Resumen de Methodology for automatic classification of atypical lymphoid cells from peripheral blood cell images
Edwin Santiago Alférez Baquero
Morphological analysis is the starting point for the diagnostic approach of more than 80% of the hematological diseases. However, the morphological differentiation among different types of abnormal lymphoid cells in peripheral blood is a difficult task, which requires high experience and skill. Objective values do not exist to define cytological variables, which sometimes results in doubts on the correct cell classification in the daily hospital routine. Automated systems exist which are able to get an automatic preclassification of the normal blood cells, but fail in the automatic recognition of the abnormal lymphoid cells. The general objective of this thesis is to develop a complete methodology to automatically recognize images of normal and reactive lymphocytes, and several types of neoplastic lymphoid cells circulating in peripheral blood in some mature B-cell neoplasms using digital image processing methods. This objective follows two directions: (1) with engineering and mathematical background, transversal methodologies and software tools are developed; and (2) with a view towards the clinical laboratory diagnosis, a system prototype is built and validated, whose input is a set of pathological cell images from individual patients, and whose output is the automatic classification in one of the groups of the different pathologies included in the system. This thesis is the evolution of various works, starting with a discrimination between normal lymphocytes and two types of neoplastic lymphoid cells, and ending with the design of a system for the automatic recognition of normal lymphocytes and five types of neoplastic lymphoid cells. All this work involves the development of a robust segmentation methodology using color clustering, which is able to separate three regions of interest: cell, nucleus and peripheral zone around the cell. A complete lymphoid cell description is developed by extracting features related to size, shape, texture and color. To reduce the complexity of the process, a feature selection is performed using information theory. Then, several classifiers are implemented to automatically recognize different types of lymphoid cells. The best classification results are achieved using support vector machines with radial basis function kernel. The methodology developed, which combines medical, engineering and mathematical backgrounds, is the first step to design a practical hematological diagnosis support tool in the near future.
