Deep learning PET/CT model matches Herder model and clinicians for lung nodule malignancy prediction

Imaging-only deep learning PET/CT model (AUC 0.78) non-inferior to guideline-recommended Herder model (AUC 0.73, p=0.005) for pulmonary nodule malignancy; matched clinicians. Herder referred more benign nodules to direct treatment (26 vs 3/81). Retrospective, single-center.

• Retrospective, single-center study of 533 nodules (268 malignant) in 436 patients; test set 161 nodules.
• Herder model referred significantly more benign nodules for potential direct treatment (26/81) vs. the deep learning model and clinicians (both 3/81), while assigning fewer malignant cases to CT surveillance.
• No external validation; limited to a single center and retrospective design.

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