Intelligent vision system for the detection of protozoa on microscope slides

A machine vision system for microscope slide analysis was developed. The apparatus was configured in a manner that enabled computer control of some tasks traditionally performed by a microscopist. Each sample slide was scanned for analysis utilizing developed software control routines. The system discretized each sample slide into a number of sub regions, which were approximately the size of the camera's field of view. Each region was saved and analyzed independently. Investigations into the physical characteristics of prepared water samples containing protozoa yielded information for the intelligent automation of the physical equipment.; This study involved the development of a system, the characterization of its performance, and an evaluation of its suitability for presumptive and confirmed classification of Giardia and Cryptosporidium on microscope slides. Occurrence and subsequent detection of these protozoa is generally a rate event. Measurements were incorporated into a neural network discrimination scheme that performed a presumptive identification of protozoa on the microscope slide. Regions testing positive for the presumptive test were subjected to a confirmation procedure that tested for the presence of internal structures. Presumptive classification performance using a neural network algorithm was evaluated. Results indicated 100 percent of cysts/oocysts present were appropriately classified. The system developed in this research provides a useful test-bed for future investigations into biological samples on microscope slides.; Three approaches were evaluated for suitability in the confirmation process. These methods were then checked for correspondence between the results indicated by an expert human observer. Most texture measurements alone were not found to be useful. A heuristic method, was computationally more efficient and performed with an 80 percent accuracy. Using a neural network classifier, performance ranged from 50 to 100 percent depending on the parameters tested. Overall, the correspondence between the system and expert suggested a strong relationship to classifications of unknown objects.