PO.BCS02.03 · 生物信息与计算
AI辅助的实体瘤Ki67评估：与病理专家相比的一致性与可靠性
AI-assisted Ki67 evaluation in solid tumors: Consistency and reliability compared to expert pathologists
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
背景：Ki67是实体瘤中的关键增殖标志物，在指导HR+/HER2-乳腺癌辅助治疗方面尤为相关。尽管其具有重要的临床意义，Ki67免疫组化（IHC）评分仍缺乏标准化。国际指南旨在减少病理医师之间的变异，而AI驱动的图像分析解决方案近期已作为快速且可靠的替代方案出现。本研究在一个大型实体瘤队列中，将使用Aiforia®（AI平台）和Halo®（有监督图像分析软件）的Ki67评分与三位独立病理医师的评分进行比较。
方法：我们对192例不同来源的肿瘤（包括乳腺癌和前列腺癌）进行了Ki67（克隆号30-9）染色。病理医师按照国际Ki67工作组（IKWG）建议接受培训，并据此对组织进行评分。基于深度学习的Aiforia®在数分钟内自动定量Ki67阳性肿瘤细胞。Halo®采用随机森林分类器对肿瘤区、非肿瘤区和背景区进行分割，并经病理医师验证。在细胞分割后，通过阈值法确定Ki67阳性率。
结果：在所有实体瘤中，Aiforia®与Halo®之间的Ki67评分显示出高度一致性（r² = 0.95）。尽管接受了标准化培训，病理医师之间的相关性较弱（A-B：r² = 0.78；A-C：r² = 0.86；B-C：r² = 0.85），但仍可接受。在所分析的19种肿瘤类型中，仅甲状腺和胃的软件相关性低于0.75，其中胃的病理医师间一致性尚可（r² > 0.80），甲状腺则较低（r² = 0.65-0.82）。病理医师之间存在显著的器官特异性变异，而软件评分保持一致。对于乳腺肿瘤（n = 16），Aiforia®与Halo®相关性很强（r² = 0.96），而病理医师之间的一致性范围为r² = 0.60至0.87。
结论：Aiforia®等基于AI的平台以及Halo®等有监督图像分析工具提供了稳健、可重复的Ki67评分，并显著降低了观察者间变异。这些技术为基于IHC的临床分析提供了有价值的辅助，并可作为仲裁工具或用于标准化实体瘤中的Ki67评估。
本文已在Microsoft Copilot的协助下修订，以符合规定的字符数限制。
查看英文原文 English abstract
Background: Ki67 is a key proliferation marker in solid tumors, particularly relevant for HR+/HER2- breast cancer when guiding adjuvant therapy. Despite its clinical importance, Ki67 immunohistochemistry (IHC) scoring lacks standardization. International guidelines aim to reduce variability among pathologists, and AI-driven image analysis solutions have recently emerged as rapid and reliable alternatives. This study compares Ki67 scoring using Aiforia® (AI platform) and Halo® (supervised image analysis software) against three independent pathologists across a large solid tumor cohort.
Methods: We stained 192 tumors of various origins, including breast and prostate, with Ki67 (clone 30-9). Pathologists were trained per International Ki67 Working Group (IKWG) recommendations and scored tissues accordingly. Aiforia®, based on deep learning, automatically quantified Ki67-positive tumor cells within minutes. Halo® employed a random forest classifier to segment tumor, non-tumor, and background regions, verified by a pathologist. After cell segmentation, Ki67 positivity was determined by thresholding.
Results: Ki67 scoring showed strong agreement between Aiforia® and Halo® across all solid tumors (r² = 0.95). Inter-pathologist correlations were weaker (A-B: r² = 0.78; A-C: r² = 0.86; B-C: r² = 0.85) despite standardized training, though still acceptable. Among 19 tumor types analyzed, only thyroid and stomach showed software correlation inferior to 0.75, with inter-pathologist agreement fair for stomach (r² > 0.80) and lower for thyroid (r² = 0.65-0.82). Organ-specific variability was notable among pathologists, while software scores remained consistent. For breast tumors (n = 16), Aiforia® and Halo® correlated strongly (r² = 0.96), whereas pathologist agreement ranged from r² = 0.60 to 0.87.
Conclusions: AI-based platforms like Aiforia® and supervised image analysis tools such as Halo® provide robust, reproducible Ki67 scoring and significantly reduce inter-observer variability. These technologies offer valuable assistance for IHC-based clinical analysis and may serve as arbitration tools or standardize Ki67 evaluation in solid tumors.
This text has been revised with the assistance of Microsoft Copilot to comply with the specified character limit.
利益披露 Disclosure
R. Gaspo, None.
X. Pichon,
Cerba Research Other, past Cerba Research employee.
M. Tliba,
Cerba Research Other, past Cerba Research employee.
S. Iglesias,
Cerba Research Other, past Cerba Research employee.
D. Kumar,
Aiforia Employment.
R. Burrer,
Cerba Research Other, past Cerba Research employee.
A. Finan-Marchi,
Cerba Research Other, past Cerba Research employee.
M. Gérus-Durand, None.