Imaging-only deep learning on 18F-FDG PET/CT scored pulmonary nodule malignancy with non-inferior performance to the guideline-recommended Herder model in 533 indeterminate nodules, posting an AUC of 0.78 vs 0.73 (p=0.005). In this single-center retrospective study, the AITO-PETCT-MP model also performed within the range of 7 clinicians on a 161-nodule test set.

Why It Matters To Your Practice

• The model used imaging data alone, without requiring the added clinical variables used in the Herder approach.

• That could make malignancy estimation more scalable and consistent when evaluating suspicious pulmonary nodules on PET/CT.

• For clinicians following BTS pathways, the study suggests AI may alter downstream management recommendations, not just discrimination metrics.

Clinical Implications

• AITO-PETCT-MP achieved AUC 0.78 (95% CI, 0.70-0.85), compared with 0.73 (95% CI, 0.65-0.80) for the Herder model.

• Experienced clinicians averaged AUC 0.80 (95% CI, 0.75-0.85), with individual performance ranging from 0.74 to 0.87.

• Using BTS follow-up categories, the Herder model referred more benign nodules for potential direct treatment than the DL model or clinicians: 26/81 vs 3/81 and 3/81.

• But the DL model and clinicians assigned more malignant nodules to CT surveillance instead of direct treatment, highlighting a tradeoff between overtreatment and delayed escalation.

Insights

• The cohort included 533 indeterminate nodules, 268 malignant, from 436 patients; mean nodule diameter was 18.4 mm.

• Reference standard was histopathology for malignant neoplasm confirmation or at least 2 years of benign follow-up in the national cancer registry.

• The key signal is not that AI clearly beat clinicians, but that an imaging-only model landed in the same performance band as expert readers while differing from guideline-model treatment triage.

The Bottom Line

• This PET/CT deep learning model matched the Herder model on discrimination and approximated expert clinician performance using imaging alone.

• Before adoption, clinicians should focus less on AUC alone and more on how AI-driven risk estimates would shift surveillance, biopsy, and treatment decisions for individual nodules.