New thesis examines how far we are from a trustworthy artificial intelligence in prostate cancer pathology in the clinic
In a new thesis from Karolinska Institutet, PhD student Xiaoyi Ji investigated the generalization and robustness of AI models for computational prostate cancer pathology across multiple sources of variation and examined possible strategies for improvement.
Prostate cancer is one of the most common cancers in men. Diagnosis relies on a pathologist examining tissue samples and grades how aggressive the cancer appears. This is skilled, time-consuming work, and pathologists sometimes reach different conclusions on the same sample, which can affect treatment decisions. In recent years, researchers have trained artificial intelligence (AI) systems to perform prostate cancer diagnosis and grading, and the best models now reach a level of accuracy comparable to experienced pathologists across different data sources.
What are the most important results in your thesis?
“It is crucial to determine whether this technology can be trusted in everyday clinical use. In my studies, we found that AI models can behave differently when analyzing tissue images from a different hospital, a new imaging device, or a different time period than those they were trained on,” says Xiaoyi Ji. “Factors that should not affect the diagnosis such as small differences in how tissue samples are prepared, stained, or digitized can still influence the result from the AI model. Training on larger and more diverse datasets improves performance to some extent, but does not fully solve the problem. Understanding when and why AI systems become unreliable, and how to prevent this, is essential before they can be safely used in clinical practice.”
Why did you become interested in this topic?
“AI in medical imaging has seen remarkable progress in recent years, and it's hard not to be excited about the potential”, continues Xiaoyi. “But as I started working in this field, I noticed a gap between the enthusiasm in research and the caution I encountered in practice. When we spoke with clinicians, equipment vendors, and other researchers, a recurring concern kept coming up: how do we know this actually works reliably? For many people on the ground, quality assurance was the missing piece – just as any engineering technology needs to be rigorously tested before it is put into use, medical AI should be held to the same standard. My supervisors and I felt this concern deserved more attention than it was getting, and that not all aspects of the problem had been fully explored. That's what motivated us to make it the focus of my thesis”.
What do you think should be done in future health care and research?
“I believe there is still some way to go before AI can be confidently integrated into clinical cancer diagnosis. There are underlying challenges that have not yet been fully identified or resolved. I would like to see further research focusing on understanding and preventing these limitations, as well as developing methods to ensure that AI systems remain reliable over time.”
Doctoral thesis
"Towards a trustworthy artificial intelligence in prostate cancer pathology: from quality assurance to clinical applicability". Xiaoyi Ji, 5 June 2026.
Facts about doctoral theses
A doctoral thesis is the final written product of a postgraduate education, which in Sweden corresponds to four years of full-time studies. It varies between different disciplines how a doctoral thesis is structured. In the field of medicine, the doctoral student usually collects three to five scientific articles and presents them together with a thesis summary or overview, a so-called ‘kappa’ (literally meaning overcoat in Swedish). After the doctoral student has passed the public defence of the thesis, he or she receives a doctoral degree (also called a PhD), which is the highest possible educational degree in Sweden. Karolinska Institutet has approximately 2,000 active doctoral students and each year approximately 350 doctoral theses are published at our university.