Abstarct: In this thesis, we purpose a system for automatic detection of cancer lesions so called nodule and segmenting them in thoracic CT images, which is very important in early detection of lung cancer. This system consists of three main stages. In the First noise cancellation, 3D interpolation of images and lung segmentation is performed; we refer to this stage as pre-processing. In the second stage candidate for nodules are detected and in the last stage from the candidate points lesions are segmented from lung tissue. In pre-processing stage, linear interpolation is performed to equalize the resolution of 3D image in all dimensions. In this stage with take full advantage of intensity histogram in each 2D slice and information in adjacent slices, Segmentation of lung parenchyma is done by adapting an appropriate threshold on histogram of each slice. The result of this process is promising, especially for juxtapleural nodules. Lung nodules are always appeared in different size, sphere-like in shape and they are very similar to blood vessels in intensity feature. So in nodule candidate detection stage, not only intensity feature, but also the geometrical shape features are used. Most of methods baxsed on shape detection are used in 3D space, which are time-consuming in this application because lots of information to be processed. Hence, a new method baxsed on 2D data analysis in three orthogonal directions is purposed which reduce the time of process and use volumetric information of images simultaneously. In this step to detect the geometrical shape of regions, after the smoothing of input images, we employed some filters such as shape index, convergence index and spherical enhancement filter, which use shape information e.g. gradient of intensity, Eigen values of hessian matrix, Gaussian and mean curvatures. Finally a feature space analysis is used to segment the nodule's content. Location vector, intensity and shape features which are taken from the last step are concatenated to form the 7-dimentional feature space. By using a non-parametric method, namely mean-shift procedure, data points are clustered in this space. It's critical to set the proper bandwidth parameter to use of this method which in this report a new bandwidth chosen method was presented. The method's capability to detect all of nodule types, low false positive rate, accuracy and fast computation shows much promise for clinical applications.