Automatic cardiac MR image segmentation using edge detection by tissue classification in pixel neighborhoods

Abstract

A highly sensitive edge detector has been developed that uses tissue classification of pixels based on analysis of data in their local neighborhoods. In conjunction with recursive region growing, it has been used successfully to define regions of interest (ROI) when applied specifically to gradient echo MR images of the heart. The detector adapts to nonuniformity by carrying out an independent analysis at each location. If two tissues are present in a neighborhood and the pixel at that location cannot be classified with the seed pixel, a region edge has been crossed and recursion is stopped. No geometric assumptions relating to object shape such as definition of a region center and radial search are required. The detector was applied to multi-slice, multi-phase images of the heart from 26 subjects. A segmentation strategy specified slice processing order, graded ROIs, and used successfully detected ROIs to guide subsequent detection. Segmentation of all images resulted in a 90.3% median edge pixel detection efficiency.