PO.BCS01.07 · 生物信息与计算

比较数字化与人工智能（AI）计算算法在量化临床样本中转移性乳腺癌（mBC）低表达/超低表达人表皮生长因子受体2（HER2）蛋白方面的应用

Comparison of digital and artificial intelligence (AI)-computational algorithms for quantifying low/ultralow human epidermal growth factor receptor 2 (HER2) protein expression in metastatic breast cancer (mBC) from clinical samples

编号 1455 展板 18 时间 4/20 09:00–12:00 区域 Section 4 主讲 Savitri Krishnamurthy, MD

分会场 Digital Pathology 2

查看 PDF 下载 PDF 🔒 查看 / 下载完整 PDF 需登录并开通下载套餐 · 查看套餐 / 开通 AACR 官方页面

作者与单位 Authors & Affiliations

Savitri Krishnamurthy¹, Dhanrajan Tiruchinapalli², Clara lam², Simon M. Collin³, Rosemary Taylor⁴, Linlin Luo⁵, Anupriya Dutta⁵, Ehab A. Elgabry⁶, Michele S. Woo⁷, Grace E. Kwon⁸, Robert Egger⁹, Jennifer A. Hipp¹⁰, Lauren Brunner⁹, Jeppe S. Thagaard¹¹, Thomas W. Ramsing¹², Henrik Høeg¹³, Wonkyung Jung¹⁴, Heon Song¹⁴, Chang Ho Ahn¹⁵, Vladimir Kravtsov¹⁶, Patrick Frey¹⁷, Ralf Banisch¹⁷, Stella Redpath²

¹Department of Pathology, Division of Pathology and Laboratory Medicine, The University of Texas MD Anderson Cancer Center, Houston, TX,²US Medical Affairs, Oncology, AstraZeneca Pharmaceuticals LP, Gaithersburg, MD,³Global Medical Affairs, AstraZeneca, Cambridge, United Kingdom,⁴Oncology Biometrics, Oncology R&D, AstraZeneca, Macclesfield, United Kingdom,⁵Global Medical Affairs, AstraZeneca Pharmaceuticals LP, Gaithersburg, MD,⁶Medical Affairs, Daiichi Sankyo Inc., Basking Ridge, NJ,⁷US Medical Affairs, Daiichi Sankyo Inc., Basking Ridge, NJ,⁸Daiichi Sankyo Inc., Basking Ridge, NJ,⁹PathAI, Inc., Boston, MA,¹⁰Pathology, PathAI, Inc., Boston, MA,¹¹Visiopharm, Hørsholm, Denmark,¹²Visiopharm, Copenhagen, Denmark,¹³R&D, Visiopharm, Hørsholm, Denmark,¹⁴Lunit, Seoul, Korea, Republic of,¹⁵Medicine Department, Lunit, Seoul, Korea, Republic of,¹⁶Machine Learning, Mindpeak GmbH, Hamburg, Germany,¹⁷Mindpeak GmbH, Hamburg, Germany

