Deep Learning to Improve Screening of Interstitial Lung Diseases

Principal Investigator

Name
Asif Abdul Hameed

Degrees
M.D.

Institution
Henry Ford Health

Position Title
Senior Researcher

Email
aabdul1@hfhs.org

About this CDAS Project

Study
NLST (Learn more about this study)

Project ID
NLST-1505

Initial CDAS Request Approval
Apr 28, 2026

Title
Deep Learning to Improve Screening of Interstitial Lung Diseases

Summary
Interstitial lung diseases (ILDs) are characterized by chronic inflammation and fibrosis that progressively impair respiratory function and lead to premature death, with a median survival of 2.5–3.5 years. Diagnostic delay remains a major barrier to improving outcomes, increasing mortality risk by up to 33%. Interstitial lung abnormalities (ILAs)—incidental CT findings suggestive of early ILD—offer a crucial opportunity for earlier diagnosis and intervention. However, ILAs remain underrecognized and underreported on all-cause and lung cancer screening (LCS) CTs, resulting in missed opportunities for early management. Given that LCS participants represent a high-risk population for ILD development, this cohort provides an ideal setting for developing and validating artificial intelligence (AI)-based approaches for early disease detection.
This proposal aims to validate a deep learning (DL)-based framework for automated detection of ILAs and compare its diagnostic performance to expert thoracic radiologists. Leveraging 4DMedical’s advanced DL-based imaging tools—Lung Texture Analysis (LTA™) for texture quantification and IQ-UIP™ for fibrosis-pattern classification—this work will optimize and validate model performance for the novel application of ILA detection within real-world LCS datasets.
Specific Aim 1 will develop a radiologist-validated ground-truth dataset of ILAs and optimize DL-based fibrosis thresholds for accurate ILA detection using annotated scans from the National Lung Screening Trial (NLST).
Specific Aim 2 will validate and compare the optimized DL model’s diagnostic accuracy with expert radiologists using an independent Henry Ford Health (HFH) LCS cohort. Diagnostic metrics including sensitivity, specificity, precision, recall, and area under the receiver operating characteristic curve (AUC) will be calculated, with κ and subgroup analyses to assess reproducibility and mitigate bias.

Aims

• Aim 1: Develop a radiologist-validated ground truth dataset of ILAs and optimize a DL-based model for accurate ILA detection.
CT scans from the National Lung Screening Trial (NLST) will be annotated independently by two thoracic radiologists blinded to clinical data. Information from LTA™ and IQ-UIP™ outputs will be mapped to reference annotations to optimize fibrosis-extent thresholds for best diagnostic performance. All ILA determinations will follow ATS/Fleischner definitions.
• Aim 2: Validate and compare the DL-based model’s diagnostic performance with radiologist interpretation in a real-world LCS cohort.
The optimized model will be applied to CTs from Henry Ford Health (HFH) LCS participants. Diagnostic metrics—sensitivity, specificity, precision, recall, and AUC—will quantify model performance. All ILA determinations will follow ATS/Fleischner definitions to ensure standardization and reproducibility. Approximately 1,100 NLST and 900 HFH scans will be curated for training and testing to achieve performance targets (95% CI ±5% sensitivity, ±3% specificity), with κ and subgroup analyses to confirm reproducibility and mitigate bias.

Collaborators

Charles Hatt 4D Medical
Asif Abdul Hameed Henry Ford Health
Krishna Thavarajah Henry Ford Health
Dharshan Vummidi Henry Ford Health
Thomas Keimig Henry Ford Health
Adam Alessio Michigan State University