A stain color normalization with robust dictionary learning for breast cancer histological images processing

Abstract

Microscopic analyses of tissue samples are crucial for confirming the diagnosis of breast cancer. The digitization of these samples has led to the development of computational systems that can assist pathologists. However, these systems may face limitations owing to color variations in the images. Normalization studies have been widely conducted to address these issues, but there is still a need for new proposals that take into account the biological properties of dyes and tissues. This study presents a novel method for normalizing hematoxylin and eosin-stained histological images by estimating the color appearance matrices and density maps of the stain. The proposed method offers contributions in terms of pixel selection and weight definition to improve the color estimation of histological images. Besides, to the best of our knowledge, no previous studies have evaluated normalized images considering both handcrafted and learning features. Breast cancer images with significant color variations were used to evaluate this approach and the results demonstrated its effectiveness and efficiency. The average values of FSIM, NIQE, and QSSIM were up to 0.9866, 3.4298, and 0.9655, respectively. Compared with other normalization techniques, the proposed method showed an increase of up to 5.9261, with the largest difference observed in the amount of noise added, as indicated by the NIQE metric. To determine the impact of normalization on feature extraction, the evaluations included an analysis of both color and deep-learned features. These experiments showed that all evaluated methods harmed the separation of breast cancer samples by color features. In contrast, the deep-learned features resulted in less complex classification problems, especially with the proposed normalization. This technique also reached one of the lowest processing times, nearly 6 s with the largest image from the databases.