Skip to content

New AI Tool Predicts Six-Year Lung Cancer Risk From a Single CT Scan

Key findings

  • Using low-dose lung CT scans from the National Lung Screening Trial (NLST), researchers at Massachusetts General Hospital trained Sybil, an algorithm designed to predict future lung cancer risk using a three-dimensional convolutional neural network
  • Sybil requires only a single low-dose CT scan and runs in the background at a radiology reading station as soon as the image is available, without the need for input of demographic or clinical data and without requiring radiologists to annotate the images
  • In a test set of 6,382 CT scans from the NLST the model achieved a one-year AUC of 0.92, a two-year AUC of 0.86 and a concordance index of 0.75 over six years of prediction
  • Accuracy was similar in two independent external data sets from Mass General and a hospital in Taiwan, and performance was similar regardless of sex, age and smoking history
  • Sybil is freely available and may allow personalization of follow-up schedules for lung cancer screening

After lung cancer screening (LCS) with low-dose CT, frequency of follow-up imaging currently depends primarily on assessment of visible lung nodules. Researchers at Massachusetts General Hospital have created a deep learning model named Sybil that predicts lung cancer risk over several years and whether or not nodules are visible.

Lecia V. Sequist, MD, MPH, a medical oncologist and program director of the Cancer Early Detection and Diagnostics Clinic at the Mass General Cancer CenterFlorian J. Fintelmann, MD, physician–scientist in the Division of Thoracic Imaging and Intervention, and colleagues describe the model in the Journal of Clinical Oncology.

About Sybil

Sybil is a freely available validated three-dimensional convolutional neural network that examines an entire chest CT scan. The outcome is six scores representing calibrated probabilities that lung cancer will be diagnosed one to six years later.

The program can run in the background on a standard radiology reading station, without the need for human input such as demographic or clinical data or image annotation.

Performance Testing

Sybil was trained on 6,839 low-dose CTs from the National Lung Screening Trial (NLST). It was then tested on three data sets:

  • 6,382 additional CT scans from the NLST
  • 8,821 CT scans from adults receiving standard-of-care LCS at Mass General between 2015 and 2021
  • 12,280 scans from adults who underwent LCS at Chang Gung Memorial Hospital (CGMH) in Taiwan between 2007 and 2019—any adult without a personal cancer history was eligible, regardless of smoking history

For accuracy of lung cancer prediction the area under the curve was:

  • NLST—0.92 at 1 year, 0.86 at 2 years and 0.75 at 6 years; concordance index over six years of prediction, 0.75
  • Mass General—0.86 at 1 year, 0.82 at 2 years and not available at 6 years (lack of follow-up data); concordance index, 0.81
  • CGMH—0.94 at 1 year, 0.87 at 2 years and 0.74 at 6 years; concordance index, 0.80

Performance was similar regardless of patient sex, age and smoking history.

Clinical Applications

Sybil may allow follow-up schedules to be personalized for each patient. For example, one potential application is to decrease follow-up scans or biopsies for patients who have low Sybil risk scores. On the other hand, patients who have a negative Lung-RADS interpretation (score of 1 or 2) but high Sybil risk scores might need to be followed more closely.

Refer a patient to the Mass General Cancer Center

Learn about the Mass General Cancer Center


Lung cancer screening is an extremely underused tool. Review the lung cancer screening guidelines with the American Lung Cancer Screening Initiative's founder, Chi-Fu Jeffrey Yang, MD, of Massachusetts General Hospital.


Massachusetts General Hospital researchers created a fully automated deep learning system to quantify and characterize muscle and adipose tissue on CT scans at multiple thoracic vertebral levels. The system is as accurate as human analysts on routine CT scans, regardless of intravenous contrast.