Replication Data for: Are nuclear masks all you need for improved out-of-domain generalization? A closer look at cancer classification in histopathology

This dataset is a processed version of the CAMELYON17 dataset used in the NeurIPS 2024 paper "Are nuclear masks all you need for improved out-of-domain generalization? A closer look at cancer classification in histopathology". It consists of patches / tiles from 50 Whole Slide Images (WSIs) (10 WSIs from each of the 5 hospitals) in the CAMELYON17 dataset that have tumour segmentation available. Tiles were picked such that each hospital has equal number of tumourous and non-tumours tiles. Each tile is of size 270x270 pixels. A tile is considered tumourous if the centre region of tile (90x90 pixels in size) has at least 1 pixel that lies inside the tumour segmentation map. The dataset also contains nuclear segmentation masks for all the tiles. Masks were generated using HoVer-Net trained on the CoNSeP dataset.

本数据集为用于 NeurIPS 2024 论文《是否核掩膜就是提高域外泛化能力的全部？病理学中癌症分类的深入研究》的 CAMELYON17 数据集的加工版本。该数据集由 50 张全切片图像（WSIs）的片段/瓦片组成，每家医院提供 10 张图像，其中包含可用于肿瘤分割的信息。瓦片的选择确保了每家医院肿瘤瓦片和非肿瘤瓦片的数量均等。每个瓦片的尺寸为 270x270 像素。若瓦片的中心区域（90x90 像素）中至少有 1 像素位于肿瘤分割映射内，则该瓦片被认定为肿瘤瓦片。此外，数据集还包含了所有瓦片的细胞核分割掩膜。掩膜是通过在 CoNSeP 数据集上训练的 HoVer-Net 生成的。