Multimodal computational models to stratify ovarian cancer patients

项目摘要/摘要 高级别浆液性卵巢癌(HGSC)是最致命的妇科恶性肿瘤，生存期为5年。 转移性疾病的发生率不到30%。我们的实验室已经确定突变过程是生存的预测因素 和对治疗的反应，以及一个预测同源重组缺陷的工作模型 苏木精-伊红(H&E)全幻灯片图像。我们诊断放射学的合作者们发现 BRCA突变状态与计算机增强对比剂质量特征之间的强健关联 体层摄影(CE-CT)。然而，这两种成像方式尚未与基因组学相结合。 改善HGSC患者分层的信息。 基于这些初步数据，我将检验CE-CT中结合中观信息和 H&E中的显微信息可用于推断已知的突变亚型，也可用于识别新患者 层级。我已经管理了118名HGSC患者的队列，他们都是基于匹配的靶向小组基因组测序， 扫描H&E全幻灯片图像，并为此对处理前的CE-CT图像进行分割。具体而言 目标1，我将开发一个机器学习模型来集成CE-CT和H&E成像来预测突变 这些无处不在的成像模式的亚型。在具体目标2中，我将开发端到端的深度学习 整合CE-CT、H&E和基因组测序的互补信息以求生存的模型 使用COX比例风险模型进行分析。我预计这项工作将(1)确定精细的分层 使用这种多模式预后信号的HGSC患者和(2)开发一台通用机器 整合CE-CT、H&E和基因组测序用于癌症患者生存分析的学习模型。 这项研究将在索赫拉布博士的指导下在纪念斯隆·凯特琳癌症中心进行 沙阿。沙阿医生和我制定的培训计划将为我将来成为一名医生做好准备- 对癌症患者预后进行机器学习研究的科学家。

Project Summary/Abstract High-grade serous ovarian cancer (HGSC) is the most lethal gynecologic malignancy, with a five-year survival rate of less than 30% for metastatic disease. Our lab has identified mutational processes as predictors of survival and response to therapy, along with a working model to predict homologous recombination deficiency from hematoxylin and eosin (H&E) whole-slide images. Our collaborators in diagnostic radiology have discovered robust associations between BRCA mutational status and qualitative features on contrast-enhanced computed tomography (CE-CT). These two imaging modalities, however, have yet to be combined with genomic information to improve stratification of HGSC patients. Based on these preliminary data, I will test the hypothesis that combined mesoscopic information in CE-CT and microscopic information in H&E can be used to infer known mutational subtypes and also to identify novel patient strata. I have curated a cohort of 118 HGSC patients with matched targeted panel-based genome sequencing, scanned H&E whole-slide images, and segmented pre-treatment CE-CT images for this purpose. In Specific Aim 1, I will develop a machine learning model to integrate CE-CT and H&E imaging to predict mutational subtype from these ubiquitous imaging modalities. In Specific Aim 2, I will develop an end-to-end deep learning model to integrate the complementary information from CE-CT, H&E, and genome sequencing for survival analysis using a Cox Proportional Hazards model. I anticipate that this work will (1) identify refined stratification of HGSC patients using this multimodal prognostic signature and (2) develop a general-purpose machine learning model to integrate CE-CT, H&E, and genomic sequencing for cancer patient survival analysis. This research will be conducted at Memorial Sloan Kettering Cancer Center under the mentorship of Dr. Sohrab Shah. The training plan that Dr. Shah and I have developed will prepare me well for a future as a physician- scientist conducting machine learning research for cancer patient prognosis.