Capsule Networks per la segmentazione di noduli polmonari

Lungs cancer is one of the most life-threatening diseases in the world. Early diagnosis of lung cancer is crucial to enable possible life-saving interventions and it relies on accurate quantification of pulmonary nodules, which may be cancerous. This thesis project aims to provide neural networks designed for detecting and segmenting lung nodules from medical images, thus this network is configured to solve the task of medical image segmentation. Medical image segmentation consists of automatically detecting and outputting the presence, the position and the volume of a specific anatomical structure in a medical image. Manual segmentation of nodules is a tedious and time-consuming task, therefore experts radiologists would save hours of manual work if they would use the proposed system. The neural networks were trained on the Lung Image Database Consortium image collection (LIDC-IDRI) dataset once the dataset was preprocessed, consequently experiments were performed on this dataset. The LIDC-IDRI consists of diagnostic and lung cancer screening thoracic computed tomography (CT) scans with marked-up annotated lesions. In particular, the project is focused on building and training a capsule network configured to solve the task of medical image segmentation in order to demonstrate that capsule networks are able to solve this type of task. This study allows us to observe and disclose the characterizing properties of capsule networks and it allows us to identify their advantages and disadvantages compared to other neural networks. Specifically, a Unet, a convolutional neural network developed for biomedical image segmentation, and a SegCaps, a convolutional-deconvolutional capsule network, have been implemented and analyzed. Therefore, as a result of this project, systems capable of solving the segmentation problem on medical images, more specifically on CT scans, have been developed and studied. Furthermore, this thesis comprises a set of opinions related to the usability of Capsule Networks for the task of image segmentation.