The use of ionizing radiation in medical imaging procedures, particularly in diagnostic radiology and nuclear medicine has significantly increased in the last decade. Although the use of ionizing radiation in medicine plays a pivotal role in healthcare, it is associated with risks of radiation-induced cancer. Therefore, its use is subject to standards of safety and stringent optimization procedures. To this end, the first step is the accurate assessment of radiation dose delivered to patients to assist the optimization of the given procedure.

The work presented in this dissertation aimed to establish an accurate and reliable methodology for monitoring and optimization of the radiation dose, initially from diagnostic imaging, but later in theragnostic nuclear oncology. This research addressed two main questions: first, to establish a framework for habitus-specific and patient-specific dose monitoring and radiation dose reduction in hybrid PET/CT imaging; and second, to develop a practical dosimetry workflow to bring the full capacity of theragnostic dosimetry-guided planning to RadioPharmaceutical Therapy (RPT).

In the first phase, four studies were carried out: A framework for Monte Carlo (MC) based dose calculation from both internal (i.e. PET) and external (i.e. CT) exposure was developed. and some methodology for construction of habitus-dependent and patient-specific computational model for dose estimation was proposed.

In the second phase, the application of deep learning in dose calculation and optimization was extended. Using the methods that have been developed in the first phase, we developed a novel deep learning-based algorithm for fast MC-based internal dosimetry. The proposed model was evaluated on the diagnostic ¹⁸FDG-PET examinations and was further extended to a betta- emitter therapeutic agent, ¹⁷⁷Lu-DOTATAE, using transfer learning. Furthermore, we applied the deep learning-based dose construction methodology developed for internal dosimetry into high dose rate brachytherapy.

Using our validated MC simulator for CT dosimetry, we developed a deep learning-based model to predict personalized voxel-level absorbed doses from anatomical density map and acquisition parameters. Also, we designed an ultra-low-dose CT examination protocol for clinical diagnosis of COVID-19 patients using a deep neural network.

On the grounds of previously developed personalized dosimetry methods, we further studied dosimetry in theragnostic in connection with targeted molecular radiotherapy. Neuroendocrine tumors (NETs) with overexpressing somatostatin receptors provide the basis for peptide receptor radionuclide therapy (PRRT) through theragnostic pair of ⁶⁸Ga/¹⁷⁷Lu-DOTATATE. The main purpose of this study was to develop machine learning models to predict therapeutic tumor absorbed dose using pre-therapy ⁶⁸Ga-DOTATATE PET/CT and clinicopathological biomarkers.