Active contour models - a multiscale implementation for anatomical feature delineation in cervical images

In the work presented here, digital colposcopic images from cervix uteri are subjected to a segmentation algorithm using active contour models at multiresolution levels, on images formed by a Gaussian pyramid (GP). The segmentation aims to outline a specific feature from the cervical images $the transformation zone (TZ), where a possible neoplasia can occur. The process includes an implementation of a boundary-searching snake, based on both image gradient features and region features. The adaptive "snake" initially is executed on the image with the lowest resolution from the GP, aiming to avoid a specific artefact in these images, known as specular reflection (SR). Further, the snake coordinates are propagated to the image with highest resolution from the GP. Under a specific set of parameters, we achieved accurate delineation of the TZ, measured by comparison with pre-defined "ground truth".