Abstract
OBJECTIVES: To evaluate and compare the efficacy of radiomics models derived from T2-weighted and/or contrast-enhanced T1-weighted (CET1) images in assessing pituitary macroadenoma consistency, and to validate their performance stability under varying MRI field strengths and scanner vendors. METHODS: A total of 133 patients with pathologically proven pituitary macroadenomas (35 fibrous, 98 non-fibrous) were retrospectively included. Three logistic regression models were constructed: a T2 model, a CET1 model, and a T2-CET1 combined model, based on features selected from coronal T2-weighted and contrast-enhanced T1-weighted (CET1) images. An external validation cohort of 40 patients (20 fibrous, 20 non-fibrous) was selected from another healthcare institution. Model performance was primarily evaluated using receiver operating characteristic (ROC) analysis. Stratified analyses were performed to compare the predictive performance of the models across different magnetic field strengths (1.5T and 3.0T) and scanner vendors. RESULTS: In the test dataset, the T2-CET1 combined model outperformed both the independent CET1 and T2 models, achieving an AUC of 0.86, accuracy of 83.3%, sensitivity of 83.3%, and specificity of 83.8%. This compares favorably with the CET1 model (AUC: 0.80, accuracy: 73.3%, sensitivity: 80.0%, specificity: 66.7%) and the T2 model (AUC: 0.79, accuracy: 76.7%, sensitivity: 76.7%, specificity: 76.7%). The combined model's superior performance extended to the external validation set, where its AUC (0.865) exceeded that of the CET1 model (0.765) and the T2 model (0.811). Performance varied by MRI field strength. For 1.5T systems, AUCs were 0.50 (CET1), 0.76 (T2), and 0.58 (combined). For 3.0T systems, the corresponding AUCs were 0.61, 0.83, and 0.56. Similarly, analysis by specific scanner model showed AUCs of 0.60 (CET1), 0.83 (T2), and 0.53 (combined) for one scanner, compared to 0.54, 0.84, and 0.52 for the other. CONCLUSIONS: Combining CET1 with T2 improves prediction performance for pituitary macroadenoma consistency. However, the T2 model demonstrates greater stability across different equipment than either the CET1 or combined models.