Deep learning model can predict lung cancer risk from a single scan

According to recently published research, deep learning assessment of a single low-dose computed tomography (LDCT) scan could provide highly accurate prediction of future lung cancer.

In a retrospective study published in the Journal of Clinical Oncology, researchers trained and developed a deep learning algorithm with a total of 35,001 LDCT scans of participants in the National Lung Screening Trial (NLST). To predict future lung cancer based on LDCT assessment alone, the deep learning algorithm had a 1-year area under the curve (AUC) of 92%, a 2-year AUC of 86%, and a 6-year concordance index (C index) of 75%, according to the study.

In subsequent external validation tests, the researchers tested the deep learning algorithm with a dataset of 13,309 LDCTs (6,392 patients) from Massachusetts General Hospital (MGH) and a dataset of 12,480 LDCTs (10,696 patients) from Chang Gung Memorial Hospital (CGMH). In Taiwan. The study authors found that the deep learning algorithm had an AUC of 86% over 1 year and a C-index of 81% over 6 years in the MGH dataset, as well as an AUC of 94% over 1 year and a C index of 80% in the MGH dataset. CGMH dataset.

While acknowledging that prospective clinical trials are needed to confirm the clinical utility of these study results, the researchers suggested that the deep learning algorithm could help reduce unnecessary follow-up imaging in patients considered as having low-risk lung nodules.

“Based on our clinical results, a potential clinical application is to use (the deep learning algorithm) to reduce follow-up scans or biopsies in patients with low-risk nodules,” Regina Barzilay wrote. , Ph.D., who is affiliated with the Department of Electrical Engineering and Computer Science and the Jameel Clinic at the Massachusetts Institute of Technology in Cambridge, Mass, and colleagues. “…In our evaluation of the NLST test set, (the deep learning algorithm) further reduced the FPR (false positive rate) to 8% for basic scans compared to 14% for Lung -RADS 1.0, while maintaining equivalent sensitivity.”

(Editor’s note: For related content, see “Nine takeaways from a recent meta-analysis on lung cancer screening with low-dose computed tomography” and “Can ultra-low-dose CT be effective for lung cancer screening in current or past smokers?”)

The researchers also noted a correlation with the deep learning algorithm between predicting high cancer risk and targeted identification of where malignant nodules would occur.

“We noted an association between (the ability of the deep learning algorithm) to correctly lateralise the location of future cancers and the likelihood of a PDPD receiving a high-risk score, indicating that when (the algorithm deep learning) predicts future high-risk lung cancer, the signal it uses localizes to specific risk regions rather than being evenly distributed across the entire chest,” Barzilay and colleagues.

Regarding the limitations of the study, the authors acknowledged the retrospective nature of the study. Noting that the study cohorts were made up of patients participating in lung cancer screening programs, they said they could not assess the ability of the deep learning model to predict lung cancer in patients. people who do not participate in lung cancer screening programs. The study authors also acknowledged suboptimal diversity in the study cohort and the lack of a true comparison model.