Heuristic optical detection of bioaerosols

Identifying and differentiating hazardous biological aerosols—airborne micro-organisms that can cause disease—from the natural background poses a significant challenge. Elastic light scattering is a promising technology for this task, by virtue of its low cost, ability to make fast measurements, and lack of a logistics trail, which means the apparatus can operate by itself for long stretches of time. Thus, one could envision deploying a bioaerosol monitor much like a smoke detector, by simply setting it up and then letting it run for months or years on end.

Recent work has demonstrated that elastic scattering signals can be measured from isolated (single) aerosolized particles, both in the laboratory and in the natural environment, by a technique called two-dimensional angular optical scattering (TAOS). The TAOS patterns from different types of aerosol particles (see Figure 1) display a remarkable richness and variety.^1–5 The real challenge has been to extract meaningful information about the scatterer (i.e., the particle) from such images.

Figure 1. Light scattering patterns, shown left to right, from a dioctyl phthalate droplet (Fq-set, narrow rings), a dried polystyrene latex sphere (Pq-set, wide rings), a single Bacillus subtilis spore (Bq-set, bowtie structure), an aggregate of diesel engine soot particles (sq-set, no special structure), and outdoor sampling (K5-set, pseudo-random patchwork).

In principle, the interpretation of light-scattering patterns requires a solution to an inverse problem in electromagnetics, i.e., given a pattern, to infer the properties of the scatterer. However, even a general solution to the forward problem that adequately describes light scattering from atmospheric aerosol particles does not exist, and a general inverse problem solver is even further beyond current abilities.

Two main types of strategy have been outlined to analyze the light-scattering patterns produced by aerosols. One is based on model-driven analysis and understanding. It involves using Maxwell's equations to solve the scattering problem starting from a knowledge of the aerosol particle's size, shape, and refractive index. Specific features of the light-scattering pattern are thus expected to correspond to physical features of the aerosol. For instance, the scatter from spherical particles is rotationally symmetric and gives rise to rings. The intensity and spacing of these rings are determined by the size and refractive index of the particle. However, for non-spherical particles it is much more difficult to perform this inversion (i.e., reconstruct the scatterer from the pattern) and only limited success has been achieved.^2–4

A second strategy is purely data driven, with no prior knowledge about the scatterer being assumed. Early attempts have included the sorting of experimental patterns by morphology (e.g., intensity peaks and valleys, statistical moments).² Computed patterns from the simulated scattering of simple shapes have been ascribed to an experimental class^{6, 7} without requiring a perfect match between observation and calculation.

Our present approach is to employ a learning machine (or ‘classifier’) operating on heuristic principles to discover hidden relations in the data set. The ultimate goal is pattern recognition: i.e., for the computer to automatically tell whether a new TAOS pattern belongs to a class or is an outlier. This approach includes an artificial learning process to design, train and validate a classifier. We have been implementing such classifiers for many years^{8, 9} and they have proven to work well, although only when applied to a relatively small number (tens) of patterns. While we envision that future intelligent classifiers will take advantage of modeling efforts to incorporate pattern features that are specific to set morphologies, we have currently begun testing idealized classifiers—which use no a-priori modeling knowledge—on thousands of experimental patterns.¹⁰

In other words, the interpretation of light-scattering patterns is recast in statistical terms: rather than identifying the scatterer from its pattern, a pattern is assigned to a class by means of an algorithm where feature extraction interacts with linear classification. In our current implementation, the feature extraction module treats the light-scattering pattern as an image, applies a windowed Fourier transform followed by non-linear operations, and yields a feature vector. The linear classification module employs principal components analysis of feature vectors and supervised training, whereby a suitable figure of merit is maximized. Both training and validation rely on sequences of training sets made up of patterns belonging to a known class.

After being validated in this way, the classifier can be applied to recognize and classify other patterns. Assignment of a light-scattering pattern to a class relies on a fusion rule. One goal of classification is to discriminate bacterial spore patterns, i.e., to correctly assign patterns to the Bacillus subtilis class (Bq-set). Figure 2 shows a typical result from an analysis of 957 outdoor dust patterns (K5-set), assumed not to contain any Bacillus subtilis spores. Each pattern in the K5 set is represented by a point on the plane of the first two principal components, and has a color-coded label based on the class to which it is assigned. Applying our classifier to this set, a total of 98 patterns (10%) were falsely recognized as Bq, while the remainder were assigned to the other two training classes, Fq and Pq.¹⁰

Figure 2. Classification of 957 two-dimensional angular optical scattering (TAOS) patterns from outdoor sampling (K5-set). The number assigned to each K5 pattern is plotted on the first two principal components plane {z₁, z₂}. Patterns assigned to the Fq or Pq training classes are color coded cyan or green respectively, while those assigned to the bacterial spore class Bq (false positives) are color coded blue. Pattern numbers shown in red correspond to the six images in the side columns.

Similarly (figures not shown), when the same classifier was applied to the sq pattern set (an aggregate of diesel engine soot particles), it yielded ≃11% false positives. Finally, when applied to the entire Bq set (composed of 91 patterns assumed to come solely from bacterial spores, most of which were not used for training), the classifier correctly recognized 73 out of the 91 Bq patterns. In other words it had ≃20% false negatives. In our ongoing work, we will seek to improve the selectivity of the classifier or at least to assess its performance bounds.

The authors gratefully acknowledge the financial support of Contract W911NF 11-1-0277 R&D 1449-BC-01 granted to the University of Milano-Bicocca by the US Army Research, Development, and Engineering Command (RDECOM) Acquisition Center.