Deep learning-based malignancy probability estimation of pulmonary nodules in PET/CT imaging.
OBJECTIVE The current BTS guidelines recommend evaluation of suspicious pulmonary nodules using [18F]FDG-PET/CT imaging, followed by Herder model risk stratification. However, it is based on limited imaging features, which may limit diagnostic accuracy. This study aims to develop a PET/CT-based deep learning (DL) model for malignancy probability estimation (AITO-PETCT-MP) and compare its performance to the Herder model and clinician performance. MATERIALS AND METHODS In a single-center retrospective study, we collected 533 indeterminate pulmonary nodules (268 malignant) with a mean diameter of 18.4 mm (SD ± 12.1) in 436 patients. Histopathological malignancy confirmation or a minimum 2-year benign national cancer registry follow-up served as the reference standard. Model diagnostic performance was compared against the Herder model and seven clinicians in a reader study on a test set of 161 nodules (80 malignant). RESULTS AITO-PETCT-MP achieved an AUC of 0.78 [95% CI: 0.70-0.85] compared to the Herder model: AUC = 0.73 [0.65-0.80] (non-inferiority: p = 0.005). On average, experienced clinicians achieved an AUC of 0.80 [0.75-0.85]. Stratifying into BTS follow-up categories, the Herder model referred more benign nodules for potential direct treatment (26/81) than AITO-PETCT-MP and clinicians (both 3/81), while AITO-PETCT-MP and clinicians assigned more malignant cases to CT surveillance instead of direct treatment. CONCLUSION AITO-PETCT-MP demonstrated non-inferior performance to the guideline-recommended Herder model, while only using imaging data. Diagnostic performance fell in the performance range of seven clinicians. Differences in BTS follow-up recommendations between the Herder model and clinicians suggest a difference in patient management compared to current clinical practice. KEY POINTS Question How well can an imaging-only deep learning model estimate pulmonary nodule malignancy probability on [18F]FDG-PET/CT compared to the established Herder model and expert clinicians? Findings The model (AITO-PETCT-MP) performed non-inferior to the Herder model (AUC 0.78 vs 0.73, p = 0.005) and comparably to seven expert readers (AUC 0.74-0.87). Clinical relevance BTS-based follow-up stratification showed the Herder model referred more benign nodules to potential direct treatment than clinicians and AITO-PETCT-MP, while assigning fewer malignant cases to surveillance. This suggests a difference between Herder recommendations and current clinical practice.