Adaptive Liver Segmentation from Multi-slice CT Scans

In this paper, an adaptive method was proposed to segment the liver from computed tomography (CT) images. Four sets of CT images were segmented by a binary classification method, support vector machine (SVM), after supervised thresholding and K means clus

  • PDF / 268,022 Bytes
  • 4 Pages / 595 x 791 pts Page_size
  • 60 Downloads / 199 Views

DOWNLOAD

REPORT


INTRODUCTION The imaging techniques of multi-slice computed tomography (CT) featuring high spatial resolution and fast scan speed provide a great clinical diagnostic tool for examining liver pathologies such as cirrhosis, liver cancer and fulminant hepatic failure. In order to extract the anatomical and pathological information of the liver, the large amount of data yielded from CT scanners requires a post image processing which is both laborious and time-consuming by manual process and visual inspection. Digital image processing and machine learning techniques may provide semiautomatic and automatic methods of liver segmentation for automatic detection of liver cancer or other hepatic diseases as well as the measurement of the liver volume for living donor liver transplantation or 3D volume rendering for computer aided surgical planning prior to hepatic resection. In literature, live-wire segmentation approaches, gray level based liver segmentation, model fitting, probabilistic atlases and level set approaches are the most commonly used techniques for liver segmentation. However, automatic liver segmentation remains unsolved and open for further investigation. Indeed the liver shape model based methods can not meet the need for describing complex liver shape while other methods suffer from the incorrect separation of the liver from neighboring organs such as muscles, spleen and stomach, due to the overlap in gray level ranges and the inter-subject variation in position and shape across the soft tissues in the abdomens. Recent studies of liver segmentation using texture analysis have shown that extraction of the texture features from CT images is a promising method for solving this challenging segmentation task by assuming

homogeneity and consistency of the texture information across multiple slices of a single organ. Among the normal texture analysis methods including Haralick co-occurrence matrices, fractal dimension, Gabor filtering and Markov random fields, the agreement of demonstrating Haralick cooccurrence matrices as the most significant texture characteristics in differentiating the liver tissue in CT scans is achieved across different independent studies in literature using texture analysis in liver segmentation. Compared to the gray level based method (first-order properties of an image), Haralick co-occurrence method gives second-order properties of images to describe the homogeneous anatomical structure of the liver allowing tissue separation between the liver and its neighboring organs at the areas where gray level ranges overlap [1]. After texture feature extraction, common image segmentation techniques, such as clustering techniques, mathematical morphology operation, and region growing, or machine learning methods, such as neural networks, SVM, and fuzzy rules are applied to separate the organs in CT images. Though unsupervised techniques can offer faster segmentation than the supervised techniques, the accuracy of these methods is still poor as the parameters and settings of these methods may vary