Fully automatic segmentation of the brain from T1-weighted MRI using Bridge Burner algorithm

Purpose: To validate Bridge Burner, a new brain segmentation algorithm based on thresholding, connectivity, surface detection, and a new operator of constrained growing. Materials and Methods: T1 -weighted MR images were selected at random from three previous neuroimaging studies to represent a spectrum of system manufacturers, pulse sequences, subject ages, genders, and neurological conditions. The ground truth consisted of brain masks generated manually by a consensus of expert observers. All cases were segmented using a common set of parameters. Results: Bridge Burner segmentation errors were 3.4% +/- 1.3% (volume mismatch) and 0.34 +/- 0.17 mm (surface mismatch). The disagreement among experts was 3.8% +/- 2.0% (volume mismatch) and 0.48 +/- 0.49 mm (surface mismatch). The error obtained using the brain extraction tool (BET), a widely used brain segmentation program, was 8.3% +/- 9.1%. Bridge Burner brain masks are visually similar to the masks generated by human experts. Areas affected by signal intensity nonuniformity artifacts were occasionally undersegmented, and meninges and large sinuses were often falsely classified as the brain tissue. Segmentation of one MRI dataset takes seven seconds. Conclusion: The new fully automatic algorithm appears to provide accurate brain segmentation from high-resolution T1-weighted MR images.