We develop methods for cancer diagnosis using AI models trained on data from digital pathology. We concentrate on diagnosing carcinoma in surgery samples and biopsies, which represents a significant part of the workload in clinical histopathology. Our main aim is to develop semi-automated systems integrated into pathologists' workflow to make the diagnosis faster and, possibly, more reliable while understanding how the AI systems work and what are their limits. As the primary source of data, we consider whole-slide microscopic images of tissue and develop methods based on various machine learning techniques for recognizing characteristic tumor patterns.

Since our main aim is to obtain reliable systems ready for routine use, we concentrate on extensive evaluation and testing and exploring various interpretability methods built into or applied on AI systems. We aim for reproducible systems that can undergo certifications necessary for clinical practice in the future and thus we incorporate provenance generation into the AI pipelines and feed experience with this process into provenance standardization activities (e.g., contributing to ISO standards development).