چكيده به لاتين
According to the statistics of the world health organization, lung cancer is one of the most fatal diseases in the world that can be mostly treated if it be diagnosed on time. Till now, many methods for on time detection of this disease have applied computer aided revealing systems. Our main goal of the present research is also to develop a new framework for automated detection of pulmonary nodules through computational analysis of computed tomography (CT) images and then evaluating the level of changes in the size of the nodules in different time intervals. As the first step of the proposed method, all the layers of the multiband CT data are fed into a pre-processing step based on partial derivation equations in order to provide several smoothed layers which have reduced noises as well to form more homogenous segments. Then, optimal areas or the same lung two lobs in one step based on morphological processing have been extracted and entered to the Statistical Region Merging (SRM) algorithm so that their different parts are separated. As a result, we will have classified images or as called segment-maps In as many as the whole number of layers of the main CT images. After that, similar labels are dedicated to the existing segments in the layers adjacent to each other, in case they have overlap with each other more than one limit of pre-determined threshold. The present borders in the segments of the layers adjacent to each other with similar labels connect to each other and through this; three-dimensional objects will be obtained as the nodule candidates. Of all these objects, four spectral features, one morphological, and one textural feature are extracted and are fed as two-class classifier based on Support Vector Machine (SVM) with polynomial kernel. The result of this classification identifies nodules and their sizes. In the next step, the size of the nodules is evaluated over time intervals and compared with radiologists' results.
The proposed method in this dissertation has been applied to several standard sets of lung CT images drawn from valid and international databases. Finally, the obtained results have been analyzed using 5 quantitative measures and compared with many other methods existing in this field. The most important of these criteria to compare with other methods are the accuracy of nodule detection, F score and also the AUC, which our proposed method using 800 images of LIDC-IDRI database, has the highest amount of these criteria in comparison to other competitors methods. These values are equal to 97.8%, 0.982 and 0.991, respectively. The comparison shows that the proposed method has better performance in detection and extraction of pulmonary nodules. But since other methods have not mentioned the duration of their processing steps, our method time criteria in this study is not compared with other methods. Also, the above criteria are significant values for 50 CT images from theELCAP database but are not compared with other methods due to the lack of data compliance. The accuracy, F-score and the AUC of the proposed method for 50 images of ELCAP database are wqual to 95%, 0.934 and 0.978, respectively.
The considerable point in comparison to other competing methods is that the proposed method is automated in the sense that most of the parameters are predefined as constant values or adaptively tuned according to the characteristics of each CT data, if possible.
Keywords : Pulmonary Nodule, Computer Aided Detection, CT X-ray images, Nonlinear Partial Differential Equations, Statistical Region Merging