Nita Mulliqi: Towards clinically reliable artificial intelligence for prostate cancer pathology

Accurate histopathological assessment of prostate biopsies is essential for determining prognosis and guiding patient treatment; however, it remains hindered by substantial intra- and inter-observer variability. 

Although deep learning has shown promise in improving diagnostic consistency, real-world deployment is limited by persistent challenges: generalising across laboratories, scanners and patient populations, as well as, handling clinically complex cases. 

More recently, large pathology foundation models have demonstrated broad capabilities, but their performance and practicality in disease-specific tasks such as Gleason grading remain insufficiently understood. 

Our work focused on developing clinically reliable AI systems for prostate pathology, emphasising generalisability and protocol-driven evaluation. To start, we developed tools to access and process proprietary whole-slide formats, enabling large-scale, scanner-agnostic training. 

The central part of the work covers our protocolised development and validation of a weakly supervised, end-to-end AI model for prostate cancer detection and Gleason grading. Trained and evaluated on ~100,000 digitised biopsies from 7,342 patients across 15 laboratories in 11 countries, this represents the largest international retrospective validation of its kind

We compared the task-specific model with pathology foundation models, assessing performance across scanners, patient populations, and clinically challenging scenarios, as well as practical considerations such as energy consumption and scalability.

Finally, part of the work was dedicated to specific clinical applications of our model, such as reducing unnecessary immunohistochemistry (IHC) staining. In a sensitivity-prioritised evaluation across international cohorts, the model identified all malignant slides without false negatives and enabled up to a 44% reduction in IHC use. 

Overall, this work outlines a pathway toward robust, efficient and clinically valuable AI tools for prostate pathology, supporting future prospective trials and real-world implementation.

Nita Mulliqi is a postdoctoral researcher in AI and digital pathology at Karolinska Institutet, where she recently completed her PhD focusing on developing generalisable and protocol-driven AI systems to support pathologists in diagnosing prostate cancer. She has led several international collaborations, including one of the largest retrospective validations of AI for prostate pathology. Her work has been recognised through awards, invited talks, and media features, and contributes to global efforts aimed at improving diagnostic consistency. Nita is also the co-founder and lead AI scientist at Clinsight AB, where she works to translate research innovations into clinical tools that can meaningfully impact patient care.