Thèse de doctorat en Pathologie et recherche clinique
Sous la direction de Alain Luciani.
Soutenue le 16-11-2017
à Paris Est , dans le cadre de Ecole doctorale Sciences de la Vie et de la Santé (Créteil ; 2015-....) , en partenariat avec Institut Mondor de Recherche Biomédicale (Créteil) (laboratoire) .
Le président du jury était Alain Rahmouni.
Les rapporteurs étaient Ivan Bricault, Nico Buls.
Development of a system helping in optimizing and tailoring computed tomography protocols
Radiation dose from Computed Tomography has become the subject of public attention leading to efforts to optimize acquisition protocols. Among the large number of clinical indications addressed in Computed Tomography, the assessment of primary liver cancer (i.e. Hepatocellular Carcinoma) is a situation of clinical concern due to its rising incidence and the complex multiphasic imaging process required.The selection of the appropriate clinical setting for this specific indication required to priorly design a unique anthropomorphic phantom. In addition to the phantom, two sets of additional fat plates were developed to tailor this research to various patient body types.Individual optimization of Computed Tomography protocols necessitates to precisely determine the X-ray attenuation of the patient before proceeding the helical acquisitions. The recurrence of vertical miscentering of patient during Computed Tomography examinations, leading to enlargement or reduction of the patient’s size on anteroposterior and posteroanterior localizers, appeared as a main obstacle to the appropriate quantification of the patient attenuation. This study introduced and validated a method overcoming this issue to provide an accurate patient’s attenuation quantification regardless of vertical centering.Although patient’s body type significantly affects the amount of X-rays needed to perform an exam, the setting of Computed Tomography acquisition and reconstruction parameters depends mainly on the clinical indication. This study introduced automatic methods of image quality assessment adapted to the unique anthropomorphic phantom dedicated to the assessment of primary liver cancer. The proposed metrics adequately quantify the radiologist’s perception of key features of image quality and hence, can be used to optimize Computed Tomography protocols adapted to the diagnosis of hepatocellular carcinoma. While this quantitative assessment of image quality revealed that a unique acquisition protocol properly optimized can provide a constant image quality for various patient’s body types, it appeared that contrast medium injection must be individually adapted to improve organ and lesion enhancement and reduce its patient-to-patient variability.The global image quality clinically needed for the assessment of primary liver cancer could not be defined using a unique descriptor; but rather as a multidimensional combination of image properties. This study investigated and demonstrated the capability of machine learning to apprehend radiologists’ preferences of image quality based on these specific image features. Trained classifiers can be used to predict if an image would be acceptable for a diagnosis and to optimize Computed Tomography protocols accordingly across Computed Tomography scanners without repeating time and money consuming clinical observer experiments.