基于半自动堵漏分割和量化超声特征构建三阴性乳腺癌早期诊断模型的研究

Triple negative breast cancer (TNBC) is characterized with aggressive degrees of malignancy and poor prognosis. Thus early diagnosis of TNBC is of great importance for treatment and prognosis. However, it often lacks typical imaging features and easily shows some features of benign lesion which is challenging for ultrasound diagnosis. This project is aimed to improve the early diagnosis of TNBC by using a machine learning method with quantitative ultrasound features. First we propose to use a novel multi-scale leak-plugging semi-automated segmentation method to solve the impact of pixel leakage at tumor margin on image segmentation quality. Furthermore, optimized computerized feature algorithms will be used for quantitative extraction and analysis of sonographic features of gray-scale and color Doppler ultrasound images. An effective feature set was screened out from many complex indicators such as margin, shape, texture and blood flow by feature selection, which provide objective and quantitative indicators for breast ultrasound diagnosis. Using appropriate machine learning methods, we will construct related prediction models of TNBC from various types of breast lesions to improve the ultrasound diagnosis of TNBC and provide some scientific basis for exploring the clinical application of a computer-aided diagnostic system.

三阴性乳腺癌（Triple negative breast cancer, TNBC）恶性程度高、预后差，早期诊断对治疗、预后意义重大。而TNBC往往缺乏典型的影像学特征，且易表现出一些良性肿瘤特征，成为超声诊断的难点问题。本项目旨在基于量化超声特征运用机器学习方法提高TNBC的早期诊断。拟提出一种多尺度半自动堵漏分割法，解决边缘像素泄漏对图像分割的影响。优化特征算法，定量提取灰阶超声及彩色多普勒超声图像特征。通过特征选择从边缘、形状、纹理、血流等众多复杂指标中筛选出有效的特征集，为乳腺超声诊断提供客观的量化指标。应用适当的机器学习方法构建TNBC与不同类型乳腺病变之间的相关预测模型，有效提高TNBC的超声诊断，为探索计算机辅助诊断系统的临床应用提供一定的科学依据。

三阴性乳腺癌（TNBC）恶性程度高、预后差，早期诊断对治疗、预后意义重大。而TNBC往往缺乏典型的影像学特征，成为超声诊断的难点问题。本项目基本按原计划顺利开展，原计划拟完成的四个研究内容均融入到每个研究方向中，首先我们对乳腺超声图像数据库进行了扩充和完善，制定了数据采集标准，扩充了多模态乳腺超声影像，为研究提供了丰富的病例基础；由于三阴性乳腺癌常表现出一些良性影像学特征，易误诊为常见的乳腺纤维腺瘤，我们基于量化超声特征（边缘、形态以及纹理特征）构建了TNBC与良性纤维腺瘤的诊断模型，提高了乳腺超声对高侵袭性TNBC的鉴别，有利于高危患者尤其是年轻女性患者的临床诊疗；此外，基于TNBC具有特殊的低血供特征，我们还应用区域性定量彩色超声特征鉴别三阴性（TN）与非三阴性（NTN）亚型，证实了区域性血流特征在两种亚型之间具有显著性差异，从而有效辅助了乳腺灰阶超声诊断；在原计划基础上，本项目还针对不同亚型乳腺癌多模态超声特征、多模态超声构建新辅助化疗后腋窝淋巴结病理状态的预测模型以及构建Ki-67表达分级预测模型进行了相关研究。并且，基于深度学习模型识别乳腺癌腋窝非前哨淋巴结受累风险、构建乳腺超声图像的诺莫图模型预测腋窝淋巴结状态以及进一步构建前哨淋巴结和非前哨淋巴结预测模型等内容进行了相关研究。本项目为乳腺癌尤其是TNBC相关问题的超声诊断研究奠定了一定的基础。