Download Radiomics Book in PDF, Epub and Kindle
"In this thesis, the major aim is to develop radiomic-based models for the accurate prediction of tumour outcomes via advanced machine learning. We first showed that the optimization of how texture features are extracted from medical images (different isotropic voxel sizes, image quantization schemes, etc.) is fundamental for best tumour outcome prediction. We then integrated the texture optimization process into a robust multivariable modeling methodology developed for the construction of radiomic-based prediction models. This multivariable modeling methodology employs logistic regression to linearly combine radiomic features. Using this methodology, we were able to develop a model that can predict the development of lung metastases in soft-tissue sarcomas with high accuracy. This model combines texture features extracted from functional FDG-PET and anatomical MRI pre-treatment images. Following this initial work, we demonstrated how the predictive properties of imaging textures composing such prediction models could be further enhanced by optimizing the way images are acquired. The proof of concept for the enhancement of the prediction of lung metastases in soft-tissue sarcomas was carried out using computerized simulations of FDG-PET and MR image acquisitions with tumour and clinical scanner models, by varying different physical parameters employed during image acquisitions. Next, in another study, we developed a strategy for personalizing treatments for soft-tissue sarcoma patients identified at diagnosis to be at higher risks of developing lung metastases (using radiomic-based prediction models); specifically, we verified the feasibility of applying double nested radiation dose boosting to the hypermetabolic and hypoxic soft-tissue sarcoma sub-regions to counteract the progression of more aggressive parts of tumours. For the purpose of radiation treatment planning, contours defining the hypermetabolic and hypoxic tumour sub-regions were obtained from FDG-PET and low-perfusion DCE-MRI functional images. Finally, in our last study, we developed a methodology allowing to integrate radiomic imaging data with clinical prognostic factors into comprehensive prediction models using a random forest algorithm. We tested our methodology in head-and-neck cancer to better assess the risk of locoregional recurrences and distant metastases, this time using functional FDG-PET and anatomical CT pre-treatment images in conjunction to clinical data. The clinically-integrated radiomic models that we developed possess high prognostic power, leading to patient stratification into two sub-groups for the risk assessment of locoregional recurrences (low, high) in head-and-neck cancer, and into three groups for distant metastases (low, medium, high).Overall, in this thesis, we demonstrated that radiomics analysis is an enabling method towards precision medicine. The different radiomic techniques and models developed in this work could have a major impact on the design of new clinical trials aiming at a better personalization of cancer treatments. One can envision different treatment regimens being delivered to patients based on different radiomic-based risk assessments of specific tumour outcomes." --