基于放射-病理组学的乳腺癌转移风险预测模型研究

Metastasis is the process of cancerous cells leaving the primary tumor to spread out to continue to grow. It is an important cancer staging and prognostic factors for breast cancer. High heterogeneity of metastatic primary tumor morphology is manifested in different scales and multimodal data. This proposal intends to integrate the sub-visual features of radiomics and pathomics of breast cancer. By analyzing the interaction and correlation of between two modal data in different metastatic states, a quantitative risk prediction model of the metastasis on breast cancer is built. Here is the procedure to build the model: for primary tumor 1) the four-dimensional segmentation model on Dynamic Contrast Enhancement Magnetic Resonance Imaging (MRI) is developed to segment the tumor area; 2) The area of the cancerous region, epithelial and stromal tissues, and cancerous nuclei are subsequently segmented. Based on segmentation results, 1) two- and three- dimensional morphological, textural, and pharmacokinetic features are extracted; 2) histomorphologic description of pathomics features are developed. Based on the interaction and correlation of between the heterogeneous data of two modes, the deep canonical correlation analysis is leveraged to representation learning on the features. Then the represented features are integrated with a random forest classifier for constructing the metastatic risk predictor of breast cancer. Finally, prospective studies are conducted clinically to validate the effectiveness of the metastatic risk prediction model in order to accurately predict the risk of metastatic breast cancer patients.

转移是指癌细胞离开原发灶向外扩散继续生长的过程，是重要的癌症分期和预后因子。转移风险高的原发肿瘤形态的高度异质性体现在不同尺度、多模态的数据中。本项目拟融合乳腺癌放射和病理组学的亚视觉特征，通过回顾性分析不同转移状态下两种模态数据的相互作用和关联，构建定量的转移风险预测模型。该模型构建流程是：针对乳腺原发灶1)构建动态增强核磁共振图像自动的四维分割模型，分割肿瘤区域；2)分割全扫描病理图像的癌变区域，上皮和间质组织及癌细胞。根据分割结果在1)核磁共振图像肿瘤区域提取二维和三维形态和纹理特征，以及药代动力学等放射组学特征；2)病理图像的癌变区域内提取基于组织形态学描述的病理组学特征。根据两种模态异构数据的相互作用和关联，运用深度典型相关自编码器对它们的组学特征重新表达，结合随机森林分类器，构建模型。最后，在临床上开展前瞻性研究，验证转移风险预测模型的有效性，以准确地预测乳腺癌患者的转移风险。

据最新癌症统计报告，乳腺癌已经跃居女性恶性肿瘤发病率的第一位，这给我国医疗卫生事业发展带来了沉重的负担。如何为乳腺癌患者制定最佳的治疗方案、改善患者的预后，是重要的临床问题。其中乳腺癌的转移风险是评估乳腺癌患者预后好坏的重要指标。影像和病理图像分别从不同的角度反映患者疾病的信息，如何综合运用影像和病理构建多模态融合模型评估乳腺癌患者的风险，是非常有价值的科学问题。本项目以乳腺癌影像和病理图像为研究对象，研究了新颖的深度学习方法，探索了基于像素级深度计算的图像分析方法。综合乳腺癌影像和病理数据：1)构建了有效的基于像素层次的图像分析方法，研究了影像和病理数据构建有效的肿瘤区域、癌变区域的自动识别和分割的方法，2）影像和病理数据中找到有价值的基于组学的标志物，能够预测患者的预后和转移的风险；3)通过定量地分析影像和病理图像，定量预测乳腺癌患者转移的风险。