Graduation Semester and Year
2017
Language
English
Document Type
Dissertation
Degree Name
Doctor of Philosophy in Electrical Engineering
Department
Electrical Engineering
First Advisor
Venkat Devarajan
Abstract
CT image based lung nodule detection is the most widely used and accepted method for detecting lung cancer. Most CT image based methods rely on supervised learning, which has a high number of false positives and need a large amount of pre-segmented training samples. These problems can be solved if an optimally small number of training sample images 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, two algorithms are proposed for 2D CT images and 3D CT images respectively. Both algorithms use size and shape characteristics of nodule candidates and, patch-based image segmentation. The proposed algorithm for 2D images has three 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 labeled by comparing the image patch, centered by the pixel, with patches from an atlas library. The most probable labels are then chosen according to a defined closest match criterion. c) a Laplacian of Gaussian blob detection method is then developed to find the segmented area of the lung nodule. The method is tested for more than 25 test slices, 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, the sensitivity of the proposed algorithm is 100% and no false positive was found. For 3D images, the algorithm is significantly changed. This algorithm has three steps: a) In the first step, nodule candidates of the current patient are detected and different features are extracted from each nodule b) The second step is the ‘atlas selection step’, in which two or three very similar lung images (atlas image) are selected from a group of atlases by a nodule-based atlas search process c) In the third and final step, the nodule based patch comparison process is developed to determine the accurate size and shape of the lung nodules. The proposed method has been proven accurate in recognizing all the non-attached nodules, which are bigger than 3 mm of radius, when applied on a population of twelve patient’s image datasets.
Keywords
Medical imaging, Atlas based segmentation, Pulmonary nodule detection, CT, Cancer detection
Disciplines
Electrical and Computer Engineering | Engineering
License
This work is licensed under a Creative Commons Attribution-NonCommercial-Share Alike 4.0 International License.
Recommended Citation
Alam, Mustafa, "Lung Nodule Detection and Segmentation Using a Patch-Based Multi-Atlas Method" (2017). Electrical Engineering Dissertations. 352.
https://mavmatrix.uta.edu/electricaleng_dissertations/352
Comments
Degree granted by The University of Texas at Arlington