A Dual-Center, Interpretable Deep Learning-Radiomics Fusion Model for Distinguishing Nasopharyngeal Carcinoma from Nasopharyngeal Lymphoma Using FDG PET/CT.
Authors
Affiliations (6)
Affiliations (6)
- Department of Nuclear Medicine, Affiliated Hospital of Guangdong Medical University, China (M.N.L., C.M.G.).
- Undergraduate student, Guangdong Medical University, China (X.G., D.R.H., X.L.L.).
- Cancer Center, the Affiliated Hospital of Guangdong Medical University, China (H.W.L.).
- Department of Nuclear Medicine, The First People's Hospital of Foshan, China (M.Y.).
- Huiying Medical Technology Co., Ltd., Room A206, B2, Dongsheng Science and Technology Park, Beijing, China (Y.J.).
- Department of Nuclear Medicine, Affiliated Hospital of Guangdong Medical University, No. 57 Ren-min South Road, Xiashan District, ZhanJiang, Guangdong, China (J.X.X.). Electronic address: [email protected].
Abstract
Accurate differentiation between Nasopharyngeal Carcinoma (NPC) and Nasopharyngeal Lymphoma (NPL) is critical for clinical management. We aimed to develop an interpretable deep learning-radiomics fusion model using positron emission tomography/computed tomography (PET/CT) and validate it externally. In this retrospective dual-center study, we enrolled 434 patients (349 internal, 85 external) with pathologically confirmed, treatment-naive NPC or NPL who underwent baseline FDG PET/CT. We extracted 1688 radiomics features and deep learning features from 3D-ResNet50. A three-step feature selection was applied, and a multimodal logistic regression model was constructed. Model performance was evaluated on an external test set, and SHapley Additive exPlanations (SHAP) analysis was used for interpretation. In this retrospective dual-center study, a total of 434 patients (434 primary lesions) were enrolled, including 349 from the internal cohort and 85 from the external cohort. Among them, 329 patients (75.8%) had NPC ( internal: 266, external: 63) and 105 patients (24.2%) had NPL( internal: 83, external: 22). The mean age of NPC patients was 50.5 ± 13.8 years, while that of NPL patients was 59.0 ± 16.5 years. The final fusion model (denoted as the ALL model) achieved an AUC of 0.833, specificity of 95.5%, and accuracy of 75.3% in external validation, outperforming all single-modality models (all P < 0.05). SHAP analysis identified key predictors: original_shape_MinorAxisLength_PET and lbp-3D-k_glrlm_RunLengthNonUniformity_PET for NPL, and wavelet-LLL_firstorder_Maximum_CT features for NPC. We developed a robust, interpretable fusion model for NPC/NPL differentiation, offering high specificity and transparent decision logic, with potential as a non-invasive clinical decision-support tool.