Deep Learning Model for Intracranial Hemangiopericytoma and Meningioma Classification

Area of Science:

• Neurosurgery
• Radiology
• Artificial Intelligence in Medicine

Background:

• Intracranial hemangiopericytoma/solitary fibrous tumors (SFT/HPC) are rare neoplasms with malignant potential, often presenting with infiltrative growth, edema, bleeding, or bone destruction.
• SFT/HPC shares radiological similarities with meningioma, necessitating accurate preoperative differentiation due to differing clinical management and prognoses.

Purpose of the Study:

• To develop and validate deep learning approaches for discriminating between intracranial SFT/HPC and meningioma using routine preoperative MRI.
• To identify distinct radiological features that differentiate SFT/HPC from meningioma.

Main Methods:

• A cohort of 236 patients with histopathologically confirmed SFT/HPC (n=144) and meningioma (n=122) was analyzed.
• Radiological features were manually extracted, and a classification model was applied.
• A deep learning model (ResNet-50) was trained on T1-contrast MRI images, with attention mechanisms visualized using class activation maps.

Main Results:

• SFT/HPC exhibited significantly more venous sinus invasion, cystic components, and heterogeneous enhancement compared to meningioma (p < 0.001).
• The deep learning model achieved a classification accuracy of 0.889 and an AUC of 0.91 on the validation set.
• Attention maps highlighted tumor bulk and solid texture features as key discriminators between SFT/HPC and meningioma.

Conclusions:

• Deep learning models can effectively differentiate intracranial SFT/HPC from meningioma based on preoperative MRI.
• The model's ability to focus on specific textural features aids in distinguishing these challenging-to-differentiate tumors.