Sybil
Deep-learning model that predicts an individual's risk of lung cancer over the next one to six years from a single low-dose chest CT, without radiologist annotations or clinical variables.
At a glance
What it does
Takes the full LDCT volume and outputs calibrated risk for years 1–6. It is a risk stratifier for screening programmes (who needs a shorter interval), not a nodule detector.
Notable uses in oncology
Validated on three independent cohorts (NLST test set, Massachusetts General Hospital, Chang Gung Memorial Hospital in Taiwan); one-year AUCs of 0.92, 0.86 and 0.94 respectively in the JCO paper.
Tasks, data types and cancers
Architecture
Training data
Trained on low-dose CT scans from the National Lung Screening Trial (NLST) with follow-up cancer outcomes; no radiologist labels beyond the trial's outcome data.
Linked datasets
- training NLST — National Lung Screening Trial Controlled access — training and held-out test
Evaluation
How to run
pip install sybil; the package downloads the ensemble weights. Input must be a full LDCT series — thick-slice or contrast CTs are outside the validated domain.
Regulatory status and intended use
Regulatory status is quoted from the source linked above and can change. Research-use-only models must not be used for clinical decisions.
Limitations and bias
- Trained on a screening population of heavy smokers; performance in never-smokers rests on one Taiwanese cohort.
- Sensitive to acquisition protocol; validate locally before any programme use.
Sources
This page is educational — it is not medical advice and does not replace consultation with an oncologist. Diagnostic and treatment decisions are made solely by specialist physicians.