Validating the ability to differentiate skull base osteomyelitis and nasopharyngeal carcinoma using radiomics parameters obtained from magnetic resonance imaging

Skull base osteomyelitis (SBO) is an infectious disease that represents the most severe form of the necrotizing external otitis spectrum. It can be confused radiologically with nasopharyngeal carcinoma (NPC) in cases with extensive skull base involvement including paraclival and parapharyngeal soft tissues or in the absence of a history of external otitis. Despite therapeutic advances, SBO can be fatal and early diagnosis and prompt treatment may improve clinical outcomes. Therefore, various strategies are being investigated to enhance the diagnostic accuracy of imaging tools, which are indispensable for early diagnosis. The aim of this study is to investigate the role of radiomics obtained from magnetic resonance imaging (MRI) in distinguishing SBO from NPC. A total of 93 patients (50 NPC and 43 SBO) from three different centers were divided into training (32 NPC and 31 SBO) and external validation (18 NPC and 12 SBO) cohorts in a ratio of approximately 7:3. Out of the 2553 radiomics features extracted through volumetric segmentation on three different series (pre-contrast T1-weighted, fat-suppressed T2-weighted, and post-contrast fat-suppressed T1-weighted), features with an intra-class correlation coefficient (ICC) <0.9 and features with zero coefficients determined by Least Absolute Shrinkage and Selection Operator (LASSO) regression analysis were eliminated. A radiomics model was established using remaining radiomics features with a linear kernel support vector machine (SVM). Among the semantic (conventional) MRI findings that were investigated, significant differences between the two groups (p < 0.05) were observed in tissue plane distortion, lateral extension, contrast enhancement, dural thickening, and lymphadenopathy. These variables exhibited no zero values in subgroup analyses within the training set and demonstrated significant associations between the two groups in univariate analyses (p < 0.05). A semantic Support Vector Machine (SVM) model was subsequently developed using these variables. Additionally, we computed ADCmean by avaraging three different ADC values obtained through the placement of separate regions of interest (ROIs) within the lesions, excluding cystic or necrotic areas. Normalized ADC (nADC) values were calculated by dividing the ADCmean values by the reference ADC values acquired from a region of interest (ROI) that covers the entire axial section of the spinal cord. The optimal cut-off values for these ADC parameters were established in the training dataset. The classification performance of these cut-off values was then assessed on both the training and external test datasets. We further trained a SVM classifier using the ADCmean values from the training set and independently evaluated its performance on the test set. The predictive performances of the SVM models were analyzed using Receiver Operating Curve (ROC) curves and Area Under the Curve (AUC) values. Notably, the radiomics model achieved a remarkable AUC value of 0.972 in the independent test dataset, outperforming both the semantic model (AUC: 0.902) and the ADC model (AUC: 0.681). The training set cut-offs were 0.954 x 10-3 mm2/s for ADC mean and 1.133 for nADC. Thus, the AUC values obtained according to these cut-off points were 0.849 and 0.897, in training set and 0.688 and 0.705, in independent test data set, respectively. In conclusion, the radiomic variables have demonstrated a solid diagnostic performance in discriminating the two conditions effectively. Although a normalization method was proposed, ADC parameters are sensitive to the dataset and are not generalizable. Nevertheless, if each center determines its own cut-off values, ADC values may play an important role in distinguishing between the two diseases.