Automated vessel segmentation using cross-correlation and pooled covariance matrix analysis

Towards the automatic computational assessment of enlarged perivascular spaces on brain magnetic resonance images: A systematic review

“…The intensity analysis we conducted on our sample, along with several methods included by this review showed that the intensity of EPVS in T2W images, although it overlaps with that of WMH, is generally lower than that of the lacunes, and supports the idea that masking out areas of similar intensity levels (e.g., removing CSF, WMH) may improve specificity of the results in EPVS identification and segmentation. However, when dilated, EPVS have intensity levels similar to those of CSF.…”

“…The literature search identified eight publications of six studies that assessed EPVS computationally, either directly or indirectly (Table ) by seven different methods. In addition, we found four studies that described techniques to automatically segment similar structures that are potentially suitable for EPVS segmentation (Table ).…”

“…Our search also identified four computational methods aimed at segmenting blood vessels and fiber tracts that could potentially be applied to identify and extract EPVS (Table ). Two of them are based on shape identification and the other 2 on intensity differences . The shape‐based identification methods considered tubular structures: one used a directional identification of tubular objects to generate shape priors for the segmentation and the other considered tubular structures with a center‐line in the three‐dimensional (3D) space and identified this center‐line from a seed point obtained automatically using an optimized energy function .…”

“…The shape‐based identification methods considered tubular structures: one used a directional identification of tubular objects to generate shape priors for the segmentation and the other considered tubular structures with a center‐line in the three‐dimensional (3D) space and identified this center‐line from a seed point obtained automatically using an optimized energy function . The intensity‐based methods used thresholding: one generated regions‐of‐interests (ROIs) for each feature (i.e., vessel) by iterative thresholding on multiple equal‐sized regions to automatically separate the vessels using the Mahalanobis distance , and the other used a single threshold to semi‐automatically delineate the boundaries of lacunar infarcts . This last approach combined the segmentation by thresholding with a shape analysis: considered the surface of each segmented object (i.e., lacune) as a mesh of triangles, linked together the centers of the voxel facets to obtain a high‐resolution mesh, and finally smoothed the mesh using an optimal decimation algorithm that reduced the number of triangles.…”

“…Although all segmentation methods used thresholding as a discriminator, 8 of 11 computational methods used it as a primary technique: combined with a region‐growing algorithm , in an adaptive thresholding model , combined with interconnected components' information , iteratively , on a scheme that considered intensity levels of known adjacent normal‐appearing white matter areas and on a “top‐hat” gray level thresholding transformation to enhance the conspicuity of EPVS and vessels for further discrimination from other features by a back‐propagated neural network. One study that identified EPVS for exclusion combined thresholding on coregistered T2W/PD/FLAIR sequences with location information to discriminate them from small subcortical hyperintensities and other considered size .…”