Artificial intelligence strategies to support nuclear medicine image analysis

The automatic segmentation of gross tumor volume (GTV) in PET-CT images using artificial intelligence techniques is one of the challenges that the biomedical field is facing in the last decade. Currently, the images are mainly contoured manually or using approximate strategies such as thresholding, region growing or level set methods in order to outline the lesion and define a surgical plan and consequent therapy in relation to the degree of aggressiveness of the tumor. The use of automatic contouring approaches can facilitate and support nuclear radiologists and oncologists in this task. This thesis project aims to implement an automatic algorithm, based on deep learning methods, that assists physicians in tumor segmentation of the head and neck area, the seventh most common type of cancer in the world, which, in recent years, has seen an increase due mainly to the high use of alcohol and tobacco. The images used for the training of the algorithm were provided by the Ospedale Maggiore di Novara, in particular PET images of 89 patients and the correlated manual contouring of GTV were used. The first phase of the implementation consists of the data pre-processing, in particular the resampling and normalization of the images with respect to the SUV values were done. Starting from the manual contouring of each patient, "binary masks" have been created, and these will be used to instruct the network in order to obtain the tumor delineation. 80% of the dataset is used for the training of the implemented convolutional neural network (CNN), 10% for the test set and the remaining 10% for the validation set. The dataset was homogeneously divided based on the size of the lesion in each binary mask. Various approaches to network training have been implemented; in particular the performance of the algorithm has been evaluated, giving the entire image or sub-areas based on the position and size of the lesion in the complete image. The best results were obtained by centering the image on the surrounded ROI and setting the box size equal to the maximum size of the lesion in the dataset used, the coefficient of dice similarity (DSC) obtained is 0.791. The result is encouraging, especially given that the PET images have intrinsic resolution limitations, standard deep learning architecture may not be suitable enough for the segmentation image process, and a fine-tuning approach can be more successful. Ultimately, this research demonstrates how the use of artificial intelligence-based approaches, such as deep learning, can make significant contributions to the head and neck tumor's segmentation task, supporting medical team.