dellacorte/PANDA-PLUS-Bench

PANDA-PLUS-Bench是一个用于评估病理学基础模型中WSI特定特征崩溃的基准数据集。它包含来自9个全切片图像(9名不同患者)的专家标注前列腺活检图像块，具有像素级的Gleason模式标注。数据集包含约2,770个图像块(每种增强条件)，分辨率为224×224像素(20倍放大)，分为4个类别：良性(0)、GP3(1)、GP4(2)和GP5(3)。提供8种不同的增强条件，包括基线(仅ImageNet标准化)、颜色抖动、灰度化、高斯噪声、重度几何变换、组合增强、Macenko标准化和H/E染色增强。

PANDA-PLUS-Bench is a benchmark dataset for evaluating WSI-specific feature collapse in pathology foundation models. It contains expert-annotated prostate biopsy patches from 9 whole slide images (9 unique patients) with pixel-level Gleason pattern annotations. The dataset includes ~2,770 patches per augmentation condition at 224×224 pixel resolution (20× magnification), classified into 4 categories: Benign (0), GP3 (1), GP4 (2), and GP5 (3). It provides 8 augmentation conditions including baseline (ImageNet normalization only), color jitter, grayscale, gaussian noise, heavy geometric, combined aggressive, Macenko normalization, and HED stain augmentation.