Sizing and classification of biological particles using ring-wedge detector and neural networks

We describe two methods that are suitable for the sizing and classification of small particles in the size range from 0.1mum to 200mum. One is called diffraction pattern sampling, which uses sampling of the diffraction pattern in the Fourier optical transform plane. The other is a direct image processing in which the microscopic images of the objects are obtained and then direct image processing is used to perform the classification.; For the diffraction sampling method, we emphasize the study of spherical particles. The Fourier transform patterns are used and they are shown to be a good approximation in the forward scattering region. As a result, we show that it is possible to size a group of moving particles using diffraction pattern sampling. A series of experiments is performed using a Ring-Wedge diffraction pattern sampling system to analyze the moving particles. The integral inversion method is used to recover the size distribution of the samples and the results are accurate. We also present a new theoretical analysis of diffraction pattern sampling for a cylindrical configuration that is valid for a 360 degree azimuthal angle. The electromagnetic field in the focal position of a linearly polarized converging cylindrical wave incident normal to the axis of symmetry of the cylinder is derived based on the Maxwell's equations. The scattered field is then studied for different cylinder parameters and for different polarizations of the incident plane wave. This derivation gives out the precise forms for the scattering at any angle and thus verifies the more commonly used forms from Fourier optics. The derivation also establishes a theoretical connection between results at infinity and results in the optical transform plane. This is a new theory for diffraction pattern sampling based on electromagnetic theory as well as Fourier optics.; For the direct image processing, we use the biological samples---diatoms as the investigation agents. A novel method for the automatic classification of biological specimens is presented. This method consists of two major parts. One is a machine vision based preprocessing, and it involves a series of carefully organized algorithms to separate the agent cells from the background and then many more algorithms to calculate the morphological feature values. The other part is a neural network classifier that takes the selected feature values as the inputs. The neural network is refined and optimized in the process and the classification results are excellent. We obtain 80% accuracy for an arbitrary grouping of 8 diatoms and 95% accuracy for a group of diatoms that are visually distinctive. The method is potentially applicable to the auto-classification of all kinds of biological specimens in the real time situation.