Characterization of a sequential pipeline approach to automatic tissue segmentation from brain MR Images

Abstract

Objective Quantitative analysis of gray matter and white matter in brain magnetic resonance imaging (MRI) is valuable for neuroradiology and clinical practice. Submission of large collections of MRI scans to pipeline processing is increasingly important. We characterized this process and suggest several improvements. Materials and methods To investigate tissue segmentation from brain MR images through a sequential approach, a pipeline that consecutively executes denoising, skull/scalp removal, intensity inhomogeneity correction and intensitybased classification was developed. The denoising phase employs a 3D-extension of the Bayes-Shrink method. The inhomogeneity is corrected by an improvement of the Dawant et al.'s method with automatic generation of reference points. The N3 method has also been evaluated. Subsequently the brain tissue is segmented into cerebrospinal fluid, gray matter and white matter by a generalized Otsu thresholding technique. Intensive comparisons with other sequential or iterative methods have been carried out using simulated and real images. Results The sequential approach with judicious selection on the algorithm selection in each stage is not only advantageous in speed, but also can attain at least as accurate segmentation as iterative methods under a variety of noise or inhomogeneity levels.Conclusion A sequential approach to tissue segmentation, which consecutively executes the wavelet shrinkage denoising, scalp/skull removal, inhomogeneity correction and intensity-based classification was developed to automatically segment the brain tissue into CSF, GM and WM from brain MR images. This approach is advantageous in several common applications, compared with other pipeline methods.

IJCARS special issue: BVM 2007 German conference on medical image processing

“…Therefore, the contributions in this issue can reflect only some of them. Nevertheless, a number of core areas are addressed: acquisition [1], registration and fusion [3,5], visualization and simulation [6], validation [4], and segmentation [2]. Primary application areas for the new or enhanced methods range from general improvements for an entire class of imaging devices or tasks [1,3], to the support of either the diagnosis and treatment planning [2,6], or the intervention [4,5], for different medical specialties (vascular diseases, abdominal and cardiac surgery, neurology and neurosurgery).…”

“…Continuous progress has been made toward sophisticated, more robust and (semi-) automatic segmentation methods. In Su Huang et al [2] a novel automatic tissue segmentation method for brain magnetic resonance images is presented.…”