2026-10-14T12:45:35Z https://www.scidb.cn/oai

10.57760/sciencedb.45789 2026-09-10T14:03:18Z

2026-09-10 Raw Dataset for Bone Metastasis Prediction in ER-Positive/HER2-Negative Breast Cancer doi:10.57760/sciencedb.45789 en This dataset was derived from a multicenter retrospective study that investigated multimodal medical data for predicting the risk of bone metastasis in patients with estrogen receptor-positive/human epidermal growth factor receptor 2-negative (ER-positive/HER2-negative) breast cancer. Data were collected from five medical centers between January 2010 and December 2024. All included patients had ER-positive/HER2-negative breast cancer and met the predefined inclusion and exclusion criteria. To facilitate validation and reproducibility of the associated research findings, the publicly released dataset includes de-identified baseline clinical data, breast ultrasound images in Digital Imaging and Communications in Medicine (DICOM) format, digital histopathology patches cropped from hematoxylin and eosin (H&E)-stained whole-slide images (WSIs), and intermediate data generated during model training, feature processing, validation, and performance evaluation. Collectively, these data constitute the basis for the development and validation of the clinical, ultrasound, and digital pathology components of the multimodal prediction models used in the study.Baseline clinical data were obtained from the electronic medical records and pathology information systems of the participating centers and were organized according to standardized data definitions and the predefined study protocol. The publicly available dataset contains only de-identified clinical and pathological variables relevant to model development and statistical analyses, outcome information, and dataset allocation information. Direct identifiers, including patient names, hospitalization numbers, national identification numbers, contact information, and detailed residential addresses, are not included. Date-related information and other variables that could potentially increase the risk of patient re-identification were removed, transformed, or minimized before public release in accordance with the study's data de-identification procedures.Breast ultrasound images were obtained from routine ultrasound examinations performed during clinical care. The images were acquired using ultrasound systems from multiple manufacturers, including GE Healthcare, Mindray, Siemens, Philips, Samsung, and Canon. The publicly released images are stored in DICOM format and retain, to the extent possible, the original pixel matrices, grayscale or color information, and non-identifying imaging parameters required for model analysis. Before data release, the DICOM images underwent systematic de-identification, including the removal or replacement of patient names, patient identifiers, dates of birth, hospitalization numbers, accession numbers, institutional information, operator information, and other potentially identifying information. File names were also anonymized. Anonymous study identifiers are used solely to link different data modalities within the study. The linkage between the original patient identities and the anonymous study identifiers is neither included in the publicly available dataset nor made externally available.Digital pathology data were derived from H&E-stained slides prepared from surgical pathological specimens. The original glass slides were digitized using a KF-FL-400 whole-slide scanner (Jiangfeng Bio-Information Technology Co., Ningbo, China) at 20× magnification. Because complete WSIs are large in file size and may contain slide labels, barcodes, macro images, and other associated information with potential identifying content, the publicly released dataset primarily provides the digital pathology patches actually used for model development and validation rather than the complete original WSIs for unrestricted public download. Histopathology patches were cropped from the original WSIs, with each patch measuring 512 × 512 pixels and stored in JPG format using anonymized file names. Slide labels, barcodes, and other patient-identifying information unrelated to model analysis are not included in the publicly released patch dataset.In addition to the clinical data, ultrasound DICOM images, and digital pathology patches, the dataset contains derived and intermediate data generated during model training and validation to improve research transparency and reproducibility. These data include, but are not limited to, anonymized dataset partition information, mappings between images and patients, feature data generated through model-based feature extraction, feature matrices used as downstream model inputs, data obtained after feature selection or transformation, patient-level ground-truth outcome labels, and patient-level prediction results generated during model training and validation. For analytical procedures involving data standardization, feature transformation, feature aggregation, or feature selection, the corresponding parameters, processing outputs, or information required to reproduce these procedures are also provided with the dataset or the associated publicly available code. All intermediate data are linked to the original study cohorts using anonymous study identifiers, enabling researchers to reproduce model training, validation, and the principal statistical analyses without access to personally identifiable information.All clinical data, ultrasound images, pathology images, and derived data originating from the same patient were assigned to the same dataset partition to prevent patient-level data leakage.This dataset is intended to facilitate the validation and reproduction of the principal findings of the associated study and to provide a reusable resource for research on bone metastasis risk prediction in ER-positive/HER2-negative breast cancer, artificial intelligence in medical imaging, digital pathology, and multimodal deep learning. None of the publicly released files contain original patient names, hospitalization numbers, personally identifying information, or the linkage between patient identities and anonymous study identifiers. Users of the dataset are required to comply with the applicable data-use license, ethical requirements, and privacy-protection regulations specified by the data repository and must not attempt to re-identify any study participant. Breast cancer; Bone metastasis; Ultrasound; Digital pathology; Deep learning Zaoqiang Zhang Wenjie Gao PUBLIC https://creativecommons.org/licenses/by-nc-nd/4.0/ dataset Science Data Bank