Venkatapathi Murugesan, Rajwa Bartek, Ragheb Kathy, Banada Padmapriya P, Lary Todd, Robinson J Paul, Hirleman E Daniel
School of Mechanical Engineering, Purdue University, West Lafayette, IN 47907, USA.
Appl Opt. 2008 Feb 10;47(5):678-86. doi: 10.1364/ao.47.000678.
We describe a model-based instrument design combined with a statistical classification approach for the development and realization of high speed cell classification systems based on light scatter. In our work, angular light scatter from cells of four bacterial species of interest, Bacillus subtilis, Escherichia coli, Listeria innocua, and Enterococcus faecalis, was modeled using the discrete dipole approximation. We then optimized a scattering detector array design subject to some hardware constraints, configured the instrument, and gathered experimental data from the relevant bacterial cells. Using these models and experiments, it is shown that optimization using a nominal bacteria model (i.e., using a representative size and refractive index) is insufficient for classification of most bacteria in realistic applications. Hence the computational predictions were constituted in the form of scattering-data-vector distributions that accounted for expected variability in the physical properties between individual bacteria within the four species. After the detectors were optimized using the numerical results, they were used to measure scatter from both the known control samples and unknown bacterial cells. A multivariate statistical method based on a support vector machine (SVM) was used to classify the bacteria species based on light scatter signatures. In our final instrument, we realized correct classification of B. subtilis in the presence of E. coli,L. innocua, and E. faecalis using SVM at 99.1%, 99.6%, and 98.5%, respectively, in the optimal detector array configuration. For comparison, the corresponding values for another set of angles were only 69.9%, 71.7%, and 70.2% using SVM, and more importantly, this improved performance is consistent with classification predictions.
我们描述了一种基于模型的仪器设计，并结合统计分类方法，用于开发和实现基于光散射的高速细胞分类系统。在我们的工作中，使用离散偶极近似法对枯草芽孢杆菌、大肠杆菌、无害李斯特菌和粪肠球菌这四种感兴趣的细菌细胞的角向光散射进行了建模。然后，我们在一些硬件约束条件下优化了散射探测器阵列设计，配置了仪器，并从相关细菌细胞中收集了实验数据。利用这些模型和实验表明，在实际应用中，使用标称细菌模型（即使用代表性尺寸和折射率）进行优化不足以对大多数细菌进行分类。因此，计算预测以散射数据向量分布的形式构成，该分布考虑了四个物种内单个细菌之间物理特性的预期变异性。在使用数值结果对探测器进行优化后，它们被用于测量已知对照样品和未知细菌细胞的散射。一种基于支持向量机（SVM）的多元统计方法被用于根据光散射特征对细菌种类进行分类。在我们的最终仪器中，在最佳探测器阵列配置下，使用支持向量机分别以99.1%、99.6%和98.5%的准确率实现了在存在大肠杆菌、无害李斯特菌和粪肠球菌的情况下对枯草芽孢杆菌的正确分类。相比之下，对于另一组角度，使用支持向量机的相应准确率仅为69.9%、71.7%和70.2%，更重要的是，这种改进的性能与分类预测一致。
