Miranda Vinícius Lima de, Gurgel-Gonçalves Rodrigo
Laboratório de Parasitologia Médica e Biologia de Vetores, Faculdade de Medicina, Universidade de Brasília, DF, Brasil.
Laboratório de Parasitologia Médica e Biologia de Vetores, Faculdade de Medicina, Universidade de Brasília, DF, Brasil.
Acta Trop. 2025 May;265:107621. doi: 10.1016/j.actatropica.2025.107621. Epub 2025 Apr 26.
Correct identification of blood-sucking bugs, such as triatomines, is important because they are vectors of Chagas' disease. Identifying these insects is often difficult for non-specialists. Deep learning is emerging as a solution for automated identification. This study evaluates the performance of three convolutional neural networks (CNNs) - AlexNet, MobileNetV2 and ResNet-50 - to identify bugs categorized by their feeding habits: 'blood-suckers', 'phytophagous' and 'predators'. A dataset of 707 dorsal view pictures was divided into training, validation, and test subsets (70 %, 10 %, and 20 %, respectively). Transfer learning was used to train the models, and Grad-CAM visualizations identified the picture regions that most influenced the predictions. All models achieved an accuracy of over 94 %, with ResNet-50 slightly outperforming the other models in terms of sensitivity and specificity. ROC and AUC analyses confirmed the reliability of these algorithms, highlighting their potential for robust bug identification. This study demonstrates the applicability of CNNs in distinguishing Triatominae from other insects, paving the way for the development of affordable vector identification tools to improve Chagas disease surveillance and control.
正确识别吸血虫，如锥蝽，非常重要，因为它们是恰加斯病的传播媒介。对于非专业人员来说，识别这些昆虫往往很困难。深度学习正在成为一种自动识别的解决方案。本研究评估了三种卷积神经网络（CNN）——AlexNet、MobileNetV2和ResNet-50——对按进食习惯分类的昆虫进行识别的性能：“吸血者”、“植食性”和“捕食者”。一个包含707张背视图图片的数据集被分为训练集、验证集和测试集（分别为70%、10%和20%）。使用迁移学习来训练模型，Grad-CAM可视化确定了对预测影响最大的图片区域。所有模型的准确率都超过了94%，ResNet-50在灵敏度和特异性方面略优于其他模型。ROC和AUC分析证实了这些算法的可靠性，突出了它们在可靠识别昆虫方面的潜力。本研究证明了CNN在区分锥蝽与其他昆虫方面的适用性，为开发经济实惠的病媒识别工具以改善恰加斯病监测和控制铺平了道路。