摘要 Abstract

中文摘要

背景基于DESTINY-Breast04（HER2-low）和-06（HER2-low/-ultralow）试验，T-DXd已被批准用于HER2-low（免疫组化[IHC] 1+或IHC 2+/原位杂交阴性）或HER2-ultralow（IHC 0且膜染色见于≤10%的肿瘤细胞）的mBC。将来自mBC活检样本、HER2 IHC评分为0/1+的全切片图像（WSI）由病理学家和使用数字病理学（DP）重新评分，以评估一致性。方法这项回顾性真实世界证据研究纳入了2020-2023年收集的384张WSI，均使用PATHWAY HER2（4B5）检测染色，评分为HER2 IHC 0（n = 246）或1+（n = 138）。三名病理学家各自使用2023年ASCO/CAP指南对每张WSI进行2次盲法阅读；若阅读结果不同，则采用协调后的评分。共识定义为3名病理学家中≥2名达成一致。同样的WSI使用4种在研的AI计算DP工具进行分析。通过总体百分比一致性（OPA）和Cohen κ测量与人工共识的一致性，并记录审查时间。结果在384张WSI中，375张经病理学家审查后HER2 IHC评分一致；9张存在不一致。在共识病例中，2/3一致出现在154张WSI（41.1%），3/3一致出现在221张（58.9%）；81张（21.6%）WSI被重新归类为IHC 0无膜染色，85张（22.7%）为IHC 0有膜染色，203张（51.4%）为IHC 1+，6张（1.6%）为IHC 2+。4种DP工具的HER2 IHC重新评分结果见表1。跨所有HER2 IHC评分类别，共识与DP辅助评分之间的OPA（95% CI）分别为74%（69-78%）、73%（68-77%）、69%（64-74%）和55%（50-60%）。Cohen κ（95% CI）范围为0.33-0.59。DP的中位审查时间比人工审查更短。结论初步分析表明，将AI计算DP工具整合到HER2 IHC临床工作流程中可能减少病理学家的审查时间。进一步的分析正在进行中，以评估DP工具与人工评分的一致性。表1. WSI的HER2 IHC重新评分（包括HER2-low/ultralow）及HER2 IHC中位审查时间 HER2 IHC 0无膜染色 n (%)；HER2 IHC 0有膜染色 n (%)；HER2 IHC 1+ n (%)；HER2 IHC 2+ n (%)；HER2 IHC 3+ n (%)；缺失结果 a n (%)；人工共识与DP工具间的OPA评分, % (95% CI)；Cohen κ (95% CI)；审查时间, 中位数（范围）, 分钟。人工共识 N = 375：81 (21.6)、85 (22.7)、203 (54.1)、6 (1.6)、0、0、不适用、不适用、6.7 (3.0-11.5)。 DP工具 RV73X N = 375：72 (19.2)、100 (26.7)、176 (46.9)、5 (1.3)、0、22 (5.9)、74 (69-78)、0.59 (0.52-0.66)、2.1 (0.3-50.2)。 DP工具 MQ52G N = 375：77 (20.5)、126 (33.6)、168 (44.8)、4 (1.1)、0、0、73 (68-77)、0.57 (0.51-0.64)、0.7 (0.1-8.5)。 DP工具 KL84Q N = 375：82 (21.9)、137 (36.5)、145 (38.7)、10 (2.7)、1 (0.3)、0、69 (64-74)、0.53 (0.46-0.60)、2.4 (0.5-25.8)。 DP工具 ZX19P N = 375：27 (7.2)、201 (53.6)、130 (34.7)、6 (1.6)、3 (0.8)、8 (2.1)、55 (50-60)、0.33 (0.27-0.39)、不可用。 a 完整数据集：375张WSI；缺失结果 = 无DP工具输出的WSI。

