Segmentation via image structure (in medical imaging)

The author describes a segmentation technique which satisfies the criteria for good segmentations in that it (i) can produce T-junctions, (ii) each section of edge has an associated confidence and (iii) requires no user-intervention or domain-dependent constants. Initial results with the technique appear to indicate that the segmentations produced compare well with the expectations of an inexpert human. The detection of edges is good and even if an edge has not been constructed, it is usually present as a rejected ridge (uphill MGP). The localisation of the constructed edges is excellent, supporting the hypothesis that the differential uphill MGPs form a superset of the perceptual edges. If the calculation of ridge and trough confidences can be improved one could be able to produce segmentations which are satisfactory to an inexpert human.<
>