• Computer Aided Detection
• lung nodules
• SVM

Lung cancer is one of the main public health issues in developed countries. The overall 5-year survival rate is only 10−16%, although the mortality rate among men in the United States has started to decrease by about 1.5% per year since 1991 and a similar trend for the male population has been observed in most European countries.
By contrast, in the case of the female population, the survival rate is still decreasing, despite a decline in the mortality of young women has been ob- served over the last decade.
Approximately 70% of lung cancers are diagnosed at too advanced stages for the treatments to be effective. The five-year survival rate for early-stage lung cancers (stage I), which can reach 70%, is sensibly higher than for cancers diagnosed at more advanced stages.
Lung cancer most commonly manifests itself as non-calcified pulmonary nodules. The CT has been shown as the most sensitive imaging modality for the detection of small pulmonary nodules, particularly since the introduction of the multi-detector-row and helical CT technologies. Screening programs based on Low Dose Computed Tomography (LDCT) may be regarded as a promising technique for detecting small, early-stage lung cancers. The efficacy of screening programs based on CT in reducing the mortality rate for lung cancer has not been fully demonstrated yet, and different and opposing opinions are being pointed out on this topic by many experts.
However, the recent results obtained by the National Lung Screening Trial (NLST), involving 53454 high risk patients, show a 20% reduction of mortality when the screening program was carried out with the helical CT, rather than with a conventional chest X-ray.
LDCT settings are currently recommended by the screening trial protocols. However, it is not trivial in this case to identify small pulmonary nodules,due to the noisier appearance of the images in low-dose CT with respect to the standard-dose CT. Moreover, thin slices are generally used in screening programs, thus originating datasets of about 300 − 400 slices per study. De- pending on the screening trial protocol they joined, radiologists can be asked to identify even very small lung nodules, which is a very difficult and time- consuming task. Lung nodules are rather spherical objects, characterized by very low CT values and/or low contrast. Nodules may have CT values in the same range of those of blood vessels, airway walls, pleura and may be strongly connected to them. It has been demonstrated, that a large percent- age of nodules (20 − 35%) is actually missed in screening diagnoses.
To support radiologists in the identification of early-stage pathological objects, about one decade ago, researchers started to develop CAD methods to be applied to CT examinations.
Within this framework, two CAD sub-systems are proposed: CAD for internal nodules (CADI), devoted to the identification of small nodules embedded in the lung parenchyma, i.e. Internal Nodules (INs) and CADJP, devoted the identification of nodules originating on the pleura surface, i.e. Juxta-Pleural Nodules (JPNs) respectively.
As the training and validation sets may drastically influence the performance of a CAD system, the presented approaches have been trained, developed and tested on different datasets of CT scans (Lung Image Database Consortium (LIDC), ITALUNG − CT) and finally blindly validated on the ANODE09 dataset.
The two CAD sub-systems are implemented in the ITK framework, an open source C++ framework for segmentation and registration of medical im- ages, and the rendering of the obtained results are achieved using VTK, a freely available software system for 3D computer graphics, image processing and visualization. The Support Vector Machines (SVMs) are implemented in SVMLight. The two proposed approaches have been developed to detect solid nodules, since the number of Ground Glass Opacity (GGO) contained in the available datasets has been considered too low.
This thesis is structured as follows: in the first chapter the basic concepts about CT and lung anatomy are explained. The second chapter deals with CAD systems and their evaluation methods. In the third chapter the datasets used for this work are described. In chapter 4 the lung segmentation algorithm is explained in details, and in chapter 5 and 6 the algorithms to detect internal and juxta-pleural candidates are discussed. In chapter 7 the reduction of false positives findings is explained. In chapter 8 results of the train and validation sessions are shown. Finally in the last chapter the conclusions are drawn.