A deep learning algorithm demonstrated “excellent performance” in estimating the malignancy risk of pulmonary nodules and proved superior to the Pan-Canadian Early Detection of Lung Cancer (PanCan) model, researchers reported in Radiology.
Although lung cancer screening can identify nodules, a major limitation is that most pulmonary nodules are benign, the researchers explained. Accurate estimation of nodules’ malignancy risk is crucial, but it remains a challenge, they added.
With this in mind, the researchers developed a deep learning algorithm using data from the National Lung Screening Trial cohort, which included 16,077 nodules, 1249 of them malignant.
For external validation of the algorithm, the researchers used 3 cohorts from the Danish Lung Cancer Screening Trial (DLCST): a full cohort containing all 883 nodules (65 malignant) and 2 cancer-enriched cohorts with size matching of benign nodules (175 nodules, 59 malignant) and without size matching (177 nodules, 59 malignant).
The algorithm was validated and compared with the PanCan model in the full DLCST cohort. The area under the receiver operating characteristic curve (AUC) was used as the performance measure for the algorithm.
The algorithm outperformed the PanCan model, with AUCs of 0.93 and 0.90, respectively (P =.046).
The researchers used the cancer-enriched cohorts to compare the algorithm’s performance with the performance of 11 clinicians, including 4 thoracic radiologists, 5 radiology residents, and 2 pulmonologists.
The algorithm’s performance was comparable to the thoracic radiologists’ performance with regard to both randomly selected benign nodules (AUC, 0.96 vs 0.90; P =.11) and size-matched benign nodules (AUC, 0.86 vs 0.82; P =.26).
“This algorithm has the potential to provide reliable and reproducible malignancy risk scores for clinicians, which may help optimize management in lung cancer screening,” the researchers wrote.
They added that the algorithm may help reduce radiologists’ workloads while decreasing costs of lung cancer screening and unnecessary diagnostic interventions. The algorithm is freely accessible to the public for research purposes (https://grand-challenge.org/algorithms/pulmonary-nodule-malignancy-prediction).
Disclosures: Several study authors declared affiliations with the radiology industry. Please see the original reference for a full list of the authors’ disclosures.
Venkadesh KV, Setio AAA, Schreuder A, et al. Deep learning for malignancy risk estimation of pulmonary nodules detected at low-dose screening CT. Radiology. Published online May 18, 2021. doi:10.1148/radiol.2021204433