Skip to Main Content
An official website of the United States government
Login Change: CDAS is now using the NIH Researcher Auth Service (RAS) to manage our existing login options. When logging into CDAS, you will be taken to RAS, where you can select your preferred login provider to access CDAS.

Parameter-Efficient Fine-Tuning of a 3D CT Foundation Model for Lung Cancer Risk Prediction

Principal Investigator

Name
Akash Pattnaik

Degrees
Ph.D.

Institution
HOPPR AI

Position Title
Research Scientist

Email
akash@hoppr.ai

About this CDAS Project

Study
NLST (Learn more about this study)

Project ID
NLST-1528

Initial CDAS Request Approval
Aug 24, 2026

Title
Parameter-Efficient Fine-Tuning of a 3D CT Foundation Model for Lung Cancer Risk Prediction

Summary
Lung cancer remains the leading cause of cancer death. Low-dose CT (LDCT) screening has been shown to reduce mortality through early detection by 25%, but realizing this benefit depends on tools that can stratify individual risk from a single scan. This project applies NLST data to a pre-trained 3D CT Foundation Model to build lung cancer risk prediction models, replicating the study design of Sybil and Sybil 1.5 for predicting biopsy-confirmed lung cancer within 6 years.

Study Design

We request all available radiologic and clinical data from NLST participants in the LDCT arm, along with the Sybil model's predicted probabilities for each participant, to support both cohort construction and direct model comparison. From this data, we will curate a sample of 15,000 NLST participants, including all lung cancers and each participant's initial LDCT and up to two annual follow-up LDCTs when available. Following Sybil's split, LDCTs from participants in the Ardila et al. test set will be assigned to our test set (n = 2,328) and held out from training. Remaining participants will be randomly assigned to a training set (n = 10,200) or a development set (n = 2,472) used as a proxy for test performance during model development. Set assignment will be made at the participant level to prevent leakage. Within each LDCT, the single series with the thinnest slices will be selected for analysis. An LDCT will be labeled positive for future cancer risk if biopsy-confirmed lung cancer was diagnosed within 6 years, regardless of nodule presence on that exam.

Analysis Plan

We will use the curated cohort to fine-tune the HOPPR 3D CT Foundation Model, which is a self-supervised ViT-L backbone pretrained on large-scale spatiotemporal data. Following Sybil's design, the model's output will be a set of six calibrated probabilities representing lung cancer risk at 1 through 6 years following a given LDCT. Performance will be evaluated using Uno's concordance (C)-index and the area under the receiver operating characteristic (ROC) curve at each year through year 6, directly benchmarked against published Sybil and Sybil 1.5 results, as well as the Sybil model's own predicted probabilities, using the matched cohort and split. The C-index will assess whether, among a randomly selected pair of LDCTs, the scan closer to a cancer diagnosis received a higher predicted risk. ROC curves will characterize sensitivity-specificity tradeoffs at each horizon, informing threshold selection depending on clinical priorities (e.g., minimizing false positives). Confidence intervals will be computed via bootstrapping with 10,000 resamples, clustering LDCTs by participant. A p-value of .05 will be considered statistically significant for all tests.

Aims

Aim 1: Build a matched NLST cohort. Obtain all available radiologic and clinical data for NLST LDCT-arm participants, together with Sybil's predicted probabilities for each participant, and curate a matched sample of 15,000 participants replicating Sybil's train (n = 10,200), development (n = 2,472), and test (n = 2,328) splits, including the original Ardila et al. test set, for direct benchmarking against Sybil and Sybil 1.5.
Aim 2: Fine-tune a 3D CT foundation model for lung cancer risk prediction, with CAC as a secondary output. Adapt the HOPPR 3D CT Foundation Model to predict 6-year lung cancer risk from a single LDCT using a frozen backbone with vision LoRA (rank = 8, α = 16, dropout = 0.15) on attention and projection layers, jointly training a coronary artery calcium (CAC) binary classification head to assess whether multi-task learning improves or complements risk stratification.
Aim 3: Benchmark against Sybil and establish new state-of-the-art. Evaluate risk predictions at 1–6 years using Uno's C-index and year-by-year AUC, with bootstrapped confidence intervals (10,000 resamples, clustered by participant), compared directly against both published Sybil and Sybil 1.5 results and Sybil's own predicted probabilities.

Collaborators

Akash Pattnaik HOPPR AI
Harris Bergman HOPPR AI
Kalina Slavkova HOPPR AI
Ameen Salahudeen Taos Bio