Lung Nodule Detection and Segmentation Using a Patch-Based Multi-Atlas Method

CT image based lung nodule detection is the most widely used and accepted method for detecting lung cancer. Most CT image based methods are based on supervised/unsupervised learning, which has a high number of false positives and needs a large amount pre-segmented training samples. This problem can be solved, if a set of optimally small number of training samples can be created, where each sample has lung nodules of similar size and shape as the target image of the actual patient. Based on this hypothesis, we propose a novel patch-based multi-atlas method with three main steps: a) a small set of atlases is selected by comparing the target image with a larger set of atlas images using a size-shape based feature vector, b) lung nodules are selected using a patch-based method, where each pixel of a target image is labelled by comparing the image patch, centered by the pixel with patches from an atlas library and choosing the most probable labels according to a defined closest match criterion and c) Laplacian of Gaussian blob detection method is used to find the segmented area of the lung nodule. We tested the method with more than 5 test images, where each test image is applied to more than 200 atlas images. For non-attached nodules in the size between 3 mm to 30 mm, sensitivity of the proposed algorithm is 100%.