Hough transforms for shape identification and applications in medical image processing

An iterative randomized Hough transform (IRHT) and a direct inverse randomized Hough transform (DIRHT) are developed for detection of incomplete curves in images with strong noise. The IRHT iteratively applies the randomized Hough transform to a region of interest in the image space. By iterative parameter adjustments and reciprocal use of the image space and parameter space, the IRHT progressively reduces the noise interference. The DIRHT combines the advantages of both the inverse and the randomized Hough transforms, and demonstrates a “curve-pass filtering” effect: curves are highlighted and curve pixels connected while a large amount of unrelated pixels can be removed. Both the IRHT and the DIRHT do not rely on specific geometric properties and other a priori information about the curve. Furthermore, a K-Mean clustering-based method is developed for object segmentation in ultrasound images. These methods show robustness and effectiveness for ellipse identification in synthesized images and fetal head identification in medical ultrasound images. The IRHT can reliably and efficiently detect pupils in infra-red mouse eye images for pupillary light reflex analysis. The results demonstrate great potential of these methods for object identification in real-world applications.