You are here
A Combined Approach for Lung Boundary Segmentation of Chest X-Ray Images.
In collaboration with the Indiana University School of Medicine, AMPATH, the largest AIDS treatment program in the third world), and Moi University in Eldoret, the National Library of Medicine is developing a digital chest x-ray screening software with a focus on detecting tuberculosis. The software is intended for deployment on low-cost transport vehicles in remote areas in Kenya to help screen populations using mobile x-ray scanners. Our software consists of various image-processing algorithms. In this poster we are presenting a lung boundary detection method using a combination of techniques based on graph-cut and elastic registration to achieve state-of-the-art performance. More specifically, the method consists of a fast image retrieval method for identifying training images (with masks) most similar to the patient CXR using Bhattacharyya shape similarity measure and image correlation, creating the initial patient-specific anatomical atlas using elastic registration using cubic B-splines, building refined lung boundaries using a graph cut optimization approach with a customized energy function. We validated the accuracy of the approach with extensive experiments on public and NLM x-ray datasets. The average accuracy of the system is above 95% on the public dataset (Jaccard similarity coefficient) and comparable with the state-of-the-art.