Skip to Main Content

An official website of the United States government

Principal Investigator
Name
Ziba Gandomkar
Degrees
Ph.D.
Institution
University of Sydney
Position Title
Researcher
Email
About this CDAS Project
Study
NLST (Learn more about this study)
Project ID
NLST-1319
Initial CDAS Request Approval
Sep 5, 2024
Title
AI-Driven Radiomic Models for Predicting Lung Cancer Risk and Disease Progression in Asbestos and Silica-Exposed Individuals: A Deep Learning Approach Integrating Clinical and Imaging Data
Summary
This project aims to develop advanced AI tools for predicting lung cancer risk and disease progression among individuals exposed to asbestos and silica, leveraging radiomic features extracted from lung CT images. Asbestos and silica exposure significantly increase the risk of developing dust-related interstitial lung disease (ILD) and lung cancer. Existing risk prediction methods primarily rely on occupational history and clinical data, which may lack the precision required to predict individual outcomes accurately. The study seeks to address this gap by integrating deep learning techniques with radiologist-guided image annotations and clinical data to enhance prediction accuracy.

The study will utilize a retrospective longitudinal cohort design, integrating data from four major sources: the National Lung Screening Trial (NLST), Heart of Australia, and I-MED. These datasets provide a substantial sample size of individuals exposed to asbestos and silica, including lung cancer cases and normal controls. The study population will consist of individuals with documented asbestos or silica exposure, with a focus on those with available CT imaging data.

The core objective of this research is twofold: (1) to develop an AI model capable of predicting lung cancer risk among exposed individuals by incorporating radiomic features and clinical data, and (2) to predict disease progression and patient outcomes through deep learning models trained on longitudinal CT scans. The Sybil deep learning architecture, which has already demonstrated strong predictive potential for lung cancer, will serve as the foundation of our model. We will modify and enhance this model by integrating radiologist-guided annotations that identify changes in CT images, such as nodules or ground-glass opacities, that may indicate future disease progression.

To improve performance, various deep learning strategies will be explored, including testing different convolutional neural network backbones and refining the attention-guided pooling layer. In addition, three types of attention mechanisms will be compared to determine which provides the most robust feature extraction for lung cancer and silicosis prediction. The model will also incorporate clinical and occupational data into a unified framework to provide a comprehensive risk assessment for each individual.

Given the smaller sample size of the asbestos and silica-exposed cohort, transfer learning will be employed to overcome data limitations. The model will initially be trained on a larger dataset of non-exposed individuals from the NLST to predict lung cancer and later fine-tuned using the dust-exposed cohort.

This study holds the potential to significantly improve risk prediction for lung cancer and disease progression in individuals exposed to hazardous dust particles. By integrating advanced AI-driven models with clinical data and radiomics, the project aims to provide personalized and accurate predictions that can inform preventive strategies, early interventions, and tailored patient management.
Aims

• Develop AI tools using radiomic features from lung CT images to predict lung cancer risk among asbestos and silica-exposed individuals.

• Develop AI models to predict disease progression and patient outcome.

• Enhance the Sybil architecture by testing different deep learning backbones and refining the attention-guided pooling layer: single attention module, separate modules for distinct features, and multi-head attention, to optimize radiomic feature extraction.

• Compare two time-to-event prediction models: DeepSurv, a semi-parametric non-linear continuous-time survival model, and DeepHitSingle, a non-parametric non-linear discrete-time survival model.

Collaborators

Dr. Ziba Gandomkar, The University of Sydney
Dr. Xuetong Tao, The University of Sydney
Dr. Moe Suleiman, The University of Sydney
Professor Patrick Brennan, The University of Sydney
Clinical Associate Professor Anthony Linton, The Asbestos Diseases Research Institute
Dr Elham Hosseini-Beheshti, The Asbestos Diseases Research Institute