Automatic recognition of different types of acute leukaemia using peripheral blood cell images

Abstract

[eng] Clinical pathologists have learned to identify morphological qualitative features to characterise the different normal cells, as well as the abnormal cell types whose presence in peripheral blood is the evidence of serious haematological diseases. A drawback of visual morphological analysis is that is time consuming, requires well-trained personnel and is prone to intra-observer variability, which is particularly true when dealing with blast cells. Indeed, subtle interclass morphological differences exist for leukaemia types, which turns into low specificity scores in the routine screening. They are well-known the difficulties that clinical pathologists have in the discrimination among different blasts and the subjectivity associated with their morphological recognition. The general objective of this thesis is the automatic recognition of different types of blast cells circulating in peripheral blood in acute leukaemia using digital image processing and machine learning techniques. In order to accomplish this objective, this thesis starts with a discrimination among normal mononuclear cells, reactive lymphocytes and three types of leukemic cells using traditional machine learning techniques and hand-crafted features obtained from cell segmentation. In the second part of the thesis, a new predictive system designed with two serially connected convolutional neural networks is developed for the diagnosis of acute leukaemia. This system was proved to distinguish neoplastic (leukaemia) and non-neoplastic (infections) diseases, as well as recognise the leukaemia lineage. Furthermore, it was evaluated for its integration in a real-clinical setting. This thesis also contributes in advancing the state of the art of the automatic recognition of acute leukaemia by providing a more realistic approach which reflects the real-life complexity of acute leukaemia diagnosis.