Deep learning models outperform Brock for incidental pulmonary nodule malignancy estimation in multicentre dataset

Two deep learning models for pulmonary nodule malignancy estimation on CT achieved area under the curve (AUC) values of 0.74 and 0.72 vs 0.63 for Brock (both p<0.01), with higher specificity (60% vs 44%) at fixed sensitivity in a multicentre dataset of 269 incidental nodules.

• Retrospective case-control dataset of 269 nodules (89 malignant) from 231 patients across three centres, enriched for malignancy.
• The screening-trained and additionally clinically trained deep learning models performed similarly, while the Brock model's AUC was 0.63.
• Consistent performance across centres, but prospective validation with real-world prevalence is needed as a next step.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.