Multiparameter MRI-based automatic segmentation and diagnostic models for the differentiation of intracranial solitary fibrous tumors and meningiomas.

BACKGROUND

Intracranial solitary fibrous tumors (SFTs) and meningiomas are meningeal tumors with different malignancy levels and prognoses. Their similar imaging features make preoperative differentiation difficult, resulting in high misdiagnosis rates. Thus, accurately distinguishing SFTs from meningiomas preoperatively is vital for surgical planning and treatment strategies.

PATIENTS AND

METHODS

A total of 252 patients (56 SFTs and 196 meningiomas) data from January 2014 to May 2024 were used to train our models . A VB-Net deep learning network was employed to refine automatic segmentation. To identify SFTs and meningiomas, the segmented data were analyzed using machine learning to construct single-sequence and multi-sequence MRI models and combined with clinical/radiological features to develop a fusion index-related model.To enhance clinical applicability, we constructed a four-category model using the predictive probabilities from secondary classification as input features.

RESULTS

The VB-Net segmentation model performed well in both the tumor cores and the whole tumor, with DSCs of 0.87 (±0.17) and 0.79 (±0.26), respectively. The integration of clinical and radiological data enhanced the model's AUC to 0.957. Stratified analysis showed that the weighted AUC value reached 0.846 in the validation set.

CONCLUSION

The comprehensive system integrating automatic segmentation with diagnostic models can differentiate SFTs from meningiomas precisely.