Improvement, Classification and Interpretation of Cancer Histological Images using Probabilistic Models

Résumé

Histopathological images are commonly used for the diagnosis of cancer and other diseases. These images are also used by Computer Assisted Diagnosis (CAD) systems, which have shown a promising performance on the diagnosis of of cancer and other diseases. However, the images obtained in different laboratories show acquisition differences that hamper the performance of AI-based CAD systems. In particular, color variation is often considered the most relevant issue when working with images from different centers. To accurately reduce color variation it is important to consider the acquisition procedure of histopathological image and, specifically, the staining protocol with two or more stains. Blind Color Deconvolution (BCD) use an observation model that takes the staining protocol into account and separate the stains mixed in the observed image, separating also color from structural information. This thesis studies Bayesian modeling and inference, and their application to BCD techniques. With the proposed approach, it is possible to combine prior knowledge, the observation model, and the data evidence, obtaining robust, high quality posterior distributions that can be used to reduce the effect of color variation on CAD systems. We propose three different Bayesian models for BCD of histopathological images: A Total Variation (TV) framework, Super Gaussian Sparse priors, and Bayesian K-Singular Value Decomposition (BKSVD), and apply them to stain separation, color normalization, stain augmentation, and cancer classification. Two additional contributions are included in this thesis: the improvement of multi-spectral satellite images and network anomaly detection. Both benefit too from the use of Bayesian modeling and inference. Furthermore, we also present three additional works in which we have collaborated but are not part of the compendium of publications presented to obtain the Ph.D. degree: a tutorial paper on the processing of histological images, a paper on blood detection using BCD, and a paper on Deep Variational BCD. The project of this thesis was awarded the 3 Minute Thesis (3MT) prize at the University of Granada, and represented the university at the international Coimbra group competition in 2021.