Validation of diagnostic artificial intelligence models for prostate pathology in a Middle Eastern cohort
Abstract
Abstract Artificial intelligence (AI) for pathology is accelerating cancer diagnostics. For global adoption, studies in currently under-represented populations – where the potential gains from AI may also be greatest – are needed. We present the first external validation study in a Middle Eastern setting, focusing on diagnosis and Gleason grading of prostate cancer. We digitised 339 prostate glass slides from Kurdistan, Iraq, representing a consecutive series of 185 patients from 2013 to 2024. Each slide was scanned on three digital pathology scanners, generating 1,017 whole-slide images. Task-specific and foundation AI models demonstrated AI-pathologist grading concordance similar to pathologist-pathologist concordance (Cohen’s quadratically weighted kappa 0.801 vs. 0.799). Cross-scanner concordance was high (kappa > 0.90) for all AI models across three scanners. The demonstrated robust performance on a low-cost compact scanner could enable validation studies in non-digitalised settings and cost-effective AI adoption in laboratories with limited sample volumes. The publicly shared dataset supports further research to facilitate broader adoption of AI pathology.
