Comparative analysis of end-to-end learning and foundation models for clinical-grade prostate pathology
Abstract Foundation models (FMs) using self-supervised pre-training are widely advocated as a universal solution for pathology, yet their advantages over task-specific (TS) models remain unclear. Here we show, using over 100,000 prostate biopsies from 6268 patients across 15 sites, that while FMs demonstrate utility in data-scarce scenarios, their performance converges with—and is in some cases surpassed by—TS models with sufficient labeled training data. Concordance with pathologists’ grading reaches quadratic Cohen’s kappas of 0.91 (95% CI 0.90-0.92; all models) on internal and 0.87 (0.86-0.88; TS), 0.86 (0.84-0.87; UNI FM) and 0.85 (0.84-0.86; Virchow2 FM) on external cohorts at full training. Task-specific training markedly reduces misgrading, misdiagnosis of challenging morphologies, and technical variability. Notably, FMs consume up to 35 times more energy. We advocate for considering both FMs and end-to-end learning for robust, resource-efficient, and clinically applicable AI pathology.
Authors
- Emiel A. M. Janssen (ORCID: https://orcid.org/0000-0002-7760-5396)
- Kristina Hotakainen (ORCID: https://orcid.org/0000-0002-0735-4763)
- Roman Łowicki (ORCID: https://orcid.org/0000-0002-7590-9329)
- J. Asenjo
- Einar G. Gudlaugsson (ORCID: https://orcid.org/0000-0003-2133-3958)
- Radzisław Kordek (ORCID: https://orcid.org/0000-0003-4724-3627)
- Svein Reidar Kjosavik (ORCID: https://orcid.org/0000-0002-7004-5533)
- Bodil Ginnerup Pedersen (ORCID: https://orcid.org/0000-0003-2792-7343)
- Nita Mulliqi (ORCID: https://orcid.org/0009-0002-7448-4960)
- Toyonori Tsuzuki (ORCID: https://orcid.org/0000-0002-4855-4366)
- Henrik Olsson (ORCID: https://orcid.org/0000-0002-2270-2017)
- Marcin Braun (ORCID: https://orcid.org/0000-0003-3804-7042)
- Paolo Libretti
- Sol Erika Boman (ORCID: https://orcid.org/0009-0001-5733-3178)
- Kimmo Kartasalo (ORCID: https://orcid.org/0000-0002-9470-4783)
- Kelvin Szolnoky (ORCID: https://orcid.org/0000-0002-0554-1872)
- Masi Valkonen (ORCID: https://orcid.org/0000-0003-3091-2484)
- Martin Eklund (ORCID: https://orcid.org/0000-0001-5032-5266)
- Päivi O. Väre
- Lars Egevad (ORCID: https://orcid.org/0000-0001-8531-222X)
- Karina Dalsgaard Sørensen (ORCID: https://orcid.org/0000-0002-4902-5490)
- Pekka Ruusuvuori (ORCID: https://orcid.org/0000-0001-9086-9591)
- Anders Blilie (ORCID: https://orcid.org/0009-0002-9546-5443)
- Geraldine Martinez Gonzalez
- Benedicte Parm Ulhøi (ORCID: https://orcid.org/0009-0000-3551-4173)
- Matteo Titus (ORCID: https://orcid.org/0009-0008-7893-7385)
- Marcello Gambacorta
- Julia Anna Mielcarz (ORCID: https://orcid.org/0009-0005-3992-4774)
- Brett Delahunt
- Hemamali Samaratunga
- Xiaoyi Ji
Institutions
- Griffith University (AU)
- University of Helsinki (FI)
- University of Turku (FI)
- The University of Queensland (AU)
- Aichi Medical University (JP)
- Tampere University (FI)
- Malaghan Institute of Medical Research (NZ)
- Aarhus University (DK)
- Stavanger University Hospital (NO)
- Science for Life Laboratory (SE)
- Aarhus University Hospital (DK)
- Karolinska Institutet (SE)
- Länsi-Pohjan Sairaanhoitopiirin Kuntayhtymä (FI)
- Aichi Medical University Hospital (JP)
- SDN Istituto di Ricerca Diagnostica e Nucleare (IT)
- Medical University of Lodz (PL)
- University of Tampere (FI)
- University of Bergen (NO)
- University of Stavanger (NO)
Publication Details
- Journal
- Nature Communications
- Published
- 2026-10-26
- DOI
- https://doi.org/10.1038/s41467-026-78996-7
- Primary Topic
- AI in cancer detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- KAUTE-Säätiö
- Innovationsfonden
- Cancerfonden
- Karolinska Institutet
- Knut och Alice Wallenbergs Stiftelse
- Vetenskapsrådet
- China Scholarship Council
- NordForsk
- Orionin Tutkimussäätiö
- Instrumentariumin Tiedesäätiö
- Science for Life Laboratory
- David och Astrid Hageléns Stiftelse
- Syöpäsäätiö
- Business Finland
- National Supercomputer Centre, Linköpings Universitet