Computerized detection and diagnosis of lung lesions using CT

The CAD system described in this thesis is able to automatically detect nodules in lung CT images and to automatically perform the diagnosis of such nodules. The system has the following novel characteristics: -it performs the automatic diagnosis without requiring any intervention by the radiologist in any of the phases it consists of; -it does not emulate a single radiologist, but a team of radiologists. In particular, the first characteristic means that the system can replace the radiologist both for detecting the presence of nodules (by distinguishing them from the other anatomical structures), and classifying them into benign and malignant. The second characteristic, on the other hand, means that the CAD system is made of a set of CAD subsystems independent of each other, each able to perform a diagnosis. More precisely, we have considered, within the radiologist’s activity, three distinct and subsequent phases: the search for ROIs, their classification into nodules and non-nodules, and nodule diagnosis. At the start we developed our project by implementing different techniques to perform the three phases of CAD. Each technique, which can work “stand-alone", has been tested on a significant set of CT scans. Subsequently we built a system that takes into consideration all the set of techniques to obtain a more robust output for each phase. Our system is composed of three modules, one for each phase and each module contains submodules to implement the techniques.