Classificazione di Carcinomi Polmonari Non A Piccole Cellule tramite Immagini PET utilizzando Reti Neurali Convoluzionali e Radiomica

Lung cancer is the leading cause of cancer death worldwide among both men and women. Non-small cell lung carcinoma (NSCLC) accounts for 80 − 85% of all cases of lung cancer. Despite the significant progress obtained in the past decade, the overall five- year survival rate for NSCLC remains poor, also because the disease is often diagnosed when it is already in the advanced stages. The term Progressive Disease (PD) defines an increase of at least 20% in the size of the tumor since the beginning of therapy. Timing is a crucial aspect of cancer treatment: it would be impractical to perform multiple PET scans on the same patient in order to determine whether a cancer would be labeled as progressive or not. Considering that medical imaging most commonly relies upon visual evaluations, one way we can tackle the problem of cancer detection, characterisation, and monitoring is through advanced computational analyses. Artificial Intelligence has shown great strides in this direction. A particularly promising research field is radiomics, one of the main protagonists in this work: this approach provides high-dimensional data describing underlying tumor pathophysiology and allowing evaluation of tumor heterogeneity. The main task of this Thesis is to perform a binary tumor classification. Specifically, we aim at categorising two classes: PD and non-PD. Our objective is two-fold. First, we would like to investigate whether a Machine Learning algorithm is able to correctly classify a patient’s tumor based on the metabolic activity captured by PET imaging. Second, we aim at comparing the performance of two different strategies: implementing state-of-the-art statistical classifiers using radiomics features and, following a Deep Learning approach to cancer classification, we employ ResNet-18, an 18-layers Deep Convolutional Neural Network directly trained on PET scans. Even though we obtain very satisfying results by exploiting Tree Classifiers reaching an accuracy of about 80%, the small cohort of patients did not allow an efficient use of Deep Learning. However, we are still able to reach 66.30% accuracy using some strategies to suffice the shortage of data, like transfer learning and data augmentation. We expect the Deep Learning model’s performance to improve with respect to the use of radiomic features and Tree Classifiers when having a larger dataset.