Massachusetts General Hospital and MIT researchers stand in front of a CT scanner at MGH, where they generated part of the validation data. Each patient's chance of developing lung cancer is calculated using unique characteristics and a deep-learning model.

Purpose

Lung cancer was the biggest cause of cancer-related death in 2020, responsible for an anticipated 1.7 million deaths worldwide. Although most people who could benefit from low-dose computed tomography (LDCT) for lung cancer screening are not being screened, LDCT is beneficial. Tools that predict future cancer risk on an individual basis have the potential to target interventions to those most likely to benefit. It was postulated by the researchers that using the volumetric LDCT data as a whole, they could construct a deep learning model to predict individual risk.

Despite new lung cancer treatments, most patients die. Low-dose computed tomography (LDCT) scans are the standard lung cancer screening method. By independently assessing the LDCT image data, Sybil may forecast a patient's likelihood of acquiring lung cancer within the next six years, which is not possible without screening.

The Jameel Clinic, the Massachusetts General Hospital Cancer Center (MGCC), and the Children's Hospital Los Angeles (CHLA) published a paper in the Journal of Clinical Oncology. Models achieving a C-index score Even better, the ROC-AUCs for Sybil's one-year predictions ranged from 0.86 to 0.94, with 1.00 being the best possible score.

Method

The team constructed Sybil using National Lung Screening Trial data (NLST). Sybil operates in real-time in a radiological reading station with one LDCT, no clinical data, and no radiologist annotations. Sybil was tested on 12,280 LDCTs from three hospitals: a holdout group of 6,282 NLST participants, 8,821 MGH, and 12,280 Chang Gung Memorial Hospital (CGMH, which included people with a range of smoking history, including nonsmokers).

Conclusion

MGH and MIT researchers pose in front of a CT scanner that generates validation data. According to their first findings, Sybil could be used in clinical practice to reduce unnecessary scans and biopsies for patients with low-risk lesions. In addition, the Lung-RADS approach was adopted as the gold standard in the United States because it increased the specificity of LDCT screening compared to the nodule evaluation algorithm used in the NLST study.

By evaluating the NLST test set, Sybil could reduce the FPR for baseline scans from 14% to 8%, compared to Lung-RADS 1.0's 16%, while keeping the same level of sensitivity. False negatives, also known as missed interval cancers, are a severe medical and legal concern for patients enrolled in LCS programmes. According to a retrospective analysis of the 44 missed interval lung tumours in the NLST, most of these cases could have been averted if not for human error.