《生命科学》 2026, 38(7): 1201-1207
人工智能在乳腺癌HER2精准判读中的应用进展
摘 要：
乳腺癌是女性最常见的恶性肿瘤之一，HER2作为其重要的分子标志物，对于指导乳腺癌靶向治疗具有重 要意义。依赖免疫组化和荧光原位杂交的传统HER2判读方法存在主观性强、重复性差等问题。近年来，人工智能 技术在病理图像分析领域迅速发展，为HER2的智能化、标准化判读提供了新思路。本文围绕人工智能在HER2判 读中的应用进展进行综述，探讨其优劣势及未来发展方向，为推动人工智能在乳腺癌精准诊断中的临床转化提供 参考。
通讯作者：袁静萍 , Email:yuanjingping@whu.edu.cn
Abstract:
Breast cancer is one of the most frequently diagnosed malignant neoplasms in women globally. As an essential molecular biomarker for breast cancer stratification, human epidermal growth factor receptor 2 (HER2) plays a decisive role in guiding individualized anti-tumor targeted therapy. Immunohistochemistry (IHC) and fluorescence in situ hybridization (FISH) constitute the routine diagnostic methods for clinical HER2 status evaluation. Manual slide interpretation based on these two detection techniques has prominent inherent defects, including marked subjective bias among pathologists, poor inter-laboratory reproducibility, long diagnostic turnaround time, and limited diagnostic performance for lesions with intratumoral spatial heterogeneity, which creates prominent diagnostic difficulties particularly for patients with HER2-low expression. The popularization of digital whole-slide imaging technology and the continuous improvement of computational pathological analysis algorithms have driven rapid progress of intelligent pathological diagnosis over the past decade, which provides a new technical path to realize objective, standardized automatic HER2 evaluation. Existing relevant research can be divided into three primary research directions: predictive models that infer HER2 expression status only from conventional hematoxylin-eosin (HE) stained sections, automatic scoring systems for IHC slices to classify HER2 expression grades, and quantitative analysis algorithms designed for FISH amplification signal counting. Multiple technical frameworks, including weakly supervised multiple instance learning, transfer learning and multi-modal contrastive learning, have been developed to extract tumor morphological characteristics, realize cross-modal feature matching and complete HER2 classification tasks. Related validation data indicate that most of these models obtain area under the curve values between 0.75 and 0.90 in internal cross-verification and external cohort testing. Nevertheless, such automated analytical approaches still possess obvious application constraints. The predictive efficacy of these models highly depends on indirect tissue morphological characteristics, and their diagnostic stability decreases sharply in specimens with high tumor heterogeneity. Other limiting factors include insufficient multi-center verified data sets, insufficient model generalization ability, and imperfect clinical transformation standards. The present review systematically sorts out the research progress of computational pathological techniques in HER2 automatic diagnosis, comprehensively compares the advantages and inherent defects of different analytical algorithms, and expounds future research orientations such as interpretable pathological models, multi-modal combined diagnosis and prospective multi-center clinical verification. This work is expected to offer theoretical references for advancing the clinical translation of digital pathological detection in precision diagnosis of breast malignant tumors.
Communication Author：YUAN Jing-Ping , Email:yuanjingping@whu.edu.cn
SciCloud投稿系统