Lung Screening Using Data Synthesis and Explainable Artificial Intelligence
Principal Investigator
Name
Matthew Hamilton
Degrees
Ph.D.
Institution
Memorial University of Newfoundland
Position Title
Assistant Professor
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-830
Initial CDAS Request Approval
Sep 14, 2021
Title
Lung Screening Using Data Synthesis and Explainable Artificial Intelligence
Summary
Lung cancer is the most commonly diagnosed cancer among Canadians. It is also the leading cause of death from cancer, accounting for 1 in 4 of all cancer deaths, having among the lowest survival rate amongst cancers. If lung cancer is found at an earlier stage, the chance of successful treatment is higher. Early detection of cancer can only help improve treatment outcomes.
Screening programs geared towards early detection have been shown to reduce mortality in lung cancer by allowing for early intervention. Administering these screening programs can require significant resources. For example, large volumes of images are produced which currently must be screened by human radiologists, consuming large amounts of time. Given the large volumes of screening required, human error may occur more frequently. Moreover, image screening based on human analysis is limited to discrimination based on features visible to the human eye and does not allow for cancer signatures (known or yet to be discovered) which simply may not be visible using conventional image screening procedures.
The work proposed here aims to produce automated systems for lung screening based on artificial intelligence. These systems will allow for high-throughput screening programs which require minimal human input, thus reducing program costs. We hypothesize that artificial intelligence systems can eventually perform better than humans at detecting cancer in screening images, having less errors and learning to detect cancer signatures much earlier than human radiologists. Another possible positive outcome is that automated screening can be shown to enable lower radiation screening images with longer screening intervals, thus reducing the patient’s exposure to harmful radiation.
Our approach to improve upon existing work is to consider production of synthetic data for training systems which generalize better across the wide swath of possible cases. We also consider approaches to produce explanable predictions as current approaches are based on "blackbox" AI.
Screening programs geared towards early detection have been shown to reduce mortality in lung cancer by allowing for early intervention. Administering these screening programs can require significant resources. For example, large volumes of images are produced which currently must be screened by human radiologists, consuming large amounts of time. Given the large volumes of screening required, human error may occur more frequently. Moreover, image screening based on human analysis is limited to discrimination based on features visible to the human eye and does not allow for cancer signatures (known or yet to be discovered) which simply may not be visible using conventional image screening procedures.
The work proposed here aims to produce automated systems for lung screening based on artificial intelligence. These systems will allow for high-throughput screening programs which require minimal human input, thus reducing program costs. We hypothesize that artificial intelligence systems can eventually perform better than humans at detecting cancer in screening images, having less errors and learning to detect cancer signatures much earlier than human radiologists. Another possible positive outcome is that automated screening can be shown to enable lower radiation screening images with longer screening intervals, thus reducing the patient’s exposure to harmful radiation.
Our approach to improve upon existing work is to consider production of synthetic data for training systems which generalize better across the wide swath of possible cases. We also consider approaches to produce explanable predictions as current approaches are based on "blackbox" AI.
Aims
-Improve current automated lung screening AI-based approaches with improved accuracy
-Produce automated screening systems which provide visual expandability of the resulting diagnostic prediction
-Identify and bridge gaps between research and producing a reliable solution for clinical use
-Produce datasets of synthetic data for the larger research community use in AI/machine learning
-Improve knowledge of imaging biomarkers for cancer targeted towards earlier detection, targeting features currently invisible to the eye of radiologists in images
Collaborators
Edward Kendall, Memorial University of Newfoundland
Bassem Elshahat, Memorial University of Newfoundland