Resumen de Caracterització morfològica de cèl·lules limfoides normals, reactives, anormals i blàstiques de sang perifèrica mitjançant processament digital d'imatges
Laura Puigví Fernández
The main objective of the present Doctoral Thesis is to obtain new quantitative features by means of digital image processing and machine learning for the differentiation of normal, reactive and malignant lymphoid cells of peripheral blood, contributing to an objective morphologic assessment.The research addresses the following two issues:(1)Using machine learning, geometric, color and texture descriptors are searched which have an explicit quantitative formulation and a reasonable qualitative interpretation in visual morphologic terms.(2)Considering cellular abnormalities established a priori, associated with specific diseases, the aim is to identify specific quantitative descriptors of morphological characteristics that cytologists recognize visually and usually express subjectively.More than 200 patients and 16 different lymphoid cell groups have been included in the research. Using the CellaVision DM96, 12,000 images have been acquired and using the microscope Olympus BX43, 9,000. Almost 2,700 features including geometric, color and texture (first and second-order statistics, granulometric, Wavelet and Gabor) have been analyzed for lymphoid cell differentiation. Six color spaces have been considered: RGB, CMYK, HSV, XYZ, Lab and Luv.The 20 most efficient features for the differentiation between reactive lymphoid cells (infections) and neoplastic cells (abnormal lymphocytes in lymphoma or lymphoid blasts in acute leukemia) have been analyzed. The three most relevant descriptors for the recognition of the 12 lymphoid cell groups considered are geometric: 1) nucleus/cytoplasm ratio, 2) nuclear perimeter and 3) cell diameter. Most of the 20 descriptors show significant differences between pairs of abnormal lymphocytes which are difficult to recognize by morphology. Five color and texture features are significant to discriminate reactive lymphocytes from abnormal lymphoid cells.Quantitative descriptors have been identified for the detection of specific cell morphologic abnormalities of certain lymphoid neoplasms, which have shown good specificity and sensitivity using the two different image acquisition systems. Regarding nuclear abnormalities, the detection of the mature and condensed chromatin seen in chronic lymphatic leukemia cells has been achieved by the correlation of the cyan of the nucleus, and the cerebriform chromatin characteristic of Sézary cells, by means of the standard deviation of the granulometric curve of the cyan component of the nucleus.Regarding the cytoplasm, hairiness descriptor has been able to detect cytoplasmic villi present in villous lymphocytes in hairy cell leukemia and splenic marginal zone lymphoma. The skewness of the histogram of the u component of the cytoplasm has shown to be useful for detecting azurophilic cytoplasmic granules seen in abnormal lymphocytes in T-cell large granular lymphocytic leukemia.The results of this Doctoral Thesis provide objectivity in the morphologic assessment of normal, reactive and neoplastic lymphoid cells. Obtaining quantitative descriptors for abnormal lymphoid cells, which are specific to certain lymphoid neoplasms with peripheral blood expression, could facilitate their detection. Hematological analyzers based on digital image analysis could benefit from the use of quantitative descriptors, such as those described herein, in order to discriminate between reactive and neoplastic lymphoid cells.
