← Back to Home

Beyond European and US Data: Prostate Pathology AI Passes External Validation in Iraqi Patient Specimens

Can AI still interpret slides reliably when the patient population, laboratory, and scanning equipment change? A study of 185 Iraqi patients helps fill a gap in Middle Eastern validation data, with grading agreement close to that between pathologists, though hurdles remain before routine diagnostic use.

By SURL BioNews

An AI system that can interpret tissue slides from European and US patients may not deliver the same performance in another healthcare setting. Patient composition, staining procedures, and imaging equipment can all change the details a model sees. A study published on October 10 in *Virchows Archiv* tested prostate cancer pathology AI on specimens from the Kurdistan region of Iraq, adding evidence for Middle Eastern populations that have long lacked validation data.

The study included 339 prostate needle biopsy slides from 185 patients between 2013 and 2024. Each slide was digitized using three scanners, producing 1,017 whole-slide images. These images came from the same set of specimens, rather than additional patients. The models were tasked with detecting cancer and assigning Gleason grades based on tissue architecture; this grading is an important basis for assessing prostate cancer risk and discussing treatment options.

The team compared a model built specifically for prostate pathology with two foundation models, UNI and Virchow2. Foundation models first learn general features from large volumes of pathology images, then apply them to specific interpretation tasks. An earlier DDLS conference abstract had introduced the same set of Iraqi specimens and noted that the patient population, laboratory setting, and scanning equipment all constituted external validation conditions. The preliminary findings in that abstract should be considered separately from the results of the published paper.

The study reported a quadratic-weighted κ of 0.801 for grading agreement between AI and pathologists, compared with 0.799 between pathologists. A preprint of the same study made public in December 2025 also listed these values, with a p-value of 0.9824 for the comparison. κ measures agreement between interpretations and cannot be directly translated into “80% diagnostic accuracy.” The absence of a significant difference also does not establish that AI and physicians are equivalent in every setting.

Differences in equipment presented another test. After the same slides were scanned using Hamamatsu, Leica, and Grundium devices, all models achieved grading agreement κ values above 0.90 across every scanner pairing, including those involving lower-cost compact equipment. This supports laboratories with limited resources using smaller-scale digital equipment to begin building local validation data. However, a model producing similar answers across different scanners does not, by itself, guarantee that those answers are correct.

This evidence also has clear limits: the study was retrospective, the number of patients was limited, and low-grade cancer samples were underrepresented. Scanning with all three devices was still performed at a single site, so the study did not cover variations in laboratory personnel and operating environments. Using physicians’ interpretations as the comparison benchmark also cannot establish whether AI can improve patient outcomes or actually reduce missed diagnoses and workload in routine practice.

The paper explicitly states that the systems tested remain research algorithms and have not received regulatory approval for clinical diagnostic use. This validation provides technical evidence of performance across settings and gives more regions reason to undertake local testing. Broader multicenter studies and evaluations of clinical workflows are needed next to translate agreement on slides into diagnostic benefits that patients actually receive.

References

  1. Virchows Archiv
  2. arXiv
  3. SciLifeLab