Automated Brain Metastases Segmentation With a Deep Dive Into False-positive Detection.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Clinical management of brain metastases post-stereotactic radiosurgery (SRS) is challenging due to labor-intensive review of magnetic resonance imaging (MRI) scans.
• Accurate contouring of metastases is crucial for effective SRS treatment planning and outcome assessment.

Purpose of the Study:

• To develop and evaluate an automated framework for contouring brain metastases on MRI scans.
• To assist in treatment planning for SRS and understand the limitations of automated contouring.

Main Methods:

• Utilized two self-adaptive nnU-Net models trained on 3D T1-weighted postcontrast MRI scans.
• Evaluated performance using positive predictive value (PPV), sensitivity, and Dice similarity coefficient (DSC) on a dataset of 845 patients (3482 metastases) for training and 206 patients (930 metastases) for testing.

Main Results:

• Achieved high performance with per-patient PPV of 90.1%, sensitivity of 88.4%, and DSC of 82.2%.
• For large metastases (≥6 mm), performance improved with PPV of 95.6%, sensitivity of 94.5%, and DSC of 86.8%.
• Physician assessment indicated 75% of contours were clinically acceptable as-is, with the remainder requiring minor edits.

Conclusions:

• The automated framework effectively detects and contours brain metastases on MRI, supporting radiologists and radiation oncologists.
• The tool aids in precise decision-making for suspicious lesions and shows potential for early lesion detection.
• The findings suggest the framework can streamline clinical workflows and improve patient management after SRS.