Development and application in clinical practice of Computer-aided Diagnosis systems for the early detection of lung cancer

Abstract

Lung cancer is the main cause of cancer-related deaths both in Europe and United States, because often it is diagnosed at late stages of the disease, when the survival rate is very low if compared to first asymptomatic stage. Lung cancer screening using annual low-dose Computed Tomography (CT) reduces lung cancer 5-year mortality by about 20% in comparison to annual screening with chest radiography. However, the detection of pulmonary nodules in low-dose chest CT scans is a very difficult task for radiologists, because of the large number (300/500) of slices to be analyzed. In order to support radiologists, researchers have developed Computer aided Detection (CAD) algorithms for the automated detection of pulmonary nodules in chest CT scans. Despite proved benefits of those systems on the radiologists detection sensitivity, the usage of CADs in clinical practice has not spread yet. The main objective of this thesis is to investigate and tackle the issues underlying this inconsistency. In particular, in Chapter 2 we introduce M5L, a fully automated Web and Cloud-based CAD for the automated detection of pulmonary nodules in chest CT scans. This system introduces a new paradigm in clinical practice, by making available CAD systems without requiring to radiologists any additional software and hardware installation. The proposed solution provides an innovative cost-effective approach for clinical structures. In Chapter 3 we present our international challenge aiming at a large-scale validation of state-of-the-art CAD systems. We also investigate and prove how the combination of different CAD systems reaches performances much higher than any best stand-alone system developed so far. Our results open the possibility to introduce in clinical practice very high-performing CAD systems, which miss a tiny fraction of clinically relevant nodules. Finally, we tested the performance of M5L on clinical data-sets. In chapter 4 we present the results of its clinical validation, which prove the positive impact of CAD as second reader in the diagnosis of pulmonary metastases on oncological patients with extra-thoracic cancers. The proposed approaches have the potential to exploit at best the features of different algorithms, developed independently, for any possible clinical application, setting a collaborative environment for algorithm comparison, combination, clinical validation and, if all of the above were successful, clinical practice.