Organ localization is an important step in medical applications such as automated image analysis, segmentation, registration, smarter workflow designs, or data mining. In this paper, we present a generic framework for automated localization of anatomical structures and organs in CT and MR image data. In the learning phase, we identify unique signatures for organs of interest from training data. Organ localization is done in three steps: body region identification, organ size estimation, and signature matching for localization. These steps enable our algorithm to be robust across patient demographics, profiles and medical conditions. Our technique does not make a priori assumptions about presence or absence of any organ or supporting structure in supplied data. We propose a cascading scheme consisting of Gabor filtering followed by Speeded-Up Robust Features (SURF) for identification of reliable interest points, and show that our objective function has stronger local minima when compared with SURF, SIFT and GIST-based methods.We find point-based correspondences for an input image with exemplar images for body region identification. We have also introduced a computationally efficient way to build histogram of 3D Uniform Local Gabor Binary Patterns using fuzzy approximations. We have used our algorithm for retrieving head, neck, liver, heart, kidneys, spleen and lungs regions from a database of 60 non-contrast CT and 80 T1-weighted MR images. The performance is assessed quantitatively on all three stages using ground-truth database sanitized by medical experts. The average error in body region estimation was 5.76mm. In organ size estimation, the average error was 8.32% of the organ size. Finally, organs were localized correctly 97.14% of the times, within an error margin of 20mm.
展开▼
