Replication and Comparative Evaluation of Deep Learning Models (Sybil and M3FM) for Lung Cancer Risk Prediction Using NLST Low-Dose CT Scans
Principal Investigator
Name
Aristotelis Tsirigos
Degrees
Ph.D
Institution
NYU Langone Health
Position Title
Full Professor
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-1481
Initial CDAS Request Approval
Nov 13, 2025
Title
Replication and Comparative Evaluation of Deep Learning Models (Sybil and M3FM) for Lung Cancer Risk Prediction Using NLST Low-Dose CT Scans
Summary
We are working on a project to replicate and study the Sybil model, a deep learning system that predicts future lung cancer risk from a single low-dose CT scan, and to extend this work by evaluating the more recent M3FM multimodal multitask model on the same NLST data.
Our goals are to:
(1) reproduce and evaluate Sybil using the NLST lung cancer selection CTs;
(2) train and evaluate an NLST-based implementation of the M3FM model (Niu et. al) for lung cancer risk, cardiovascular disease (CVD) diagnosis and mortality risk, and related chest abnormalities; and (3) compare performance and behavior of these models.
Analyses will be performed on secure NYU institutional servers, and all NLST data will remain restricted to our research group (Prof. Aristotelis Tsirigos and myself) and will not be redistributed.
Aims
Aim 1: Recreate and evaluate the Sybil pipeline , from preprocessing through inference, using the NLST Lung Cancer Selection subset, and compare our results to the original publication.
Aim 2: Recreate and evaluate the M3FM model (Niu et. al) using LDCT scans plus linked clinical variables needed and mentioned in the research paper (demographics, smoking history, disease history, family history, outcomes) to jointly model lung cancer risk, CVD outcomes, and chest abnormalities.
Aim 3: Systematically compare Sybil and M3FM on shared tasks (e.g., multi-year lung cancer risk) to understand how multimodal multitask learning changes performance and calibration.
Aim 4: Use insights from these models to inform future work on NYU Langone lung cancer screening cohorts.
Hypothesis: Deep learning models can identify subtle imaging and clinical patterns in low-dose CT screening data that predict future lung cancer and cardiovascular outcomes. Multimodal multitask models such as M3FM may provide improved risk estimates over single-task image-only models like Sybil, while using the same underlying NLST cohort.
Collaborators
Aristotelis Tsirigos New York University (Affiliated with NYU Langone)
ELEKTRA MANOLAKOS New York University (Affiliated with NYU Langone)
Aristotelis Tsirigos New York University (Affiliated with NYU Langone)