A Fusion Model of MRI Deep Transfer Learning and Radiomics for Discriminating between Pilocytic Astrocytoma and Adamantinomatous Craniopharyngioma.

Abstract

Rationale and objectives: This study aimed to develop and validate a fusion model combining MRI deep transfer learning (DTL) and radiomics for discriminating between pilocytic astrocytoma (PA) and adamantinomatous craniopharyngioma (ACP) in the sellar region.

Methods: This study included 348 patients with histologically confirmed PA (n = 139) and ACP (n = 209). Data were randomly divided into training and testing cohorts in a 7:3 ratio. Pre-trained ResNet50 network was utilized to extract DTL features from T1WI, T2WI, and CET1, while radiomics features (Rad) were extracted from manually delineated images of the same modalities. The fusion feature set (DLR) was constructed by integrating these features. Semantic features were used to develop clinical models. Pearson rank correlation and The least absolute shrinkage and selection operator regression were used for feature selection, and K-nearest neighbor algorithm was applied to establish the model. The performance of the model was evaluated using receiver operating characteristic curve. DeLong's test was performed to assess differences between models, and decision curve analysis was conducted to evaluate the clinical utility of the models.

Results: The DLR model achieved AUC values of 0.945 (95% CI, 0.9149-0.9760) in the training cohort and 0.929 (95% CI, 0.8824-0.9762) in the testing cohort, significantly higher than those of models using DTL features, Rad features, or clinical features alone.

Conclusion: The fusion model based on MRI deep transfer learning and radiomics (DLR) demonstrated high accuracy and clinical utility in discriminating between PA and ACP, providing an effective tool for the non-invasive diagnosis of these two diseases.