Transformer-based multimodal deep learning network for predicting postoperative bone metastasis in ER-positive/HER2-negative breast cancer: a multi-institutional retrospective study
Accurate risk stratification is essential for guiding treatment decisions and postoperative management in breast cancer. However, unimodal models are limited in their ability to fully exploit complementary information from clinical, radiological, and histopathological data associated with metastatic progression. We aimed to develop and validate Breast Cancer Bone Metastasis Network (BCBM-Net), a transformer-based multimodal deep learning network for predicting postoperative bone metastasis in patients with estrogen receptor (ER)-positive, human epidermal growth factor receptor 2 (HER2)-negative breast cancer. In this multi-institutional retrospective study, 1,240 patients from five clinical centers were included in training, internal test, and independent external test cohorts. BCBM-Net integrated selected clinicopathological variables, ultrasound (US)-derived features, and pathomics features extracted from whole-slide images (WSIs). Modality-specific fully connected encoders were used to learn nonlinear representations and project heterogeneous features into a shared latent space, followed by adaptive fusion using a multi-head attention transformer. Model performance was evaluated using receiver operating characteristic analysis, DeLong tests, and net reclassification improvement (NRI) analysis and was compared with clinical, US-based, and pathomics unimodal models. BCBM-Net achieved consistently higher discriminative performance than unimodal models across all cohorts, with area under the receiver operating characteristic curve (AUC) of 0.873, 0.866, and 0.855 in the training, internal test, and external test sets, respectively. DeLong tests and NRI analyses further supported the added predictive value of multimodal fusion. BCBM-Net effectively integrates routinely available clinical, US, and histopathological information to provide individualized assessment of 5-year postoperative bone metastasis risk. Its consistent performance in independent external centers supports the value of transformer-based multimodal fusion for cancer prognostication. BCBM-Net could provide a practical framework for risk-adapted surveillance and postoperative management, although prospective and multinational validation is required before clinical implementation.
Authors
- Wenjie Gao (ORCID: https://orcid.org/0000-0001-9456-8884)
- Zhihuai Deng (ORCID: https://orcid.org/0009-0005-0190-0777)
- Youxi Lin
- Zaoqiang Zhang (ORCID: https://orcid.org/0009-0008-5613-0783)
- Bo Gao (ORCID: https://orcid.org/0000-0003-4578-0234)
- Xianjian Qiu
- Guoxi Huang
- Wangim Choi
- Jianan Chen
- Zijia Lai
- Kaiwei Zhang
- Zizhao Wu
- Lu Ouyang
- Yanbo Chen
- Piao Huang
- Zidong Wu
- Jiayuan Zheng
Institutions
- Sun Yat-sen University (CN)
- Jinan University (CN)
- Southern University of Science and Technology (CN)
- Sun Yat-sen Memorial Hospital (CN)
- ShenZhen People’s Hospital (CN)
- Zhujiang Hospital (CN)
- Panyu District Central Hospital (CN)
- Third Affiliated Hospital of Sun Yat-sen University (CN)
- Southern Medical University (CN)
- Guangzhou Medical University (CN)
Publication Details
- Journal
- BMC Medicine
- Published
- 2026-11-16
- DOI
- https://doi.org/10.1186/s12916-026-05352-3
- Primary Topic
- Radiomics and Machine Learning in Medical Imaging
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- National Natural Science Foundation of China
- Natural Science Foundation of Guangdong Province