Comprehensive Lung Cancer Assessment via Deep Learning

Principal Investigator
Name: Qian Li
Degrees: Ph.D.
Institution: VoxelCloud, inc.
Position Title: Research Scientist
Email: qianl90@gmail.com

Project Information
Study: NLST
Project ID: NLST-386
Title: Comprehensive Lung Cancer Assessment via Deep Learning

Summary

Lung cancer is the leading cause among all cancer-related death both in the US and worldwide. The National Lung Screening Trial (NLST) has demonstrated that the lung cancer related death can be reduced by 20% via low-dose CT (LDCT) screening, therefore, it is recommended by both the US Preventive Services Task Force and the US center for Medicare and Medicaid Services. As LDCT becomes a standard for lung cancer screening in practice, millions of chest CT images need to be examined by professionals on a daily basis. The volume and necessity of the task highlight the need for computer-aided diagnosis (CAD) systems to assist radiologists comprehensively. Recently, much effort has been made on lung nodule detection, malignancy diagnosis, characteristics assessment and more, driven by the blooming success of deep learning in computer vision. However, such specialized CAD systems target one angle at a time and result in very limited usability in practice. In this proposal, we aim to develop a comprehensive lung cancer assessment system, a complete toolkit for radiologists, that achieves lung nodule detection, localization, segmentation and characterization per single scan, and nodule progression tracking over multiple scans. With the extensive collection of CT images and clinical datasets from NLST, we believe the recent advancements in artificial intelligence, especially deep learning techniques, will significantly contribute to both the lung cancer research and clinical practices. And such a system will assist radiologist with clinical decision making, in a comprehensive and truly intelligent way.

Aims

Aim #1: To developed a comprehensive and well integrated system that achieves lung nodule detection, localization, segmentation and characterization simultaneously.
Aim #2: To automatically track and quantify nodule progression from multiple CT scans.
Aim #3: To extend the capability of the proposed model to detect non-nodule related abnormalities, based on available expert diagnosis as ground truth.

Collaborators

Demetri Terzopoulos, Ph.D., UCLA
Nima Tajbakhsh, Ph.D., Research Scientist, VoxelCloud, Inc.