Deep learning-radiomics-SUVmax integration from <sup>18</sup>F-FDG PET/CT predicts synchronous distant metastasis in nasopharyngeal carcinoma.
Authors
Affiliations (3)
Affiliations (3)
- Department of PET/CT Center, Jiangsu Cancer Hospital& Jiangsu Institute of Cancer Research & The Affiliated Cancer Hospital of Nanjing Medical University, Nanjing, China.
- School of Mathematics, Nanjing University, Nanjing, China.
- Department of Nuclear Medicine, Nanjing Drum Tower Hospital, The Affiliated Hospital of Nanjing University Medical School, Nanjing, China.
Abstract
This study aimed to develop and evaluate deep learning (DL) models of primary tumor (T) and cervical metastatic lymph nodes (CMLNs) using fluorine-18 fluorodeoxyglucose positron emission computed tomography (<sup>18</sup>F-FDG positron emission tomography/computed tomography (PET/CT)) imaging for predicting synchronous distant metastasis (SDM) in nasopharyngeal carcinoma (NPC) patients, integrating radiomic features, primary tumor maximum standardized uptake value (SUVmax-T) and clinical features. A two-center retrospective cohort of 218 patients (105 SDM, 113 non-SDM) was analyzed. Three DL architectures (ResNet18, DenseNet121, EfficientNet-B0) were trained on CT-only, PET-only, and fused CT+PET images. Nine models combining DL features, radiomic features of T + CMLNs or T-only, clinical data, and SUVmax-T using Multilayer Perceptron (MLP) and Random Forest (RF) classifiers were established. Multiple MLP benchmark models were built based on separate T/N staging, clinical data, SUVmax-T and deep features for comparative analysis. Performance was assessed via accuracy, precision, recall, F1-score, and ROC-AUC. Permutation feature importance analysis identified core predictors of the optimal RF model, and the optimal hybrid model underwent external multicenter validation. Dual-modal PET/CT networks surpassed single-modality models, with ResNet18 yielding the highest internal ROC-AUC of 0.804, superior to DenseNet121 (0.791) and EfficientNet-B0 (0.779). Integrating radiomic features (T), clinical parameters and SUVmax-T continuously elevated model discrimination, and the MLP hybrid model reached the peak internal ROC-AUC of 0.839 (recall = 0.747), outperforming the RF model (ROC-AUC = 0.787, recall = 0.777). Models built only on T/N staging reached an internal AUC of 0.476 and an external AUC of 0.513, while integrating deep features boosted their predictive AUC to 0.658 (internal) and 0.688 (external). External validation of the MLP hybrid model showed an ROC-AUC of 0.701 and a PR-AUC of 0.571. SUVmax-T was identified as the most important predictive feature in the RF hybrid model. Integrating ResNet18 (CT+PET) features with radiomic features (T), SUVmax-T and clinical data in an MLP framework yielded the highest performance, with SUVmax-T identified as an SDM predictor. Comparative benchmark analysis verified that this multimodal hybrid model achieved better discriminative performance than models relying solely on conventional T and N staging.