Objective: To establish a deep learning-based visual model for intelligent recognition of Oncomelania hupensis, the intermediate host of Schistosoma japonicum, and evaluate the effects of different training strategies for O. hupensis image recognition.
Methods: A total of 2 614 datasets of O. hupensis snails and 4 similar snails were generated through field sampling and internet capture, and were divided into training sets and test sets. An intelligent recognition model was created based on deep learning, and was trained and tested. The precision, sensitivity, specificity, accuracy, F1 score and Youden index were calculated. In addition, the receiver operating characteristic (ROC) curve of the model for snail recognition was plotted to evaluate the effects of "new learning", "transfer learning" and "transfer learning + data enhancement" training strategies on the accuracy of the model for snail recognition.
Results: Under the "transfer learning + data enhancement" strategy, the precision, sensitivity, specificity, accuracy, Youden index and F1 score of the model were 90.10%, 91.00%, 97.50%, 96.20%, 88.50% and 90.51% for snail recognition, which were all higher than those under both "new learning" and "transfer learning" strategies. There were significant differences in the sensitivity, specificity and accuracy of the model for snail recognition under "new learning", "transfer learning" and "transfer learning + data enhancement" training strategies (all P values < 0.001). In addition, the area under the ROC curve of the model was highest (0.94) under the "transfer learning + dataenhancement" training strategy.
Conclusions: This is the first visual model for intelligent recognition of O. hupensis based on deep learning, which shows a high accuracy for snail image recognition. The "transfer learning + data enhancement" training strategy is helpful to improve the accuracy of the model for snail recognition.
［摘要］ 目的 建立一种基于深度学习技术的日本血吸虫中间宿主湖北钉螺视觉智能识别模型, 评价不同训练策略用于钉螺图像识别的效果。方法 通过现场采集和互联网抓取构建钉螺及 4 种相似螺类数据集 2 614 幅, 将其分为训练集和测试集。基于深度学习技术建立智能识别模型, 并对模型进行训练及测试, 计算模型识别钉螺的精确率、敏感性、特异性、准确率、F1 值、约登指数; 绘制受试者工作特征 (receiver operating characteristic, ROC) 曲线, 分析“全新学习”、“迁移学习”、“迁移学习+数据增强”等 3 种不同训练策略对模型识别钉螺准确性的影响。结果 “迁移学习+数据增强”训练策略下, 模型识别钉螺的精确率、敏感性、特异性、准确率、约登指数和F1值分别为 90.10%、91.00%、97.50%、96.20%、88.50%、90.51%, 均高于“全新学习”、“迁移学习”策略; “全新学习”、“迁移学习”、“迁移学习+数据增强”训练策略下, 模型识别钉螺的敏感性、特异性和准确率差异均有统计学意义 (P 均 < 0.001)。“迁移学习+数据增强”训练策略下, 模型 ROC 曲线下面积最大 (0.94)。结论 首次建立了基于深度学习技术的湖北钉螺视觉智能识别模型, 钉螺图像识别准确性较高。“迁移学习+数据增强”训练策略有助于提高模型识别钉螺的准确性。.
Keywords:
Artificial intelligence; Computer vision; Deep learning; Intelligent recognition; Machine learning;