Building a foundation model of lung health in CT data
Principal Investigator
Name
Ahmed Hosny
Degrees
Ph.D
Institution
Ambient, Inc.
Position Title
CTO
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-1309
Initial CDAS Request Approval
Aug 23, 2024
Title
Building a foundation model of lung health in CT data
Summary
This project aims to develop a robust foundation model for assessing lung health using Computed Tomography (CT) data. Lung diseases, such as chronic obstructive pulmonary disease (COPD), lung cancer, and interstitial lung diseases, pose significant global health challenges. Early and accurate diagnosis is crucial for effective treatment and management. However, the complexity and variability in lung CT images make automated analysis challenging.
Our approach leverages the latest advancements in artificial intelligence (AI) and deep learning to create a scalable model that can analyze lung CT scans with high accuracy. The data we are collecting will be used to to train foundation models incorporating transformer architectures. Such models will be trained in a self-supervised manner (without labels) to learn an embedding representation of lungs.
We will then work to correlate these lung representations with overall survival, incidences of lung cancer, as well as cardiovascular disease. We will also perform saliency mapping and interpretability studies to identify lung regions or pathologies that correlate with clinical outcomes and disease prognosis.
By the end of this project, we expect to deliver a state-of-the-art AI tool that significantly improves the efficiency and accuracy of lung health assessments, contributing to better patient outcomes and advancing the field of medical imaging.
Our approach leverages the latest advancements in artificial intelligence (AI) and deep learning to create a scalable model that can analyze lung CT scans with high accuracy. The data we are collecting will be used to to train foundation models incorporating transformer architectures. Such models will be trained in a self-supervised manner (without labels) to learn an embedding representation of lungs.
We will then work to correlate these lung representations with overall survival, incidences of lung cancer, as well as cardiovascular disease. We will also perform saliency mapping and interpretability studies to identify lung regions or pathologies that correlate with clinical outcomes and disease prognosis.
By the end of this project, we expect to deliver a state-of-the-art AI tool that significantly improves the efficiency and accuracy of lung health assessments, contributing to better patient outcomes and advancing the field of medical imaging.
Aims
Aim 1. Develop a self-supervised foundation model that can accurately encode and represent the visual appearance of lungs on CT imaging.
Aim 2. Correlate lung representations with clinical outcomes.
Aim 3. Test on external validation data from public sources.
Collaborators
Ahmed Hosny - Ambient, Inc.
Justin Johnson - Ambient, Inc.