PO.PR01.03 · 预防研究
利用人工智能从组织病理学图像中高保真地检测和分类卵巢癌
High-fidelity detection and classification of ovarian cancer from histopathological images using Artificial Intelligence
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
引言：卵巢癌是一种致命的妇科恶性肿瘤，带来了沉重的经济负担。准确的组织病理学亚型分类对于精准治疗和预后判断至关重要。传统的病理诊断依赖于主观解读，容易受到显著的观察者间变异性的影响。利用残差架构的卷积神经网络能够从组织病理学图像中实现自动化、客观的特征提取。本研究探讨基于ResNet18的卵巢癌亚型分类，以提高诊断准确性并支持循证临床决策。
方法：获取了来自不列颠哥伦比亚大学卵巢癌亚型分类数据集的匿名组织病理学图像，涵盖五种卵巢癌亚型（n=513）。该数据集被随机按比例划分为训练集（60%）、验证集（20%）和测试集（20%）。图像在使用ResNet18架构进行训练之前，经过了全面的预处理和增强。诊断性能通过在验证集和测试集上的准确率、精确率-召回率、F1和F2评分以及受试者工作特征曲线下面积（AUROC）进行评估。经过训练的模型被部署到一个普遍可访问的跨平台应用程序中，以供全球专家验证。
结果：
ResNet18在各卵巢癌亚型中展现出稳健的判别能力，训练准确率达到>98%，验证准确率达到>82%。精确率-召回率分析显示出优异的性能，尤其是对于子宫内膜样亚型（0.95）。这些结果表明ResNet18具有增强组织病理学工作流程的潜力，有助于实现公平的卵巢癌亚型分层。
结论：基于ResNet18的分析提供了自动化、准确的卵巢癌诊断。在各组织学亚型中的高验证准确率和强AUROC评分表明其在标准化诊断解读方面的有效性。将其整合到临床工作流程中可以减少诊断变异性并支持精准指导的治疗。在前瞻性多中心队列中的进一步验证将把它转化为一种有效的诊断工具。
查看英文原文 English abstract
Introduction: Ovarian cancer is a lethal gynecological malignancy with a substantial economic burden. Accurate histopathological subtype classification is critical for precision therapy and prognostication. Traditional pathological diagnosis relies on subjective interpretation, which is vulnerable to significant interobserver variability. Convolutional neural networks leveraging residual architectures enable automated, objective feature extraction from histopathological images. This study investigates ResNet18-based ovarian cancer subtype classification to accelerate diagnostic accuracy and support evidence-based clinical decision-making.
Methods: Anonymized histopathological images from the University of British Columbia Ovarian Cancer Subtype Classification dataset encompassing five ovarian cancer subtypes (n=513) were procured. The dataset was randomly proportioned into training (60%), validation (20%), and testing (20%) cohorts. Images underwent comprehensive preprocessing and augmentation prior to training using the ResNet18 architecture. The diagnostic performance was assessed using accuracy, precision-recall, F1 and F2-scores, and area under the receiver operating characteristic curve (AUROC) on both validation and test sets. The trained model was deployed in a universally accessible, cross-platform application for expert validation globally.
Results:
ResNet18 demonstrated robust discriminative capability across ovarian cancer subtypes, achieving >98% training accuracy and >82% validation accuracy. Precision-recall analysis showed excellent performance, especially for the endometrioid subtype (0.95). These results indicate ResNet18's potential to augment histopathology workflows, facilitating equitable ovarian cancer subtype stratification.
Conclusion: ResNet18-based analysis provides automated, accurate ovarian cancer diagnosis. High validation accuracy and strong AUROC scores across histological subtypes indicate efficacy in standardizing diagnostic interpretation. Integration into clinical workflows could reduce diagnostic variability and support precision-guided therapy. Further validation in prospective multicenter cohorts will translate this into an effective diagnostic tool.
利益披露 Disclosure
E. Krishnan, None..
S. Ahmed, None..
T. Wasim, None.