Automated lung cancer detection and characterization via artificial intelligence
Principal Investigator
Name
Jianing Pang
Degrees
PhD
Institution
VoxelCloud, inc.
Position Title
Research Scientist
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-186
Initial CDAS Request Approval
Dec 22, 2015
Title
Automated lung cancer detection and characterization via artificial intelligence
Summary
Lung cancer is the most deadly type of cancer both in the US and worldwide. It is estimated that lung cancer will cause 158,040 deaths in the US in 2015, which accounts for nearly 40% of all cancer deaths in the country (NCI, 2015). Worldwide, it caused 1.69 million deaths in 2012, nearly 20% of the total cancer deaths in the world (WHO, 2014).
Early detection of lung cancer is crucial for improving patient outcome, since early stage lung cancer has much better prognosis than those found at later stages. In recent years, low-dose computer tomography (CT) has become a powerful tool for non-invasive lung cancer detection and characterization. However, locating, measuring, and interpreting lung nodules found in the CT images remain a labor-intensive task and requires highly trained readers. Since the 80s, researchers have proposed to use computer programs to aid clinicians in accomplishing such task (Giger, 1988). However, most previous attempts suffer from low diagnostic performance and limited functionality, which limited their clinical applicability.
In this proposal, we aim to utilize the latest advancements in artificial intelligence, specifically deep learning techniques, to leverage the large amount of annotated NLST data (clinical history, physical exam, laboratory data, and imaging exams, etc.) and develop an automated, predictive model for localizing, measuring, and characterizing lung nodules (e.g. Shin 2015). We believe such an automated system will be invaluable in alleviating the radiologists’ workload and providing crucial clinical decision support.
Early detection of lung cancer is crucial for improving patient outcome, since early stage lung cancer has much better prognosis than those found at later stages. In recent years, low-dose computer tomography (CT) has become a powerful tool for non-invasive lung cancer detection and characterization. However, locating, measuring, and interpreting lung nodules found in the CT images remain a labor-intensive task and requires highly trained readers. Since the 80s, researchers have proposed to use computer programs to aid clinicians in accomplishing such task (Giger, 1988). However, most previous attempts suffer from low diagnostic performance and limited functionality, which limited their clinical applicability.
In this proposal, we aim to utilize the latest advancements in artificial intelligence, specifically deep learning techniques, to leverage the large amount of annotated NLST data (clinical history, physical exam, laboratory data, and imaging exams, etc.) and develop an automated, predictive model for localizing, measuring, and characterizing lung nodules (e.g. Shin 2015). We believe such an automated system will be invaluable in alleviating the radiologists’ workload and providing crucial clinical decision support.
Aims
Aim #1: To develop a deep learning-based system for automated lung nodule detection, feature extraction, and characterization, with available expert diagnosis as the ground truth.
Aim #2: After accomplishing aim #1, we aim to further develop the algorithms that predict the patient prognosis leveraging the available outcome data.
Collaborators
Demetri Terzopoulos, PhD, UCLA
Fereidoun Abtin, MD, UCLA