Feature-based stereological volume estimation in sectional images

Volumetric measurement in digital slice images is a tedious process, if classification of pixels into background and object needs to be carried out manually. Using a stereological approach by classifying only a subset of sample locations decreases the amount of work but a large number of sample locations may still be required for achieving a sufficiently high accuracy. We combined the stereological volume estimation with a feature-guided component for carrying out a pre-classification. Slice-by-slice coherence and appearance features in the image are used for the pre-classification. The former accounts for structures that are found more by location rather than appearance in the image. Appearance features are the grey level and local roughness at object locations. We tested the algorithm using phantom images from MR as well as CT and MR patient images. Volumetric measurements for the phantom images agreed within 0.1 to 1.2% with the exact volume. Intra-observer variability, measured on the patient images, was less than 1.5%. The number of locations, that had to be re-classified, was between 13% and 33% of the total number of sample location, thus significantly decreasing the amount of user-interaction