Metodi di classificazione automatica per l'analisi di candidati noduli in CT polmonari

The Lung Cancer is one of the so-called ¿big killers¿. The mortality increased in the past years; for that reason many screening programs were started in the world, with the purpose of showing that early diagnosis can help in reducing the mortality. Computer Aided Detection systems (CAD) operating on digital images provided by devices such as Computer Tomography (CT) are very useful to help the radiologist in the investigation and discovery of cancer nodules. These tools can be used for screening activities and/ or for specific patient study. The currently available CADs show some limitations due mainly to: the detection of a great number of False Positive (FP), a weak True Positive (TP) recognizing capability and low operational efficiency principally due to man-machine interface and weak generalization capability. Therefore it is required to investigate and work on these areas to improve CAD performances. This thesis analyzes and modifies the CAD for Lung Cancer Nodules recognition (handling 3D images), already developed as a part of the INFN (National Institute for Nuclear Physics - Italy) MAGIC-5 project (Medical Applications on a Grid Infrastructure Connection), in order to: evaluate and increase performances: discovery possible causes of (TPs) findings loss and implement some solutions, reduce false positives (FPs) detection, suggest and develop and test modifications to improve the recognition capability. The CAD analysis entails the study of the standard techniques for digital image processing such as: noise reduction, segmentation, region of interest (ROI) definition, filtering, feature extraction, classification and statistical methods to evaluate numerical results (ROC, FROC curves). Also the standard Digital Imaging and Communication in Medicine (DICOM) will be analyzed to develop application module operating on 3D imaging units called voxels. Moreover to evaluate CAD performances a preliminary study of: the virtual ant colonies based segmentation algorithm (for non linear system) and the ANN (Artificial Neural Network) based classification method will have to be carried out in depth. The end user application OSIRIX shall be used to display and analyze DICOM medical imaging data, before and after the CAD processing. Several public databases, related to lung CT images, will be used in this work: LIDC ( Lung Image Database Consortium), ITALUNG-CT and ANODE09. Specific software modules (for CAD modifications) will be developed using the C++ programming language and the CERN ROOT C++ library for data analysis. The current release of CAD software will be analyzed, partially modified and upgraded. The present document is composed by the following main sections: Lung Analysis and Medical Imaging (section 2 and 3); CAD System Processing (section 4); Channeler Ant Model (CAM) based CAD description (section 5); CAM ¿ CAD Improvements and Results (section 6). The last section (7) describes Conclusion and suggestions for possible future works.