基于深度学习检测脉管浸润在早期浸润性乳腺癌预后预测中的价值

Lymphovascular Invasion (LBVI) is a crucial step in the metastatic cascade and has a significant prognostic value in early breast cancer. Objective and effective methods are urgently needed to improve the detection rate of LBVI in Hematoxylin-eosin (HE) images and reduce the false negative rate as much as possible. Recently researches showed that there exists great potential for machine learning in pathology image analysis in the era of digital pathology. In the previous study, we used a shallow machine learning based on handcrafted features to analyze HE images of invasive breast cancer, and proposed new pathologic prognostic parameters. In recent years, deep learning based on a data-driven method has shown its advantages in pathology image processing. This project intends to explore the prognostic value of LBVI detected by deep learning in early breast cancer. First, HE sections of invasive breast cancer are digitized, and then the pathologists label LBVI in whole slide images. The multi-scale incremental growing segmentation model is constructed based on the labeled image set. Realize automatic detection of LBVI using the above mention method. Then, retrospective studies are conducted to analyze whether deep learning could reduce the false negative rate of LBVI screening. And figure out the prognostic significance of LBVI in lymph node negative (N0) invasive breast cancer. Finally, in a prospective study, analyze the correlation between LBVI and the baseline level of circulating tumor cells in N0 invasive breast cancer. Explore the value of LBVI in early invasive breast cancer treatment decisions. This study may provide a new idea for breast cancer prognoses study, and offer an alternative strategy for LBVI detection.

脉管浸润（LBVI）是转移级联中至关重要的步骤，在早期乳腺癌中预后意义重大。临床亟需客观有效的检测手段来提高苏木素-伊红（HE）图像中LBVI的检出率，尽可能降低假阴性率。病理数字化时代，机器学习在病理图像分析上表现出巨大潜力。前期研究中，我们采用基于人工设计特征的浅层机器学习自动分析乳腺癌HE组织病理图像，并提出新型病理预后参数。本课题拟探索深度学习检测LBVI在早期浸润性乳腺癌预后预测中的价值。首先，基于专家标记的图像集构建多尺度渐进式生长分割模型，实现自动检测LBVI。然后，回顾性研究分析深度学习能否降低LBVI筛查的假阴性率。明确LBVI在淋巴结阴性（N0）浸润性乳腺癌中的预后意义。最后，前瞻性研究分析N0期浸润性乳腺癌中LBVI与术后循环肿瘤细胞基线水平的相关性，探索LBVI在早期浸润性乳腺癌治疗决策中的价值。提出一个客观有效的LBVI筛查方法，为乳腺癌预后研究提供新思路。

随着肿瘤三级预防中早筛的落实，早期无淋巴结转移乳腺癌诊断率逐年升高。筛选预测乳腺癌淋巴结转移（Lymph node metastasis，LNM）或远处转移的风险因素，对乳腺癌预后评估及治疗决策如腋窝淋巴结管理具有重要意义。根据肿瘤转移级联理论，癌细胞浸润血管或淋巴管（Lymph blood vessel invasion，LBVI）、进入循环是乳腺癌转移早期的关键事件。临床亟需客观有效的检测手段来提高乳腺浸润性导管癌苏木素-伊红（Hematoxylin-eosin，HE）染色组织病理图像中LBVI的检出率。病理数字化时代，深度学习在病理图像分析上表现出巨大潜力。在本项目中，首先，我们建立了基于深度学习的自动分割并定量评估HE组织病理图像中LBVI的规范化流程：组织切片数字化、LBVI注释、数据集随机化、构建并验证基于专家经验的知识迁移学习（Expert-experience embedded knowledge transfer learning，EEKT）模型、特征提取。EEKT模型在测试集中分割LBVI的DSC值约为0.930。传统病理报告仅仅报告LBVI的有或无；而我们基于EEKT模型分割结果，提取了LBVI的计数、定位及大量形态学特征，以定量特征代替传统形态描述语言。相较于传统病理报告，LBVI定量特征在预测LNM风险中更具优势。其次，我们在一项回顾性研究中分析了乳腺微浸润癌（Microinvasive breast cancer，MBC）中LNM情况及其对预后的影响。结果显示MBC中LNM发生率为6.4%，多因素分析还显示LNM是乳腺癌相关生存（Breast cancer-specific survival，BCSS）的不良预后因素中预测权重最大的风险因素，N3期（LNM个数≥10）MBC的10年BCSS仅为49.3%。表明在超早期浸润性乳腺中即出现了肿瘤转移行为。长久以来，乳腺癌细胞到底是经淋巴管还是经血管转移至远处尚存争议。为探索淋巴管癌浸润及淋巴管内皮细胞（Lymphatic endothelial cells，LECs）在乳腺癌发展中的作用，我们分析了现有组织病理技术在标记人乳腺癌组织中淋巴管及LECs的性能，并建立了稳定标记成像方法。为探索早期浸润性乳腺癌中，淋巴管及血管癌浸润预后预测价值提供了技术支持和理论依据。