Diagnostic value in distinguishing benign from malignant parotid and submandibular gland tumors through computed tomography findings and texture analysis

Objective: This study aims to evaluate the diagnostic value of single-phase contrast-enhanced computed tomography (CT) texture analysis-radiomics features, conventional CT radiological findings, and the combined use of these features in classifying tumors in the parotid or submandibular gland as benign or malignant. Methodology: This retrospective study included 101 cases between October 2012 and December 2023 with masses reported on contrast-enhanced CT in the parotid or submandibular gland, confirmed by pathology following surgery or biopsy. For radiomics analysis, a two-dimensional (2D) region of interest (ROI) encompassing the lesion boundaries was manually segmented on a single CT slice using Slicer software (v.5.6.2). Radiomic features were extracted using the PyRadiomics library (v.3.1.0) in Python (v.3.9.10), resulting in a total of 116 original radiomic features for each lesion, including shape features (N = 23), first-order statistical features (N = 18), and second-order statistical features: GLCM features (N = 24), GLDM features (N = 14), GLRLM features (N = 16), GLSZM features (N = 16), and NGTDM features (N = 5). A logistic regression radiomics classification model was trained using features selected through feature reduction algorithms. Additionally, classification models were trained using conventional radiological-demographic features and combined features selected from both models. Model calibrations were evaluated using the Hosmer-Lemeshow test, and calibration plots were generated. Model performances were assessed using ROC analysis and compared using the DeLong test. Results: The mean AUC values for the radiomics model, the conventional radiological-demographic model, and the combined model were 0.889, 0.920, and 0.955, respectively. The mean sensitivities were 0.712, 0.760, and 0.828, while the mean specificities were 0.864, 0.915, and 0.911. The mean accuracies for these models were 0.808, 0.856, and 0.881, respectively. The combined model emerged as the best classification model. A comparison of the AUC values using the DeLong test revealed that the difference between the radiomics model and the combined model was statistically significant (p-value < 0.05). Conclusion: In our study, the combined model was the best model for evaluating the malignancy of salivary gland tumors and is thought to help guide tumor management in clinical practice.