A comparison of 18 F-FDG PET-based radiomics and deep learning in predicting regional lymph node metastasis in patients with resectable lung adenocarcinoma: a cross-scanner and temporal validation study.

Objective: The performance of 18 F-FDG PET-based radiomics and deep learning in detecting pathological regional nodal metastasis (pN+) in resectable lung adenocarcinoma varies, and their use across different generations of PET machines has not been thoroughly investigated. We compared handcrafted radiomics and deep learning using different PET scanners to predict pN+ in resectable lung adenocarcinoma.

Methods: We retrospectively analyzed pretreatment 18 F-FDG PET from 148 lung adenocarcinoma patients who underwent curative surgery. Patients were separated into analog (n = 131) and digital (n = 17) PET cohorts. Handcrafted radiomics and a ResNet-50 deep-learning model of the primary tumor were used to predict pN+ status. Models were trained in the analog PET cohort, and the digital PET cohort was used for cross-scanner validation.

Results: In the analog PET cohort, entropy, a handcrafted radiomics, independently predicted pN+. However, the areas under the receiver-operating-characteristic curves (AUCs) and accuracy for entropy were only 0.676 and 62.6%, respectively. The ResNet-50 model demonstrated a better AUC and accuracy of 0.929 and 94.7%, respectively. In the digital PET validation cohort, the ResNet-50 model also demonstrated better AUC (0.871 versus 0.697) and accuracy (88.2% versus 64.7%) than entropy. The ResNet-50 model achieved comparable specificity to visual interpretation but with superior sensitivity (83.3% versus 66.7%) in the digital PET cohort.

Conclusion: Applying deep learning across different generations of PET scanners may be feasible and better predict pN+ than handcrafted radiomics. Deep learning may complement visual interpretation and facilitate tailored therapeutic strategies for resectable lung adenocarcinoma.