Study
NLST
(Learn more about this study)
Project ID
NLST-649
Initial CDAS Request Approval
Mar 16, 2020
Title
Lung tumor early detection and prognosis with machine learning approach
Summary
We focus on automatic segmentation of large-scale medical images for early tumor detection by extracting keyimaging features of tumors at an early stage. Early tumor detection is important in developing treatment plans which may increase chances of patient survival. It, when performed by medical experts, is time consuming and inaccurate since subtle image change may not be visually discernible by a human expert but may allow for machine learning based technique. To identify key features and automate the detection procedure is challenging due to large variation of tumor volume, location and characteristics across different individuals and a small signal-noise ratio. After the tumor location is identified, the tumor progression will be further monitored, possibilities about whether a tumor is benign or malignant will be predicted, and the cancer stage will be classified, based on which, the best possible clinical treatment could be determined.
Aims
Mainly we have two aims here:1. The first one is the early tumor location detection. Before we extract the medical image feature, a uniform coordinate system will be established using image registration technique. We next will apply different voxel-based machine learning techniques to segment the normal tissue, necrosis, active tumor and surrounding edema area.2. The second one is the tumor prognosis. We monitor the tumor progression using longitudinal dataset and find the key feature that determines the tumor metastasis. Various statistical and computational methods including deep neural Networks, XGBoost and random forests will be investigated. A reliable prediction can enhance the targeted treatment which uses drugs or other substances to more precisely identify and attack cancer cells, while doing little damage to normal cells, thus reducing the side effect of therapy as well as healthcare costs.
Collaborators
Hongtu Zhu, University of North Carolina at Chapel Hill
