Artificial intelligence-powered copilots for precision diagnosis and surgical assessment of histological growth patterns in resectable colorectal liver metastases: a prospective study

作者：Ruichong Lin, Yongjian Chen, Yanchun Li, Yujie Tan, Chao Wang, Zehua Wang, Mengyang Sun, Lin Wang, Yufei Wu, Qiyun Ou, Lui Ng, Xiaoxi Zhang, Weidong Pan, Zongyan Li, Zuxiao Chen, Zheyu Zheng, Xiaoming Huang, Lei Zhang, Jingsong Sun, Zaopeng He, Nannan Li, Yunfang Yu, Dawei Zhang · 发表于：International Journal of Surgery · 年份：2025 · DOI：10.1097/js9.0000000000002922 · 被引用次数：6 · 研究领域：AI in cancer detection、Radiomics and Machine Learning in Medical Imaging、Medical Imaging and Analysis

BACKGROUND: Colorectal cancer (CRC) is a leading cause of mortality in China, with metastasis significantly contributing to poor outcomes. Histopathological growth patterns (HGPs) in colorectal liver metastasis (CRLM) provide vital prognostic insights, yet the limited number of pathologists highlights the need for auxiliary diagnostic tools. Recent advancements in artificial intelligence (AI) have demonstrated potential in enhancing diagnostic precision, prompting the development of specialized AI models like COFFEE to improve the classification and management of HGPs in CRLM patients. METHODS: This study developed a Transformer-based deep learning model, COFFEE, for the precise classification of colorectal cancer subtypes using whole-slide images (WSIs) from 431 patients diagnosed with colorectal cancer liver metastasis. The model was pretrained using DINO on 1442 WSIs from the TCGA-COAD cohort, utilizing a Vision Transformer (ViT) architecture to extract 384-dimensional feature vectors from 256 × 256 pixel patches. The proposed model integrates a Transformer-based Multiple Instance Learning (TransMIL) framework, which effectively aggregates spatial and morphological information through multi-head self-attention and Pyramid Position Encoding Generator (PPEG) modules. This design enables efficient handling of large instance sequences within WSIs, allowing for accurate binary and four-class classification. The model was validated on 972 WSIs from a recent dataset, demonstratin...