Preoperative Brain Tumor Imaging: Models and Software for Segmentation and Standardized Reporting.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate brain tumor characterization is crucial for patient prognosis and treatment decisions.
• Current methods for tumor detection and reporting lack standardization and automation.
• This study addresses the need for reliable tools in neuro-oncology.

Purpose of the Study:

• To develop and validate automated tumor segmentation models for common brain tumor types.
• To create user-friendly software for standardized clinical report generation.
• To improve the efficiency and consistency of preoperative MRI analysis.

Main Methods:

• Trained tumor segmentation models using the AGU-Net architecture on four cohorts (up to 4,000 patients).
• Evaluated segmentation performance using comprehensive voxel-wise and patient-wise metrics (Dice, volume difference, etc.).
• Developed Raidionics and Raidionics-Slicer software for model deployment and report generation.

Main Results:

• Achieved high segmentation performance across glioblastomas, lower grade gliomas, meningiomas, and metastases (Dice 80-90%).
• Demonstrated homogeneous performance across tumor types with patient-wise recall of 88-98% and precision around 95%.
• Raidionics software enables rapid tumor segmentation (16-54s) and report generation (5-15 min).

Conclusions:

• The developed automated segmentation models and software provide a standardized and efficient solution for brain tumor analysis.
• Open-access models and code facilitate wider adoption and further research in neuro-oncology.
• Future work includes developing a single performance score and automatic tumor classification.