Normalização de cores de imagens histológicas coradas com H&E usando aprendizagem de dicionário
Arquivos
Data
2024-09-09
Autores
Silva, André Fernando Quaresma da 
Orientadores
Tosta, Thaína Aparecida Azevedo
Tipo
Trabalho de conclusão de curso
Título da Revista
ISSN da Revista
Título de Volume
Resumo
O câncer é um dos principais problemas de saúde no Brasil e no mundo, destacando-se por sua alta incidência e mortalidade. O diagnóstico da doença envolve várias etapas, incluindo a preparação e análise microscópica de amostras de tecidos, comumente coradas com Hematoxilina e Eosina (H&E). A obtenção de imagens histológicas oferece uma vantagem relevante em comparação com a análise manual das lâminas histopatológicas físicas. Isso porque a digitalização possibilita o uso de ferramentas computacionais que simplificam o compartilhamento de informações e permitem a aplicação de algoritmos para a análise e processamento de dados. No entanto, a variação inadequeda de cores nas imagens de tecido digitalizadas representa um desafio significativo para o desenvolvimento de sistemas de diagnósticos auxiliados por computador. A aplicação de métodos de normalização em estudos de processamento de imagens visa corrigir a aparência das amostras histológicas digitalizadas, com o propósito de apoiar patologistas e métodos de análise de imagens histológicas. O principal objetivo deste trabalho é investigar o uso de técnicas de aprendizagem de dicionário para normalização dos corantes H&E em imagens histológicas. A ferramenta hpcNMF foi utilizada para estimar a matriz de aparência de cores dos corantes H&E, permitindo a normalização das cores e a preservação das estruturas dos tecidos. Foram analisadas diferentes técnicas de normalização, comparando seu desempenho em bases de imagens histológicas públicas com variações de cores dos corantes H&E. Os resultados obtidos indicam que as técnicas de aprendizagem de dicionário investigadas na ferramenta hpcNMF foram mais eficientes na preservação das características dos tecidos do que as abordagens de normalização presentes na literatura, com valores máximos de FSIM, PSNR, QSSIM e SSIM de aproximadamente 0,86, 41,48, 0,91 e 0,96, respectivamente. A análise da influência da normalização sobre a segmentação de núcleos celulares demonstrou que a qualidade visual das imagens normalizadas não apresenta correlação simples com os resultados quantitativos da segmentação. Assim, este trabalho apresenta questões em aberto para o desenvolvimento de pesquisas futuras diversas.
Cancer is one of the leading health issues in Brazil and worldwide, characterized by its high incidence and mortality. Diagnosing the disease involves several stages, including the preparation and microscopic analysis of tissue samples, commonly stained with hematoxylin and eosin (H&E). Obtaining histological images offers a significant advantage over manual analysis of physical histopathological slides, as digitization enables the use of computational tools that simplify information sharing and allow the application of algorithms for data analysis and processing. However, the inadequate color variation in digitized tissue images represents a significant challenge for the development of Computer-Aided Diagnosis systems. The application of normalization methods in image processing studies aims to standardize the appearance of digitized tissue samples to support pathologists and histological image analysis methods. The primary objective of this study is to investigate the use of a dictionary learning technique for H&E stain normalization in histological images. The hpcNMF tool was used to estimate the appearance matrix of H&E stain colors, allowing for color normalization while preserving tissue structures. Various normalization techniques were analyzed, comparing their performance across several public histological image datasets with H&E stain color variations. The results obtained indicate that the dictionary learning techniques investigated using the hpcNMF tool were more efficient in preserving tissue characteristics than the normalization approaches in the literature, with maximum values of FSIM, PSNR, QSSIM, and SSIM of nearly 0.86, 41.48, 0.91, and 0.96, respectively. The analysis of the influence of normalization on cell nuclei segmentation demonstrated that the visual quality of the normalized images does not simply correlate with the quantitative results of segmentation. Therefore, this study raises open questions for the development of future research.
Cancer is one of the leading health issues in Brazil and worldwide, characterized by its high incidence and mortality. Diagnosing the disease involves several stages, including the preparation and microscopic analysis of tissue samples, commonly stained with hematoxylin and eosin (H&E). Obtaining histological images offers a significant advantage over manual analysis of physical histopathological slides, as digitization enables the use of computational tools that simplify information sharing and allow the application of algorithms for data analysis and processing. However, the inadequate color variation in digitized tissue images represents a significant challenge for the development of Computer-Aided Diagnosis systems. The application of normalization methods in image processing studies aims to standardize the appearance of digitized tissue samples to support pathologists and histological image analysis methods. The primary objective of this study is to investigate the use of a dictionary learning technique for H&E stain normalization in histological images. The hpcNMF tool was used to estimate the appearance matrix of H&E stain colors, allowing for color normalization while preserving tissue structures. Various normalization techniques were analyzed, comparing their performance across several public histological image datasets with H&E stain color variations. The results obtained indicate that the dictionary learning techniques investigated using the hpcNMF tool were more efficient in preserving tissue characteristics than the normalization approaches in the literature, with maximum values of FSIM, PSNR, QSSIM, and SSIM of nearly 0.86, 41.48, 0.91, and 0.96, respectively. The analysis of the influence of normalization on cell nuclei segmentation demonstrated that the visual quality of the normalized images does not simply correlate with the quantitative results of segmentation. Therefore, this study raises open questions for the development of future research.