The PAR dataset: Prostate biopsy whole slide images from an underrepresented Middle Eastern population

Technical Summary: The PAR Dataset

Problem Statement
The paper addresses critical gaps in the development and validation of artificial intelligence (AI) for digital pathology. Existing publicly available histopathology datasets are scarce and predominantly represent Western, Caucasian populations, limiting the generalizability of AI models to underrepresented regions such as the Middle East. Furthermore, current datasets often suffer from three specific limitations: (1) they typically rely on a single pathologist's reference standard, hindering the assessment of grading variability; (2) they lack data from underrepresented populations where pathology services are not yet fully digitized; and (3) they usually provide whole-slide images (WSIs) from only one scanner type, preventing the evaluation of cross-scanner robustness. These factors contribute to a "chicken-and-egg" problem where the lack of diverse, multi-scanner data prevents the validation of AI in low-resource settings, which in turn discourages investment in digitization.

Methodology
The authors present the PAR dataset, derived from prostate core needle biopsies collected at PAR Hospital in Erbil, Iraq, between 2013 and 2024. The data curation process involved the following steps:

• Cohort Selection: From an archive of 30,056 cases, the authors filtered for prostate core needle biopsies, merged addenda reports, and confirmed the availability of original formalin-fixed and paraffin-embedded (FFPE) tissue blocks. This resulted in a final cohort of 185 patients (339 unique glass slides).
• Slide Preparation: Original diagnostic slides were not used due to potential degradation. Instead, new H&E-stained slides were prepared from the archival FFPE blocks to ensure consistent, high-quality staining.
• Multi-Scanner Digitization: Each of the 339 glass slides was scanned using three distinct whole-slide scanners to generate three versions per slide:
  1. Hamamatsu NanoZoomer 2.0 HT (0.22 µm/pixel, .ndpi format)
  2. Leica Aperio GT 450 DX (0.26 µm/pixel, .svs format)
  3. Grundium Ocus40 compact scanner (0.25 µm/pixel, .svs format)
• Annotation and Reference Standard: Three pathologists independently assessed the WSIs using the Cytomine platform. Two were general pathologists (one from Iraq, one from Norway) and one was a uropathologist (Norway). They assigned slide-level Gleason Scores (GS) and International Society of Urological Pathology (ISUP) grades. No pixel-level annotations were provided. The grading was performed blindly, with no immediate reconciliation between pathologists.

Key Contributions
The primary contribution is the public release of the PAR dataset via the BioImage Archive (Accession: S-BIAD2323) under a CC BY 4.0 license. The dataset includes:

• 1,017 WSIs: Representing 339 slides from 185 patients, covering three scanner formats.
• Multi-Pathologist Labels: Independent Gleason Scores and ISUP grades (0–5) from three pathologists for the majority of slides (Pathologist I graded all 339; Pathologist II graded 337; Pathologist III graded a stratified subset of 59).
• Demographic Representation: Data originating from an underrepresented Middle Eastern population (Kurdistan Region, Iraq).
• Technical Diversity: Native file formats without conversion, enabling studies on cross-scanner robustness and color normalization.

Results and Data Characteristics
The dataset comprises 185 patients with an age distribution skewed toward older demographics (79 patients aged ≥70 years). The ISUP grade distribution across the 339 slides shows a wide range of malignancy, from benign (164–208 cases depending on the pathologist) to high-grade ISUP 5 (21–86 cases). The dataset structure includes three component folders for the scanner types, an annotations folder containing the independent grading tables, a file list for cross-scanner pairing, and documentation. Technical validation confirmed that all files are openable via standard tools (OpenSlide, Cytomine, ASAP) and that the annotation dictionary is consistent.

Significance
The paper positions the PAR dataset as a resource to support the development and validation of pathology AI models across globally diverse populations. By providing data from a region with limited prior digitization, scanned with multiple scanner types (including a compact, cost-effective model), and annotated by multiple pathologists, the dataset enables:

• Cross-population validation: Testing model generalizability to Middle Eastern populations.
• Robustness evaluation: Assessing AI performance across different scanner manufacturers and models.
• Grading concordance analysis: Studying inter-pathologist variability using multiple independent reference standards.
• Benchmarking: Addressing the broader crisis in biomedical AI benchmarking caused by limited public datasets and institution-specific biases.

The authors state that making this dataset available aims to help bridge inequities in population representation and technological access, moving the field toward more inclusive and reliable AI for pathology.