Computerized platform for interactive annotation and topological characterization of tumor associated vasculature for predicting response to immunotherapy in lung cancer

点击翻译按钮获取中文摘要

SUMMARY: The tumor microenvironment (TME) vascular network harbors a compelling amount of anatomical and physiological information embedded on the imaging scale. Although techniques like Radiomics have shown significant promise in several medical imaging applications, such approaches are limited to capturing properties such as lesion morphology and texture, and cannot comprehensively characterize or visualize the properties of the aberrant TME vasculature. We hypothesize that angiogenesis manifests as characteristic topological and geometrical patterns of vasculature in the nodule periphery, and is associated with disease progression and outcome. In this project, we propose to leverage these topological and geometrical constructs in building adaptive segmentation, quantification, and visualization tools for tumor associated vasculature. To demonstrate the clinical efficacy of these new tools in therapy response assessment, we propose to target unmet clinical needs in response prediction of lung immunotherapy. Fewer than 20% non-small cell lung cancer (NSCLC) patients treated with immune checkpoint inhibitors (ICIs) respond favorably. Additionally, the associated costs are extremely high. Molecular markers and metrics evaluating changes in tumor size have not been very effective in predicting and monitoring response to ICIs. Intra- and peritumoral radiomic features have been recently shown to outperform traditional biomarkers in outcome prediction. None of the existing markers, however, consider the tumor associated vasculature in the clinical assessment of TME despite strong evidence of its role in determining disease progression and response to therapy. One critical obstacle is the lack of an efficient and easy-to-use 3- dimensional (D) vasculature annotation tool for clinicians. Despite rich literature, it is difficult to train an automatic segmentation model due of the highly heterogeneous and complex 3D morphology of vasculature. This is especially challenging near nodule periphery, where the pathological vasculature exhibits abnormal yet clinically relevant geometry and topology. We aim to 1) build a human-in-the-loop vasculature visualization and segmentation framework based on topological active learning, 2) characterize the topology and geometry of the extracted vessels to obtain a set of novel vascular radiomic markers, and 3) use the developed suite of quantitative vascular biomarkers to establish a risk scoring system for predicting clinical benefit for NSCLC patients undergoing ICI therapy. Specifically, these tools will be optimized to identify patients who will benefit from ICIs on pre-treatment CT. A major strength of our work is to provide clinicians an intuitive informatics platform to visualize topological and geometrical attributes of aberrant vasculature, thereby enabling them to better understand the role of vessel architecture in disease progression from a phenotypic perspective. The team will train these biologically interpretable radiomic tools using a learning set of N=120 NSCLC patients treated with ICI therapy at Stony Brook University Hospital. The developed tools will then be validated on a cohort of N=300 patients, treated at University Hospitals Cleveland Medical Center.