Abstract

The PR-IHC-40X dataset comprises a high-resolution collection of region-of-interest (ROI) images and corresponding ground-truth (GT) annotations for progesterone receptor (PR) immunohistochemistry (IHC) analysis in breast cancer pathology. We obtained 50 glass slides from the University of Malaya Medical Centre (UMMC) and digitized them into whole-slide images (WSIs) at 40× magnification using a 3DHistech Pannoramic DESK scanner. Pathologists annotated ROIs on the collaborative Cytomine platform, which formed the basis of dataset extraction. Ground-truth masks were generated in a multi-stage process: binary nuclei masks for segmentation were first created with a StarDist deep learning model and refined by manual correction, while the classification ground truth was first determined using a CNN-based approach and then modified by diaminobenzidine (DAB) intensity thresholding into four expression classes: Strong (red), Moderate (yellow), Weak (green), and Negative (blue). The classification outputs were re-corrected in a loop against the pathologists’ feedback and the manually checked results. There were approximately 32,000 nuclei within 250 ROI images that were manually checked and validated by senior pathologists individually. Each ROI comes with its binary segmentation mask and four-class color annotations, which make it a reliable dataset for deep learning research on nuclei segmentation, PR expression classification, and Allred scoring. To ensure a fair and reproducible evaluation, the dataset is released with a predefined slide-level (patient-wise) partition into training, testing, and evaluation subsets so that no slide contributes regions of interest to more than one subset and data leakage across subsets is avoided.

IPC Classification

Keywords