A foundation model for predicting outcomes of neoadjuvant chemotherapy in breast cancer

作者：Ruichong Lin, Zifan He, Jingjing Han, Zehua Wang, Yongjian Chen, Luhui Mao, Qing Peng, Zebang Zhang, Tang Li, Zhenjun Huang, Haizhu Chen, J L Liang, Olivia Monteiro, Daniel T. Baptista‐Hon, Yanqiong Liu, Yunjie Zeng, Zhenhui Zhao, Huiqin Xu, Kaicong Zhang, Xingbin Hu, Xiaoxi Zhang, Puxiang Lai, Liancheng Yang, Tong Li, Lijun Zheng, Juanjuan Yong, Wei Ren, Kai Chen, Herui Yao, Yunfang Yu · 发表于：International Journal of Surgery · 年份：2025 · DOI：10.1097/js9.0000000000003999 · 被引用次数：2 · 研究领域：Breast Cancer Treatment Studies、Medical Imaging Techniques and Applications、Breast Lesions and Carcinomas

BACKGROUND: Although neoadjuvant chemotherapy (NAC) is a widely adopted approach in the treatment of breast cancer, personalizing the intensity of subsequent adjuvant therapy remains a major clinical challenge due to tumor heterogeneity and the lack of reliable biomarkers. Existing strategies fall short in identifying patients who could benefit from intensive adjuvant chemotherapy, particularly among non-pCR cases. To address this, we developed a foundation model that integrates histopathology and clinical data to support individualized treatment decisions. METHODS: We collected whole-slide images and clinical data from 1,543 patients with non-metastatic invasive breast cancer who underwent planned NAC prior to surgery, across three cohorts: training cohort ( n = 756), validation cohort ( n = 560), and test cohort ( n = 227). A hybrid AI-pathology model was developed, combining a convolutional neural networks branch and a transformer-based foundation model pretrained on The Cancer Genome Atlas (TCGA) whole-slide images. This architecture enabled robust feature extraction from histopathological slides. These features were integrated with clinical data in a multimodal framework to predict pathological complete response (pCR) and disease-free survival (DFS). RESULTS: The proposed hybrid AI-multimodal model demonstrated superior performance for pCR, achieving an area under the receiver operating characteristic curve (AUC) of 0.999 in both the training and validation cohorts. Inte...