Staining Resilient Algorithms for Cancer Image Analysis in Digital Pathology

Medical imaging and image-based computer-assisted tools and techniques promise advances in cancer diagnosis, but there are still many unresolved challenges in technical aspects, data standardization, staining variability, and regulatory and ethical approval. These challenges in clinical medicine and pathology include the need to develop methods that are invariant and insensitive to staining variations. In modern clinical practice, digital pathology is indispensable, facilitating efficient image management and analysis through advanced technology that surpasses traditional optical microscopy methods. Technological advances and a growing emphasis on precision medicine have opened the way for developing methods for quantitative pathology assessments. These include methodologies such as whole slide imaging and artificial intelligence-driven solutions, which enable the exploration and extraction of information beyond the capabilities of human visual perception.

In this dissertation, Md. Ziaul Hoque focuses on finding solutions to three research questions by introducing robust algorithms that facilitate stained tissue sample image analysis. The main contributions can be classified into three parts: the first contribution is to solve staining challenges with methods that can quantify individual components of histochemical stains for optimal removal of staining variability. The second contribution is to find solutions for histopathology image registration challenges minimizing the misalignment errors and the risks of inter-laboratory variations. Finally, the third contribution is to optimize and aid staging cancer through nuclei detection and invasion depth estimation with minimal inter-observer and intra-observer variations. The performance evaluations of the proposed methods have been thoroughly performed on eight datasets, and these methods outperform state-of-the-art methods in terms of several performance evaluation metrics. This cancer image analysis research involves the development of advanced algorithms and tools that have far-reaching implications for healthcare and set the stage for more efficient cancer treatment. The findings will serve as a basis for further research on the application and interpretability of AI-driven tools in support of digital pathology and computer-aided cancer diagnosis.

This dissertation was conducted at Center for Machine Vision and Signal Analysis in the Faculty of Information Technology and Electrical Engineering at the University of Oulu, Finland. This research was financially supported by the Academy of Finland Identifying Trajectories of Healthy Aging via Integration of Birth Cohorts and Biobank Data, and the Academy of Finland 6G Flagship projects.