Abstract: Objective　To evaluate the performance of the artificial intelligence (AI)⁃enabled snail identification system for recognition of Oncomelania hupensis robertsoni and Tricula in schistosomiasis⁃endemic areas of Yunnan Province. Methods　Fifty O. hupensis robertsoni and 50 Tricula samples were collected from Yongbei Township, Yongsheng County, Lijiang City, a schistosomiasis⁃endemic area in Yunnan Province in May 2024. A total of 100 snail sample images were captured with smartphones, including front⁃view images of 25 O. hupensis robertsoni and 25 Tricula samples (upward shell opening) and back⁃view images of 25 O. hupensis robertsoni and 25 Tricula samples (downward shell opening). Snail samples were identified as O. hupensis robertsoni or Tricula by schistosomiasis control experts with a deputy senior professional title and above according to image quality and morphological characteristics. A standard dataset for snail image classification was created, and served as a gold standard for recognition of snail samples. A total of 100 snail sample images were recognized with the AI⁃enabled intelligent snail identification system based on a WeChat mini program in smartphones. Schistosomiasis control professionals were randomly sampled from stations of schistosomisis prevention and control and centers for disease control and prevention in 18 schistosomiasis⁃endemic counties (districts, cities) of Yunnan Province, for artificial identification of 100 snail sample images. All professionals are assigned to two groups according the median years of snail survey experiences, and the effect of years of snail survey experiences on O. hupensis robertsoni sample image recognition was evaluated. A receiver operating characteristic (ROC) curve was plotted, and the sensitivity, specificity, accuracy, Youden's index and the area under the curve (AUC) of the AI⁃enabled intelligent snail identification system and artificial identification were calculated for recognition of snail sample images. The snail sample image recognition results of AI⁃enabled intelligent snail identification system and artificial identification were compared with the gold standard, and the internal consistency of artificial identification results was evaluated with the Cronbach's coefficient alpha. Results　A total of 54 schistosomiasis control professionals were sampled for artificial identification of snail sample image recognition, with a response rate of 100% (54/54), and the accuracy, sensitivity, specificity, Youden's index, and AUC of artificial identification were 90%, 86%, 94%, 0.80 and 0.90 for recognition of snail sample images, respectively. The overall Cronbach's coefficient alpha of artificial identification was 0.768 for recognition of snail sample images, and the Cronbach's coefficient alpha was 0.916 for recognition of O. hupensis robertsoni snail sample images and 0.925 for recognition of Tricula snail sample images. The overall accuracy of artificial identification was 90% for recognition of snail sample images, and there was no significant difference in the accuracy of artificial identification for recognition of O. hupensis robertsoni (86%) and Tricula snail sample images (94%) ([χ2] = 1.778, P > 0.05). There was no significant difference in the accuracy of artificial identification for recognition of snail sample images with upward (88%) and downward shell openings (92%) ([χ2] = 0.444, P > 0.05), and there was a significant difference in the accuracy of artificial identification for recognition of snail sample images between schistosomiasis control professionals with snail survey experiences of 6 years and less (75%) and more than 6 years (90%) ([χ2] = 7.792, P < 0.05). The accuracy, sensitivity, specificity and AUC of the AI⁃enabled intelligent snail identification system were 88%, 100%, 76% and 0.88 for recognition of O. hupensis robertsoni snail sample images, and there was no significant difference in the accuracy of recognition of O. hupensis robertsoni snail sample images between the AI⁃enabled intelligent snail identification system and artificial identification ([χ2] = 0.204, P > 0.05). In addition, there was no significant difference in the accuracy of artificial identification for recognition of snail sample images with upward (90%) and downward shell openings (86%) ([χ2] = 0.379, P > 0.05), and there was a significant difference in the accuracy of artificial identification for recognition of snail sample images between schistosomiasis control professionals with snail survey experiences of 6 years and less and more than 6 years ([χ2] = 5.604, P < 0.025). Conclusions　The accuracy of recognition of snail sample images is comparable between the AI⁃enabled intelligent snail identification system and artificial identification by schistosomiasis control professionals, and the AI⁃enabled intelligent snail identification system is feasible for recognition of O. hupensis robertsoni and Tricula in Yunnan Province.

Key words: Oncomelania hupensis robertsoni, Tricula, Artificial intelligence, Intelligent recognition, Yunnan Province

摘要： 目的　评价人工智能（artificial intelligence，AI）识螺系统识别云南省血吸虫病流行区湖北钉螺滇川亚种与拟钉螺的效能。方法　2024年5月，于云南省血吸虫病流行区选取丽江市永胜县永北镇采集湖北钉螺滇川亚种和拟钉螺样本各50只，通过智能手机拍摄100张螺样图像，包括25只湖北钉螺滇川亚种和拟钉螺壳口向上的正面图像及25只湖北钉螺滇川亚种和拟钉螺壳口向下的背面图像。由具有副高级及以上职称的血吸虫病防治专家根据图像质量和形态特征将螺样鉴别为“湖北钉螺滇川亚种”或“拟钉螺”，建立螺类图像分类标准数据集，并将该鉴别结果作为“金标准”。采用基于智能手机微信小程序的AI识螺系统对100张螺样图像进行识别。于云南省18个血吸虫病流行县（市、区）血吸虫病防治站、疾病预防控制中心随机选取血吸虫病防治专业技术人员对100张螺样图像进行人工识别测评，并按照查螺年限中位数将其分为两组，分析查螺年限对钉螺识别的影响。将AI识螺系统及人工识别结果分别与“金标准”进行比较，绘制受试者工作特征（receiver operating characteristic，ROC）曲线，计算灵敏度、特异度、准确率、约登指数及曲线下面积（area under the curve，AUC）以评价识别结果与“金标准”的一致性。采用克朗巴哈α系数评价人工识别结果的内部一致性。结果　累计选取54名血吸虫病防治专业技术人员进行人工识别测评，应答率为100%（54/54）；人工识别准确率、灵敏度、特异度、约登指数、AUC分别为90%、86%、94%、0.80和0.90。人工测评结果内部总体一致性克朗巴哈α系数为 0.768，其中湖北钉螺滇川亚种和拟钉螺图像内部一致性克朗巴哈α系数分别为0.916 和 0.925。螺样图像人工识别准确率为90%，其中湖北钉螺滇川亚种（86%）与拟钉螺识别准确率（94%）差异无统计学意义（[χ2] = 1.778，P > 0.05）。壳口向上（88%）与壳口向下螺样图像人工识别准确率（92%）差异无统计学意义（[χ2] = 0.444，P > 0.05）；查螺年限≤ 6年组和> 6年组人工识别螺样图像准确率分别为75%和90%，差异有统计学意义（[χ2] = 7.792，P < 0.05）。AI识螺系统识别湖北钉螺滇川亚种准确率为88%，与人工识别差异无统计学意义（[χ2] = 0.204，P > 0.05），灵敏度为100%、特异度为76%、AUC为0.88；对壳口向上和壳口向下螺样图像识别准确率分别为90%和86%，差异无统计学意义（[χ2] = 0.379，P > 0.05）。AI识螺系统识别准确率高于查螺年限≤ 6年组人员（[χ2] = 5.604，P < 0.025）。结论　AI识螺系统对螺样图像识别准确性与专业技术人员人工识别水平接近，可应用于云南省湖北钉螺滇川亚种与拟钉螺识别。

关键词: 湖北钉螺滇川亚种, 拟钉螺, 人工智能, 智能识别, 云南省