查看英文原文 English abstract

Background Based on DESTINY-Breast04 (HER2-low) and -06 (HER2-low/-ultralow) trials, T-DXd is approved for HER2-low (immunohistochemistry [IHC] 1+ or IHC 2+/in situ hybridization negative) or -ultralow (IHC 0 with membrane staining in ≤10% of tumor cells) mBC. Whole slide images (WSIs) from mBC biopsy samples scored HER2 IHC 0/1+ were rescored by pathologists and using digital pathology (DP) to evaluate concordance. Methods This retrospective real-world evidence study included 384 WSIs collected 2020-2023, stained with PATHWAY HER2 (4B5) assay scored as HER2 IHC 0 (n = 246) or 1+ (n = 138). Three pathologists each performed 2 blinded readings per WSI using 2023 ASCO/CAP guidelines; if readings differed, a reconciled score was used. Consensus was agreement by ≥2 of 3 pathologists. The same WSIs were analysed with 4 AI-computational DP tools in development. Concordance vs manual consensus was measured by overall percentage agreement (OPA) and Cohen κ, with review time recorded. Results Of 384 WSIs, 375 had aligned HER2 IHC scores by pathologist review; 9 had discordance. Among consensus cases, 2/3 agreement occurred in 154 WSIs (41.1%) and 3/3 in 221 (58.9%); 81 (21.6%) WSIs were reclassified as IHC 0 absent membrane staining, 85 (22.7%) as IHC 0 with membrane staining, 203 (51.4%) as IHC 1+, and 6 (1.6%) as IHC 2+. HER2 IHC rescoring results with the 4 DP tools are shown in Table 1. OPA (95% CI) between consensus and DP-assisted scores across all HER2 IHC score categories was 74% (69-78%), 73% (68-77%), 69% (64-74%), and 55% (50-60%). Cohen κ (95% CI) ranged from 0.33-0.59. Median review times were shorter with DP vs manual review. Conclusion Preliminary analysis suggests integrating AI-computational DP tools into HER2 IHC clinical workflows may reduce pathologist review time. Further analysis is underway to assess concordance of DP tools with manual scoring. Table 1. HER2 IHC Re-scores (including HER2-low/ultralow) of WSIs and Median Time to Review HER2 IHC 0 absent membrane staining n (%) HER2 IHC 0 with membrane staining n (%) HER2 IHC 1+ n (%) HER2 IHC 2+ n (%) HER2 IHC 3+ n (%) Missing results a n (%) OPA score between manual consensus and DP Tool, % (95% CI) Cohen κ (95% CI) Review time, median (range), minutes Manual consensus N = 375 81 (21.6) 85 (22.7) 203 (54.1) 6 (1.6) 0 0 Not applicable Not applicable 6.7 (3.0-11.5) DP tool RV73X N = 375 72 (19.2) 100 (26.7) 176 (46.9) 5 (1.3) 0 22 (5.9) 74 (69-78) 0.59 (0.52-0.66) 2.1 (0.3-50.2) DP tool MQ52G N = 375 77 (20.5) 126 (33.6) 168 (44.8) 4 (1.1) 0 0 73 (68-77) 0.57 (0.51-0.64) 0.7 (0.1-8.5) DP tool KL84Q N = 375 82 (21.9) 137 (36.5) 145 (38.7) 10 (2.7) 1 (0.3) 0 69 (64-74) 0.53 (0.46-0.60) 2.4 (0.5-25.8) DP tool ZX19P N = 375 27 (7.2) 201 (53.6) 130 (34.7) 6 (1.6) 3 (0.8) 8 (2.1) 55 (50-60) 0.33 (0.27-0.39) Not available a’ Full dataset: 375 WSIs; Missing results = WSIs without a DP tool output

利益披露 Disclosure

S. Krishnamurthy, None. D. Tiruchinapalli, AstraZeneca Pharmaceuticals LP Employment. C. lam, AstraZeneca Employment, Stock Option. S. M. Collin, AstraZeneca Pharmaceuticals Ltd Employment, Stock Option. R. Taylor, AstraZeneca Pharmaceuticals Ltd. Employment, Stock Option. L. Luo, AstraZeneca Pharmaceuticals Ltd. Employment, Stock Option. A. Dutta, AstraZeneca Pharmaceuticals LP Independent Contractor. E. A. Elgabry, Daiichi Sankyo Employment. M. S. Woo, Daiichi Sankyo Inc. Employment, Stock, Travel, Other, Owner of Daiichi Sankyo Restricted Stock Units (RSU) and Stock Appreciation Rights (SAR), Business travel supported by company as part of employment. G. E. Kwon, Daiichi Sankyo, Inc. Employment. R. Egger, PathAI, Inc. Employment. J. A. Hipp, PathAI, Inc. Employment, Stock. L. Brunner, PathAI, Inc. Employment. J. S. Thagaard, Visiopharm Employment. T. W. Ramsing, VisioPharm Employment. H. Høeg, Visiopharm Employment. W. Jung, Lunit Employment. H. Song, Lunit Employment. C. Ahn, Lunit Employment, Stock, Stock Option. V. Kravtsov, Mindpeak GmbH Employment. P. Frey, Mindpeak GmbH Employment, Stock Option. R. Banisch, Mindpeak GmbH Employment. S. Redpath, AstraZeneca Pharmaceuticals Ltd Employment, Stock Option.

← 返回 AACR 2026 检索