Why estimate lung cancer risk in advance?
Low-dose chest CT is an effective screening test for lung cancer, but most people who could benefit do not take it. Tools able to assess an individual's future risk could help focus effort on those who would benefit most.
The team's idea was to have an AI read the whole volume of the scan and draw an individual risk from it, without asking for any extra information about the person.
How was Sybil built and tested?
The researchers trained a deep-learning model (an AI that learns from examples) on scans from a large American screening trial, the National Lung Screening Trial. Sybil analyses the whole 3D volume of a single scan. It needs no clinical data and no annotations from a radiologist, and it can run in real time, in the background, on the reading workstation.
It was then checked on three independent datasets: 6,282 scans set aside from the NLST, 8,821 scans from Massachusetts General Hospital (United States) and 12,280 scans from Chang Gung Memorial Hospital (Taiwan), which also included people who had never smoked.
What are the results?
For one-year risk, the area under the curve (a score from 0.5 for chance to 1 for perfection) is 0.92 on the NLST, 0.86 at Massachusetts General Hospital and 0.94 at Chang Gung. Over 6 years, the concordance index (which measures whether the model correctly ranks who develops a cancer before who does not) is 0.75, 0.81 and 0.80.
The authors conclude that Sybil can accurately estimate an individual's future lung cancer risk from a single scan, paving the way for more personalised screening. The model and its annotations are public.
Key figures
- 0.92 / 0.86 / 0.94area under the curve for one-year lung cancer risk (NLST, Massachusetts General Hospital, Chang Gung)
- 0.75 / 0.81 / 0.80concordance index over 6 years for the same three datasets
- 27,383scans in total across the three validation datasets (6,282 + 8,821 + 12,280)
What does this change for patients?
In time, a tool like this could help tailor follow-up: watch more closely the people whose estimated risk is higher, and invite people to screening more effectively. The authors stress that this must be confirmed by prospective clinical trials.
For now, Sybil is not a diagnostic test. It estimates a risk, not a cancer that is already there.
What are the limits?
The study is entirely retrospective: the scans were analysed after the fact. The people studied were already in a screening programme, so we do not know whether Sybil spots cancers that arise outside screening. In Taiwan, the researchers did not have detailed smoking data.
The groups studied do not include enough Black or Hispanic people to be sure the tool works for everyone. The authors write that Sybil's clinical applications still need to be studied.
Chest scan: preparing your follow-up with Mima
Sybil is a research effort: your scan is not read with this tool in routine care. In time, it could help adapt follow-up, once clinical trials have confirmed it.
With Mima, you can note your scan dates and results and prepare your questions. A concrete question: "How often should I have another scan, and why?"
Frequently asked questions
Does Sybil detect a cancer that is already there?
No, it estimates the risk of lung cancer in the following years from a single screening scan. A diagnosis relies on other tests and on doctors' judgement.
Does Sybil need medical information about me?
No. According to the study published in the Journal of Clinical Oncology in 2023, it needs only a scan, with no clinical data and no radiologist annotations.
Is Sybil already used in routine care?
The study does not say it is used in routine care. It concludes that its clinical applications still need to be studied.
Is the model available to researchers?
Yes. The authors state that their model and annotations are publicly available.
The original study
Mikhael PG, Wohlwend J, Yala A, et al. “Sybil: A Validated Deep Learning Model to Predict Future Lung Cancer Risk From a Single Low-Dose Chest Computed Tomography.” Journal of Clinical Oncology, 12 January 2023. DOI : 10.1200/JCO.22.01345
Read the full study in the Journal of Clinical OncologyThis article summarises a scientific study and is not medical advice.