A multimodal deep learning approach for future lung cancer risk prediction based on low-dose screening CT and individual risk factors
Principal Investigator
Name
Paolo Soda
Degrees
PhD
Institution
Università Campus Bio-Medico di Roma
Position Title
Full Professor
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-1408
Initial CDAS Request Approval
Mar 18, 2025
Title
A multimodal deep learning approach for future lung cancer risk prediction based on low-dose screening CT and individual risk factors
Summary
Lung cancer screening using low-dose computed tomography (LDCT) has significantly improved early detection, yet challenges remain in balancing sensitivity and specificity. The high false positive rate leads to unnecessary invasive diagnostic procedures, while fixed screening intervals may contribute to missed cancers. This project aims to develop a predictive model to estimate an individual’s six-year lung cancer risk at the time of their first screening.
Our approach will leverage multimodal deep learning to integrate LDCT imaging with baseline clinical characteristics and known risk factors. Image preprocessing can be complemented with denoising methods that leverage generative AI. By learning complex interactions between imaging and non-imaging data, the model will provide a more personalized risk assessment compared to traditional risk models. This enhanced risk prediction could enable adaptive screening strategies, optimizing follow-up intervals based on personalized risk rather than fixed schedules.
Our approach will leverage multimodal deep learning to integrate LDCT imaging with baseline clinical characteristics and known risk factors. Image preprocessing can be complemented with denoising methods that leverage generative AI. By learning complex interactions between imaging and non-imaging data, the model will provide a more personalized risk assessment compared to traditional risk models. This enhanced risk prediction could enable adaptive screening strategies, optimizing follow-up intervals based on personalized risk rather than fixed schedules.
Aims
The expected impact of this research includes:
- Reducing false positives, thereby minimizing unnecessary invasive procedures.
- Optimizing screening intervals, decreasing missed cancer incidences while reducing unnecessary scans.
- Enhancing personalized lung cancer management, improving early detection and resource allocation.
Collaborators
Fatih Aksu, Università Campus Bio-Medico di Roma
Camillo Maria Caruso, Università Campus Bio-Medico di Roma
Valerio Guarrasi, Università Campus Bio-Medico di Roma
Rosa Sicilia, Università Campus Bio-Medico di Roma
Pierangelo Veltri, Università della Calabria