PANDA-PLUS: Improved dataset of prostate whole slide images from PANDA Challenge with pixel-level expert annotations

Artificial intelligence (AI)-based prostate cancer detection through whole slide images (WSIs) offers promising potential to address the global pathologist shortage while improving clinical consistency. Digital slides and improving image analysis methods encourage the creation of tools to aid in WSI classification. Despite promising advances, these tools are still limited by available training data. Current publicly available datasets, such as Kaggle's PANDA Challenge, while large in scale, rely on slide-level labels that may introduce noise and limit model reliability. Others contain detailed annotations, but are smaller in size due to manual processing efforts. In this work, we introduce PANDA-PLUS, a 546-image dataset derived from PANDA images with improved pixel-level annotations, as well as an accompanying annotation pipeline that reduces pathologists' time commitment. We present a detailed comparative analysis between PANDA-PLUS and PANDA using Gleason score and ISUP grade, supported by agreement values, κ, and PABAK under multiple weighting schemes. The results demonstrate consistently lower grading in PANDA-PLUS, with disagreement patterns especially pronounced at higher grades. We also demonstrate through single rater grading of various annotation granularities how slide- and patch-level labels may distort grading proportions and alter image scores. PANDA-PLUS not only improves annotation granularity and reduces label noise but also exposes potential grading errors in the original PANDA dataset. We present PANDA-PLUS's annotations as an improved alternative to the PANDA labels and conclude that it represents a step forward in the development of higher-quality public datasets for clinical AI applications in prostate cancer pathology.