Jaco Hagoort, Karim Salam & Jan M. Ruijter

Automatic segmentation of optical density images

Attached document(s)

Annexes

• AS_18_n2_265_HAG.pdf (3.2M)

Abstract

Measurement of optical densities in different parts of an organ or tissue requires an objective segmentation of these different morphological structures in images of histochemically stained sections. However, in biological tissues, the relevant morphological structures often contain a wide range of pixel values which hampers the use of segmentation procedures based on thresholds, on region-growing or on a-priory knowledge of the shape of, the segments. Segmentation based on multivariate statistics requires customised programs or dedicated hardware. Therefore, an automated segmentation procedure based on statistical criteria has been developed that could be used in the context of a readily available image processing package. The procedure is based on minimising the pixel value variation within each segment while maintaining the spatial continuity of the different segments in the image. In this procedure the original image is thresholded to make it binary and then subjected to binary operations (erosion and dilatation) to connect the spatial noise. The original image is then masked with the resulting binary image a11d its inverted complement. Of both partial images the variation in pixel values is measured and the pooled variation is calculated. The relative decrease of this pooled variation compared to the variation in the original image is used to decide whether or not the segmentation of the original image is considered relevant. If so, the image is split and the procedure is applied recursively on both resulting images until the reduction in variation no longer justifies a further segmentation. Reproducibility of the developed segmentation procedure was tested by applying it to pairs of normal and inverted images and to images with a compressed pixel value histogram. The proposed segmentation algorithm can be used for the segmentation of a known organ or tissue into functional zones, using a combination of fast image processing functions and statistical decisions.

To cite this article

About: Jaco Hagoort

About: Karim Salam

About: Jan M. Ruijter