Automatic metastatic brain tumor segmentation for stereotactic radiosurgery applications.

Area of Science:

• Medical Imaging
• Computational Biology
• Radiotherapy

Background:

• Accurate delineation of metastatic brain tumors on contrast-enhanced T1-weighted (T1c) MRI is essential for stereotactic radiosurgery (SRS).
• Manual segmentation is time-consuming and prone to inter-observer variability.
• Developing an automated, efficient, and accurate segmentation strategy is a significant clinical need.

Purpose of the Study:

• To develop and validate an automatic segmentation strategy for metastatic brain tumors on T1c MRI.
• To improve the efficiency and accuracy of tumor delineation for SRS applications.
• To establish a robust method that can be integrated into clinical workflows.

Main Methods:

• A four-step automatic segmentation strategy: pre-processing (skull removal), initial contouring (multi-scaled adaptive thresholding and super-voxel clustering), contour evolution (regional active contour), and contour triage (geometric characterization).
• Validation using a virtual phantom with varying noise levels (Gaussian, Rician) and real-world datasets (BRATS challenge, clinical cases).
• Performance evaluation using metrics like Dice coefficient (DC), Normalized Mutual Information (NMI), Structural Similarity Index Measure (SSIM), and Hausdorff Distance (HD).

Main Results:

• On numerical phantoms, the strategy achieved a Dice coefficient (DC) of 0.98 ± 0.01 and Hausdorff distance (HD) of 2.2 ± 0.8 mm.
• Validation on the BRATS dataset yielded a DC of 0.89 ± 0.08, outperforming challenge algorithms.
• Clinical dataset evaluation showed a DC of 0.86 ± 0.09 and HD of 8.8 ± 12.6 mm compared to physician-drawn contours.

Conclusions:

• The developed automatic segmentation strategy provides accurate and efficient metastatic brain tumor delineation.
• The method demonstrates superior performance compared to existing algorithms on benchmark datasets.
• This automated approach shows significant potential as a valuable clinical tool for SRS applications.