基于显微高光谱成像技术判别食源性致病菌种类的方法研究

A Novel Classification Method of Foodborne Bacterial Species Based on Hyperspectral Microscopy Imaging Technology

如何实现对食源性致病微生物的早期快速检测是全球食品安全领域面临的挑战之一。传统的致病菌生化检测方法虽然准确性高,但是过程复杂、耗时漫长,易错过控制疫情爆发的窗口期。该研究提出一种基于暗场显微高光谱成像技术的检测方法,能够借助显微镜技术突破传统光谱成像的灵敏度和分辨率极限,并利用可见/近红外光谱为单个致病菌细胞添加高分辨率的图像和光谱信息。以空肠弯曲杆菌、大肠埃希氏菌O157:H7和鼠伤寒沙门氏菌为检测对象,采用显微高光谱成像技术进行数字化表征和数据采集,结合双向长短期记忆网络(Bi-LSTM)算法对各致病菌细胞的图像和光谱数据进行建模分类。结果显示,显微尺度的致病菌光谱数据呈现可判别的分布规律,新采用的Bi-LSTM网络在光谱数据集中表现优异,在三种致病菌的分类任务中取得了91.0%的平均准确率,而传统的线性判别分类器(LDA)和支持向量机分类器(PCA-SVM)的平均准确率分别为80.1%和88.5%。但是,仅依赖光谱数据进行致病菌种类判别仍然存在较为严重的假阳性问题,尤其是在大肠埃希氏菌O157:H7和鼠伤寒沙门氏菌的分类中存在错误分类。图像信息的加入则能够显著改善各分类的识别准确率,其中Bi-LSTM分类器取得了高达98.1%的准确率,LDA和PCA-SVM均取得了95.3%的准确率。研究结果表明,显微高光谱成像技术在食源性致病菌的特异性光谱和图像表征中具有优势,提出的Bi-LSTM网络能够直接处理高维的图谱特征,两种技术的融合在食源性致病菌细胞级别的早期检测应用中展现潜力。

The early and rapid detection of food-borne pathogens still challenges global food safety.Although routine tests such as selective agars served as the gold standard for decades,these methods are time consuming,often exceeding the best control time for foodborne bacterial outbreak.A hyperspectral microscopic imaging(HMI)technology coupled with an artificial intelligence algorithm was proposed to detect the common foodborne bacteria.The HMI technology has a natural trait in generating robust high-resolution spatial and spectral characterization at the cellular level.In this study,the Bi-directional long short-term memory(Bi-LSTM)was employed in the classification of Campylobacter jejuni(C.jejuni),Escherichia coli O157:H7(E.coli),Salmonella Typhimurium(S.Typhimurium)based on morphological and spectral features extracted from single cell hypercubes.Compared with traditional linear discriminant analysis(LDA,80.1%)and principal components analysis with support vector machine(PCA-SVM,88.5%)classifiers,our proposed Bi-LSTM achieved the highest accuracy of 91.0%on the spectral dataset.Serious false-positive problems occurred in recognising E.coli and S.Typhimurium.However,with the involvement of morphological features,the discriminability of all classifiers was significantly improved.The proposed Bi-LSTM classifier achieved the highest accuracy of 98.1%based on the morphological-spectral feature dataset,while the LDA and PCA-SVM all achieved an accuracy of 95.3%.Our study demonstrated the applicability of HMI technology for foodborne bacterial cell characterization.Furthermore,with the advantage of Bi-LSTM in instantly processing the high-dimensional spatial-spectral features,the intelligent HMI shows great potential for rapid detection of foodborne pathogens.