统的准确率为77.18%，灵敏度90.79%，特异度为69.49%；人机协同的方式准确率为92.89%，灵敏度99.63%，特异度89.09%。对照组织病理学结果，发现人工阅片准确率为79.82%，灵敏度74.20%，特异度95.80%；人工智能辅助阅片系统准确率为93.45%，灵敏度93.73%，特异度92.66%；人机协同的方式准确率为95.36%，灵敏度95.21%，特异度95.80%，无论从细胞学还是组织学方面对照，人机协同阅片复核方式均具有更高的诊断准确率及灵敏度，假阴性率下降。 结论： 人工智能TPS辅助阅片系统实现了尿液脱落细胞学智能化阅片，人工阅片协同人工智能TPS辅助阅片系统可有效提高阅片的灵敏度及准确率，降低漏诊风险。.