Enhanced thresholding

There are many techniques for improving the ROC performance of threshold-based segmentation:

• Seeding and region growing. Simple thresholding makes no use of the geometric connectivity information present in an image. Using a seeding and region-growing technique, a segmented pixel set is created by initially selecting one or more pixels from the image (called the seed pixels). The seeds will usually correspond to some anatomical region or landmark and are often specified interactively by the user. The region-growing algorithm will then add to the segmented pixel set all the pixels that are r-connected to the initial seed pixels and also fall within the threshold limits. To be r-connected to one another, two pixels must share at least r corner points ( ). The algorithm recursively adds to the segmented pixel set all the pixels that are connected to the current members of the pixel set. The process stops when no more pixels can be added to the segmented set. This method tends to reduce the total number of false positives because only those pixels that can be r-connected to the original seed are classified. Isolated pockets of pixels that are not part of the object being classified but fall within the range are eliminated.
• Hysteresis. In this technique, a different threshold rule is used to grow a region than is used to select the seeds. Typically, the growing threshold is set to a lower value than the seed-selection threshold. This technique tends to produce segmented pixel sets with more clearly defined borders. Because the seed pixels have a higher threshold value, the seeds are less likely to correspond to false positive regions. This also tends to reduce the number of isolated pockets of false positive regions and produces sharper object borders.
• Subdivision. Rather than use a single value throughout the image, the threshold can be made into a spatial function of the region of interests [18]. This implies that the region of interest can be viewed with a standard orientation and scale.
• Pattern recognition and texture analysis. Many tissues, such as bone marrow, have a regular pattern or texture that can be detected using techniques such as statistical clustering [29]. Often, regions can be grouped together according to image intensity values, and measurements regarding the distribution and concentration of various groups can be used to classify tissue types.
• Spectral analysis. This technique is useful when images can be acquired in different bands (for example, color or temperature components), with multiple sensors, or with multiple imaging parameters (for example, T1- and T2-weighted MR images) [60].