Método computacional para segmentação não supervisionada de imagens histológicas de linfoma

• Autor(a) principal: Tosta, Thaína Aparecida Azevedo
• Data de Publicação: 2016
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFU
• Texto Completo: https://repositorio.ufu.br/handle/123456789/12594; http://doi.org/10.14393/ufu.di.2016.17
• topic: Segmentação nuclear; Imagens histológicas de linfoma; Linfoma de células do Manto; Linfoma folicular; Leucemia linfoide crônica; Algoritmo genético; Computação; Processamento de imagens; Câncer - Diagnóstico; Linfoma - Diagnóstico; Nuclear segmentation; Histological images of lymphoma; Mantle cell lymphoma; Follicular lymphoma; Chronic Lymphocytic leukemia; Genetic algorithm; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

Histological image analysis represents a major evolutionary step in modern medicine. Associated with this step, computational methods are being widely developed to help specialists during the analysis of these images to determine diagnostics, prognostics and appropriate treatments in accordance with the condition of the patient. However, when it is performed by specialists, this task becomes time-consuming and susceptible to inter- and intra-pathologist variability. To improve this traditional practice for diagnostics of Mantle Cell Lymphoma, Follicular Lymphoma and Chronic Lymphocytic Leukemia, this study proposes a method for the unsupervised segmentation of nuclear components in indicative cells of such neoplasias using histological images stained with Hematoxylin-Eosin. The proposed method was divided into preprocessing, segmentation and post processing. In the preprocessing step, the techniques used in histogram equalization and Gaussian filter were applied to the channels from RGB color model. In the segmentation, a thresholding technique was applied combining the methods of fuzzy 3-partition entropy and genetic algorithm. Finally, for the improvement of the segmentation results, morphological operations and the valley-emphasis technique were used. For evaluating the developed method, histological images of lymphoma with magnification 20x were selected and manually segmented by a specialist. Those reference images (gold standard) allowed the extraction of quantitative measures in order to compare this method with different techniques proposed in the literature. Furthermore, a qualitative evaluation was conducted leading to relevant and improved results over those from compared studies. Its application was also analysed considering the steps of feature extraction and classification of the lesions, obtaining results of accuracy close to 100%