Validation of Deep-Learning Models for Autosegmentation of Brain Metastases

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Radiotherapy planning for brain metastases is currently a manual and lengthy process.
• Automated segmentation using deep learning can potentially streamline this workflow.
• nnU-Net and MedNeXt are leading deep learning frameworks for medical image segmentation.

Purpose of the Study:

• To validate four pre-trained deep learning models (nnU-Net and MedNeXt) for brain metastases segmentation.
• To assess model performance on an external cohort of 243 patients.
• To compare the efficacy of nnU-Net versus MedNeXt frameworks for this task.

Main Methods:

• Utilized a dataset of 243 patients with pre-treatment T1-weighted contrast-enhanced 3D brain MRI scans.
• Employed four pre-trained deep learning models based on nnU-Net and MedNeXt architectures.
• Evaluated segmentation performance using established metrics.

Main Results:

• Models trained on combined datasets exhibited improved performance.
• nnU-Net based models showed a slight performance advantage over MedNeXt models.
• All validated models demonstrated potential for automated brain metastases segmentation.

Conclusions:

• Pre-trained nnU-Net and MedNeXt models are viable tools for automated brain metastases segmentation.
• nnU-Net models offer slightly better performance compared to MedNeXt for this specific application.
• Further validation with follow-up imaging is recommended to assess long-term performance and clinical utility.