To create an automated segmentation model of meningioma gross tumor volume (GTV) using pre-radiotherapy MRI scans.

Meningioma is the most frequent primary intracranial brain tumor.  Radiotherapy is typically employed in the context of residual disease to delay or prevent recurrence. Manual delineation of meningioma is often prone to errors and can be time-consuming. However, this task can be augmented by machine learning. Automated segmentation models have the potential to improve efficiency and consistency in treatment planning. For this model, an open-source dataset from the BraTS 2024 challenge was utilized, which includes deidentified pre-radiotherapy MRI scans for patients with meningioma.

The code implements an automated MRI image segmentation workflow using the nnU-Net framework. It begins by fetching the BraTS dataset from Synapse, extracting it, and running the nnU-Net model to generate predicted segmentation masks. The model is configured for 3D full-resolution inference using nnU-Net Trainer. After predictions, the code visualizes multiple MRI slices with the predicted masks overlaid.

The nnU-Net model demonstrated excellent performance in segmenting brain tumors from MRI scans. It achieved a Dice Similarity Coefficient (DSC) of 0.67, indicating significant overlap between the predicted tumor masks and the ground truth annotations. Visualization of the predicted masks overlaid on MRI slices clearly showed well-defined and accurate tumor boundaries. Additionally, the model generalized effectively across different MRI modalities, including T1-weighted, T2-weighted, and FLAIR images. In every case, the model reliably distinguished tumor regions from healthy tissue, even in challenging cases with irregular tumor shapes.  The Gradio interface enabled real-time segmentation, processing each MRI scan in under 3 minutes with minimal computational resources.

Our nnU-Net model, with limited computing power, efficiently delineates the meningioma GTV. This simplifies treatment planning. Future research should confirm the model with larger, multimodal imaging datasets. This will improve its accuracy and predict meningioma recurrence risks.