Determining Leishmania Infection Levels by Automatic Analysis of Microscopy Images
share
Analysis of microscopy images is one important tool in many fields of biomedical research, as it allows the quantification of a multitude of parameters at the cellular level. However, manual counting of these images is both tiring and unreliable and ultimately very time-consuming for biomedical researchers. Not only does this slow down the overall research process, it also introduces counting errors due to a lack of objectivity and consistency inherent to the researchers own human nature. This thesis addresses this issue by automatically determining infection indexes of macrophages parasite by Leishmania in microscopy images using computer vision and pattern recognition methodologies. Initially images are submitted to a pre-processing stage that consists in a normalization of illumination conditions. Three algorithms are then applied in parallel to each image. Algorithm A intends to detect macrophage nuclei and consists of segmentation via adaptive multi-threshold, and classification of resulting regions using a set of collected features. Algorithm B intends to detect parasites and is similar to Algorithm A but the adaptive multi-threshold is parameterized with a different constraints vector. Algorithm C intends to detect the macrophages and parasites cytoplasm and consists of a cut-off version of the previous two algorithms, where the classification step is skipped. Regions with multiple nuclei or parasites are processed by a voting system that employs both a Support Vector Machine and a set of region features for determining the number of objects present in each region. The previous vote is then taken into account as the number of mixtures to be used in a Gaussian Mixture Model to decluster the said region. Finally each parasite is assigned to, at most, a single macrophage using minimum Euclidean distance to a cell nucleus, thus quantifying Leishmania infection levels.
READ FULL TEXT
