New methods for leak detection and contour correction in seeded region growing segmentation

Abstract

Segmentation, i.e. the labelling of objects in image data, is a crucial step in many medical imaging processing tasks, e.g. operation planning, radio therapy or diagnostics. Seeded region growing is a basic yet effective method for semi-automatic segmentation. Its major drawback is that poor contrast at the organ edges may result in leaks, letting the area grow far beyond the region of interest. We developed a new method to reliably detect the origins of these leak regions and correct the segmentation results. Our approach is based on the observation that leaks are normally characterized by a narrow bottleneck connection to the seed area. We detect this bottleneck by specifying a single point somewhere within the erroneous area and tracing a path back to the seed point, moving along the skeleton of the segmented region. The point on the skeleton with the minimal distance to the contour marks the bottleneck. To cope with minor deviations from the desired result (not featuring a characteristic bottleneck), we optimized our freehand tool to minimize the necessary user interaction. Instead of separately cutting or merging parts of the segmentation, the new tool allows modifications by replacing entire parts of the contour in a single step.