- A deep learning artificial intelligence model detected malignant lung cancer on low-dose chest computed tomography (CT) scans with performance matching or exceeding that of radiologists.
Why this matters
- If clinically validated, this model can improve accuracy and optimize screening.
- Deep learning algorithm had 3 main components:
- model that analyzed whole-CT volumes with cancer;
- detection model for suspicious region-of-interest; and
- independent risk prediction model which also incorporates previous scans if available.
- National Lung Cancer Screening Trial (NLST) dataset was used for the development (n=14,851; 578 patients with biopsy-confirmed lung cancer).
- On the test dataset (n=6716 cases; 86 cancer cases), the model achieved an area under the curve (AUC) of 94.4%.
- 6 radiologists reviewed subset test NLST dataset (n=507; 83 cancer cases):
- When CT volumes from only current year were used, the model achieved AUC of 95.9% with sensitivity and specificity superior to the average radiologist.
- When CT volumes from both the current and previous years were used, the model’s AUC was 92.6% and it matched average reader sensitivity and specificity.
- In an independent cohort (n=1739; 27 cancer cases), the model’s AUC was 95.5% with sensitivity and specificity of 81.5% and 89.3%, respectively for lung imaging reporting and data system category ≥3 nodules.