Truncated Normal Mixture Prior Based Deep Latent Model for Color Normalization of Histology Images

The variation in color appearance among the Hematoxylin and Eosin (H&E) stained histological images is one of the major problems, as the color disagreement may affect the computer aided diagnosis of histology slides. In this regard, the paper introduces a new deep generative model to reduce the color variation present among the histological images. The proposed model assumes that the latent color appearance information, extracted through a color appearance encoder, and stain bound information, extracted via stain density encoder, are independent of each other. In order to capture the disentangled color appearance and stain bound information, a generative module as well as a reconstructive module are considered in the proposed model to formulate the corresponding objective functions. The discriminator is modeled to discriminate between not only the image samples, but also the joint distributions corresponding to image samples, color appearance information and stain bound information, which are sampled individually from different source distributions. To deal with the overlapping nature of histochemical reagents, the proposed model assumes that the latent color appearance code is sampled from a mixture model. As the outer tails of a mixture model do not contribute adequately in handling overlapping information, rather are prone to outliers, a mixture of truncated normal distributions is used to deal with the overlapping nature of histochemical stains. The performance of the proposed model, along with a comparison with state-of-the-art approaches, is demonstrated on several publicly available data sets containing H&E stained histological images. An important finding is that the proposed model outperforms state-of-the-art methods in 91.67% and 69.05% cases, with respect to stain separation and color normalization, respectively.