Faculty talk: Kimmo Kartasalo

On February 5, 2026, Kimmo Kartasalo (Karolinska Institutet and ELLIS Institute Finland/University of Turku) will give a lecture on Foundation models and their limitations in diagnostic computational pathology, as part of ELLIS Institute Finland's faculty day. 

Date and time 

February 5, 2026, 15:00-15:45

Location

Lecture hall TU2, Maarintie 8, 02150 Espoo (Aalto University campus)

Abstract

Deep learning in computational pathology has progressed from accelerating cancer diagnostics to enabling novel prognostic and predictive analyses. Generally applicable, task-agnostic foundation models (FMs) have recently received considerable attention in the field, but their clinically relevant advantages over more conventional task-specific models remain unclear. We recently conducted the largest study on AI for prostate cancer diagnosis and grading to date, analyzing over 100,000 prostate biopsies from more than 7,000 patients across 15 international sites. Comparing two FMs to an end-to-end trained task-specific model in a multiple instance learning setting, we found that while FMs performed well in data-scarce settings, their advantages diminished with sufficient labeled data. Extensive task-adapted training improved overall diagnostic accuracy and reduced variability across digital pathology instruments. Our recent findings based on unique time-series data further suggest that pathology FMs may encode image features that vary over time, potentially leading to inconsistent diagnostic outputs over prolonged use. These potential limitations require further scrutiny for safe and scalable clinical adoption of FM-based diagnostic models.

Bio

Kimmo Kartasalo is an incoming PI at ELLIS Institute Finland and Associate Professor at University of Turku. He has served as an Assistant Professor and a Fellow of the national Data-Driven Life Science (DDLS) program at Karolinska Institutet in Stockholm, Sweden, since 2024. He has a background in biomedical engineering and image analysis, and defended his doctoral thesis at Tampere University, Finland, in 2021 before joining Karolinska Institutet as a post-doc. Kartasalo's group develops machine learning solutions for digital pathology with an applied focus to enhance diagnostic accuracy, efficiency, and reproducibility. The group combines deep learning with large-scale image processing to create decision-support tools that assist pathologists in handling growing workloads. Expanding beyond image analysis, they integrate molecular and clinical data for multi-modal models that predict disease progression and optimal treatments. The group works actively with a wide international network of clinical collaborators to reach these aims and to ensure that the solutions are applicable across diverse clinical settings and patient populations. Kartasalo is also a co-founder of Clinsight, a spin-off company aiming at translating some of the diagnostic models into clinically applicable software products.