Please use this identifier to cite or link to this item:
|Title:||Multiorgan detection and segmentation strategies for 2D ultrasound images of the thyroid gland||Authors:||Subbarao, Nikhil Narayan||Keywords:||DRNTU::Engineering||Issue Date:||2016||Source:||Subbarao, N. N. (2016). Multiorgan detection and segmentation strategies for 2D ultrasound images of the thyroid gland. Doctoral thesis, Nanyang Technological University, Singapore.||Abstract:||The thyroid gland is an endocrine gland that is responsible for iodine metabolism and hormone regulation in the human body. The thyroid gland is susceptible to disorders such as Goitre, Hashimotos Thyroiditis, cancer, etc., which often present with lumps in the neck. Any physician who examines the gland, first palpates the region where the lump is present and orders for a follow up investigation to make a diagnosis. This is done with the help of imaging modalities such as ultrasound, computed tomography (CT) and magnetic resonance imaging (MRI). The choice of imaging modality is dependent on the severity of the condition prevalent with the disorder. Imaging with CT or MRI modalities are recommended when the treatment of the disorder involves surgically operating the gland. But in a clinical setting, which is where a majority of patients are treated, ultrasound is the gold standard imaging modality used to screen the thyroid gland for disorders. The aim of this research is to devise automatic landmark based multiorgan detection and segmentation algorithms for freehand 2D ultrasound images of the thyroid gland. Three new methods are proposed to automatically detect and segment multiple organs in the US images of the thyroid gland. The organs that are considered for this research are: (a) the thyroid gland; (b) the carotid artery; (c) the trachea; and (d) the muscles. All of our methods make use of speckles and imaging artefacts in the image, which are otherwise considered to be noise, as sources of information to detect and segment the organs in ultrasound images. The speckle related pixels are determined by the application of Hessian based blob detectors on the ultrasound image. We demonstrate the application of our methods to perform guided interventions, volumetric analysis and computer aided diagnosis. In the first of the three methods, the speckle related pixels are clustered into three echogenic classes based on its relative brightness in a kernel. An energy based model is then used to detect the carotid artery in the labelled image (image whose pixels are labelled with the cluster number). The carotid artery is used as a landmark to detect and segment the remaining organs in the image. In the second method, an agglomerative clustering constrained with a similarity metric is proposed to cluster the speckle related pixels into an unknown number of echogenic classes. The total number of class labels generated at the end of the clustering determines the number of echogenic classes in the image. This process is viewed as a quantization of the speckle related pixels into levels that are determined by the tissue echogenicity. The enhancement artefact is then detected in the quantized image which is in turn used as a landmark to detect the carotid artery. The carotid artery is used as a landmark to detect the remaining organs. All organs are segmented by the application of region based local phase methods. In the last algorithm, the speckle related pixels are classified into their respective echogenic classes by the use of local phase based methods. The carotid artery and the trachea are detected from the binary image containing the hypoechoic pixels as foreground pixels. The pixels of the carotid artery and trachea along with the pixels that lie in the region between them are used to train a random forests classifier to classify the remaining pixels that belong to the hyperechoic tissue into the thyroid gland and background pixels. The proposed methods are compared with each other and with state-of-the-art methods to prove the efficacy of the algorithms in performing medical image analysis. Manual segmentations obtained from two trained sonographers are used as ground truth to validate the algorithms. In all, 1091 images from five databases of annotated images are used in our experiments. The images are pre-processed by a novel annotation removal and image restoration algorithm that restores images to a high quality (average PSNR >38dB). All of our methods yield good quality segmentation results with Probabilistic Rand Index (PRI) > 0.83 and Boundary Error (BE) < 1mm. Analysis of the results using the Dice co-efficient as the metric shows that the proposed methods performs better than the state-of-the-art methods.||URI:||http://hdl.handle.net/10356/69070||Fulltext Permission:||open||Fulltext Availability:||With Fulltext|
|Appears in Collections:||EEE Theses|
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.