Distinguishing atypical parotid carcinomas and pleomorphic adenomas based on multiphasic computed tomography radiomics nomogram: a multicenter study

Objective This study aimed to develop, validate, and test a comprehensive radiomics prediction model using clinical data and contrast-enhanced multiphasic computed tomography (CT) scans for differentiating between atypical parotid carcinomas (PCAs) and pleomorphic adenomas (PAs) within a multicenter cohort. Materials and methods The study involved 218 patients diagnosed with either PAs (n=162) or atypical PCAs (n=56) (no invasion of adjacent tissues or lymph node metastases) across three anonymized hospitals, divided into a training set (n=175) and a validation set (n=43). Clinical features and radiological findings were used to develop a clinical model. Radiomics features were extracted from multi-phase contrast-enhanced CT, with feature selection achieved through statistical methods and the least absolute shrinkage and selection operator (LASSO). Radiomics signature were developed using a Light Gradient Boosting Decision Tree (LightGBM) model. A radiomics nomogram integrating significant clinical risk factors with the radiomics signature was created, with external validation conducted on an independent dataset of 32 patients from two additional hospitals. Results In the training set, the multiphase models (model A+P , model A+V and model A+P+V ) demonstrated significantly superior predictive performance compared to the arterial-phase-only model (model A ) (DeLong’s test, p=0.04–0.02). However, no significant differences emerged between the models in the validation or independent testing sets (p > 0.05). Based on recall and F1-score evaluations in the independent testing set, model A+P was selected for integration with clinical risk factors to develop a radiomics nomogram. This nomogram demonstrated excellent diagnostic performance, achieving AUCs of 1.000 (training), 0.854 (validation) and 0.783 (independent testing), accuracies of 1.000, 0.864 and 0.750, and F1-scores of 1.000, 0.914 and 0.826, respectively. Key discriminative features — cluster shade, run-length non-uniformity and first-order mean, extracted via wavelet or exponential filters — significantly differentiated atypical PCAs from PAs. Conclusion The CT-based radiomics nomogram, supplemented by machine learning, effectively differentiates atypical PCAs from PAs, presenting a non-invasive diagnostic tool that could guide treatment decisions and reduce the need for invasive procedures.