目的 探讨基于深度学习技术建立日本血吸虫中间宿主湖北钉螺视觉智能识别模型并评价其分类识别效能。 方法 2019年3月—2020年10月,根据江苏省湖沼型、山丘型和水网型地区钉螺分布规律,选择南京、镇江、扬州、苏州、常州、无锡和盐城等7个地区为样品采集现场,每个地区随机选择3个钉螺孳生环境采集螺样及螺样图像,由6名血吸虫病防治专家根据图像质量与螺类特征进行筛选、分类后建立螺类图像分类标准数据集,并将该结果作为金标准,建立的数据集按照7 ∶ 3原则划分为训练集和测试集（分内部和外部测试集）。采用MobilenetV2、ResNet50、Inception-ResNet-V2等3种主流卷积神经网络模型在训练集完成模型训练;以分类标准数据集作为金标准,绘制受试者工作特征（ROC）曲线,比较3种模型在内部测试集上的灵敏度、特异性、准确率、诊断一致率（Kappa值）和约登指数;选出最优识别模型再与20名查螺工作人员对外部测试集判断结果进行对比, 评价模型在内部测试集和外部测试集中识别钉螺结果的准确性。结果 共采集钉螺和方格短沟蜷、细钻螺、真管螺、拟钉螺等4种与钉螺相似的螺类图像3 224幅,经过筛选后,螺类图像分类标准数据集共纳入2 719幅图像,涵盖钉螺图像774幅和4种相似螺类的非钉螺图像1 945幅。其中,训练集占70.2%（1 910/2 719）,测试集占29.8%（749/2 719）。内部测试中,MobilenetV2、ResNet50和Inception-ResNet-V2模型识别钉螺与金标准一致性Kappa值分别为0.78、0.83、0.88;Inception-ResNet-V2模型的灵敏度、特异性、准确率、约登指数和ROC曲线下面积（AUC）均最高,分别为92.00%、97.16%、96.13%、0.89、0.95;3种模型识别的灵敏度和特异性差异无统计学意义（χ² = 3.892、4.948,P > 0.05）,准确率差异有统计学意义（χ² = 8.607,P < 0.05）。外部测试中,最优模型Inception-ResNet-V2模型识别和工作人员鉴别与金标准一致性较好,Kappa值分别为0.80和0.83,AUC分别为0.88和0.92,差异无统计学意义（P > 0.05）。 结论 基于深度学习技术的钉螺视觉智能识别模型具有良好的准确性。

Objective To investigate the performance of a deep learning-based visual intelligent recognition model for the intermediate host of Schistosoma japonicum, the snail Oncomelania hupensis, and evaluate its efficacy in recognition and classification. Methods According to the distribution pattern of O. hupensis in the topographic type of lake marsh, hill and water network in Jiangsu Province, seven regions including Nanjing, Zhenjiang, Yangzhou, Suzhou, Changzhou, Wuxi and Yancheng were selected as the fields for sample collection from March 2019 to October 2020. From each of the fields, 3 snail breeding environment sites were randomly selected for collection of snail samples and images by smart phone. The image quality and morphologic features of the sampled snails were screened and classified by six experts of schistosomiasis control to establish a standard data set of snail image classification. The data set was compiled into training set and test set (comprised of internal and external test). Three major convolutional neural network models including MobilenetV2, ResNet50 and Inception-ResNet-V2 were used to exercise the model training in the training data set. Taking the standard snail image classification data set as the gold standard, the receiver operating characteristic (ROC) curves was applied to compare the sensitivity, specificity, diagnostic concordance rate (Kappa value), and Yorden index of three model types with the internal test data; comparison was performed between the best-fit recognition model and the findings of snail search staff through the external test set, evaluating the accuracy of models in snail recognition with the internal and external test set. Results Totally, 3 224 images of 4 types of snails similar to O. hupensis were collected: Semisulcospira cancellata, Opeas gracile, Euphaedus, and Tricula. After screening, 2 719 images were included into the standard snail image classification data set, among them 774 were of O. hupensis and 1 945 of 4 being similar to but not O. hupensis. The concordance Kappa value from the snail recognition models of MobilenetV2, ResNet50 and Inception-ResNet-V2 model with the gold standard was 0.78, 0.83 and 0.88 respectively. The sensitivity, specificity, accuracy, Yorden index, and area under the ROC curve (AUC) of the Inception-ResNet-V2 were found highest, being 92.00%, 97.16%, 96.13%, 0.89 and 0.95, respectively. There was no statistically significant difference in the sensitivity and specificity among the three models (χ² = 3.892, 4.948, P > 0.05), while the difference in accuracy was statistically significant (χ² = 8.607, P < 0.05). In the external test, the best-fit model Inception-ResNet-V2 and the findings by staff showed better concordance with the gold standard, having the Kappa values of 0.80 and 0.83, and the AUCs 0.88 and 0.92, respectively, of which the difference was not statistically significant (P > 0.05). Conclusion The deep learning-based visual intelligent recognition model for O. hupensis snailshowed evident accuracy.