Center for Global Health, School of Public Health, Nanjing Medical University, Nanjing, 211166, China.
J Microbiol. 2021 Jun;59(6):563-572. doi: 10.1007/s12275-021-1013-z. Epub 2021 Mar 29.
Fungi of the genus Aspergillus are ubiquitously distributed in nature, and some cause invasive aspergillosis (IA) infections in immunosuppressed individuals and contamination in agricultural products. Because microscopic observation and molecular detection of Aspergillus species represent the most operator-dependent and time-intensive activities, automated and cost-effective approaches are needed. To address this challenge, a deep convolutional neural network (CNN) was used to investigate the ability to classify various Aspergillus species. Using a dissecting microscopy (DM)/stereomicroscopy platform, colonies on plates were scanned with a 35× objective, generating images of sufficient resolution for classification. A total of 8,995 original colony images from seven Aspergillus species cultured in enrichment medium were gathered and autocut to generate 17,142 image crops as training and test datasets containing the typical representative morphology of conidiophores or colonies of each strain. Encouragingly, the Xception model exhibited a classification accuracy of 99.8% on the training image set. After training, our CNN model achieved a classification accuracy of 99.7% on the test image set. Based on the Xception performance during training and testing, this classification algorithm was further applied to recognize and validate a new set of raw images of these strains, showing a detection accuracy of 98.2%. Thus, our study demonstrated a novel concept for an artificial-intelligence-based and cost-effective detection methodology for Aspergillus organisms, which also has the potential to improve the public's understanding of the fungal kingdom.
曲霉菌属真菌广泛分布于自然界，其中一些会导致免疫功能低下人群发生侵袭性曲霉病（IA）感染和农产品污染。由于曲霉菌种的显微镜观察和分子检测是最依赖操作者且耗时最长的工作，因此需要自动化且具有成本效益的方法。为了应对这一挑战，我们使用深度卷积神经网络（CNN）来研究其对各种曲霉菌种进行分类的能力。利用解剖显微镜（DM）/立体显微镜平台，用 35×物镜扫描平板上的菌落，生成具有足够分辨率以进行分类的图像。共采集了在富集培养基中培养的七种曲霉菌种的 8995 个原始菌落图像，并进行自动裁剪以生成 17142 个图像块作为训练和测试数据集，其中包含每个菌株的典型分生孢子梗或菌落形态。令人鼓舞的是，Xception 模型在训练图像集上的分类准确率达到 99.8%。在训练后，我们的 CNN 模型在测试图像集上的分类准确率达到 99.7%。基于 Xception 在训练和测试过程中的表现，该分类算法进一步应用于识别和验证这些菌株的一组新的原始图像，其检测准确率达到 98.2%。因此，我们的研究展示了一种基于人工智能且具有成本效益的曲霉菌检测方法的新概念，这也有可能提高公众对真菌王国的认识。
