Résumé :
Lors d’un traitement par radiothérapie, les patients oto-rhino-laryngé (ORL) passent plusieurs examens à visée diagnostique tels qu’un scanner X (CT) et une IRM (Imagerie à Résonnance Magnétique). Lors du CT le patient est positionné à l’identique au traitement. La dosimétrie est calculée sur les images du CT et l’IRM permet, au-delà du diagnostic, d’ajuster le contour des volumes tumoraux et à risque. Néanmoins, l’IRM pourrait être utilisée pour l’ensemble de la planification afin de réduire les erreurs liées au recalage intermodal et limiter l’examen à une seule imagerie non irradiante. Un CT synthétique (sCT) remplacerait l’actuel CT pour la dosimétrie. Les méthodes les plus récentes permettent de construire ce sCT par apprentissage automatique, mais elles nécessitent des bases de données suffisamment grandes et diversifiées. En effet, elles sont peu sensibles aux variations anatomiques de chaque patient et peuvent donner des résultats biaisés dans le cas de variations non présentes dans la base de données d’apprentissage. De plus, les patients atteints de cancers de la zone ORL montrent des localisations tumorales très variées et sont, en général, âgés de plus de 50 ans avec pour la plupart des plombages, couronnes ou implants dentaires. Ces plombages engendrent des artefacts dont l’impact diffère en fonction de leurs caractéristiques et de leur localisation par rapport à la tumeur et aux organes à risques (OAR) mais aussi en fonction de la modalité d’imagerie utilisée. Leur caractère imprévisible accentue le besoin de méthodes qui se détachent de l’anatomie des patients pour la génération du sCT.
Mots-clés : radiothérapie, tête et cou, sCT, dosimétrie, planification IRM, séquence UTE, séquence qDixon
Title : Quantitative MRI for head and neck radiotherapy planning
During a radiotherapy treatment, head and neck patients undergo several diagnostic CT (Computed Tomography) scan and MRI (Magnetic Resonance Imaging) examinations. Currently the clinical routine includes a CT scan in treatment position and an MRI in diagnostic position. The dosimetry is performed on CT images and MRI is used as a support for volumes contours. However, to reduce errors due to inter-modalities registration and limit the acquisition to one non-irradiant modality, MRI could be used for the whole process. A synthetic CT (sCT) based on the MRI would replace the CT for the treatment planning. Emerging methods allow to obtain this sCT with machine learning, but a large database filled with diverse cases is necessary. These methods are not very sensitive to anatomical variations and the database can be biased by peculiarities different from the database characteristics. Moreover, head and neck patients can have varied tumor location (larynx, pharynx, nasopharynx, oral cavity, etc.) and are mostly over 50 years old with, for most of them, fillings, crowns or dental leads. These leads generate artifacts with different impacts depending on their characteristics and location, in relation to the tumor and the organs at risk (OAR) but also according to the type of imaging modality performed. Their unpredictable nature is an additional argument in favor of methods which are not influenced by the anatomy of the patient for the generation of sCT.
In this context, this thesis aims to propose quantitative MRI methods based only on parametric or multi-parametric images to ultimately consider establishing an MRI-only workflow for head and neck radiotherapy. In the first part of this thesis, a new method is developed based on the hydrogen content of the tissues. Proton density allows to reach this content, and it can be estimated with an ultra-short echo time (UTE) MRI sequence. Some studies have shown that there is a link between the hydrogen content in the tissues and its mass attenuation coefficient and mass stopping power which are directly linked to electron density. This link allowed us to generate a synthetic CT (sCT) for doses calculation, which was compared to the reference dosimetry. The comparisons performed on 25 patients showed similar dosimetric results with the literature (mean dose difference <2%). However, the mean absolute error (MAE) between the CT and the sCT is important due to the implants signal and difficulties with bone segmentation.
In the second part of this thesis, the impact of the quality of the sCT on the dosimetry is assessed. Several types of sCT are generated, with a different quantity of pixels assigned to the bone tissue with a density assignment method. The protocol performed on 24 patients shows the lowest dose differences for the sCT without bone assignment (<2% for most volumes), and similar results with the most accurate and the over-estimated bone class. A CT “without bone” was also evaluated, the bone tissues are segmented with a threshold and assign the density of water (0 Hounsfield Units, HU), it gave a dose difference lower than 1% for almost all OAR and the tumor volumes. The impact of the energy of the CT scan on the contrast between tissues is also validated with a phantom. The energy of a classical CT scan is around 120 kV while the energy of radiotherapy treatment is around 6 MV. For this level of energy, the HU differences between tissues decreases, even between bone and soft tissues, resembling the MRI contrast. The importance of correct assignment of bone regions is no longer as significant as the contrast on CT images would suggest. Finally, the methods focusing on a perfect reproduction of a CT do not seem essential for dosimetry, a conversion of MR images intensities into electron densities would allow a dosimetry closer to the one generated on the reference CT.
Keywords : Radiotherapy, head and neck, sCT, dosimetry, MRI treatment planning, UTE sequence, Dixon sequence
Jury :
Johanne Bezy | Maître de Conférences, Université de Rennes | Rapporteure |
Stanislas Rapacchi | Chargé de recherche, CNRS | Rapporteur |
Eric Deutsch | PU-PH, Université Paris-Saclay | Examinateur |
Olivier Beuf | DR CNRS, INSA Lyon | Directeur de thèse |
Benjamin Leporq | Chargé de recherche, Inserm | Co-encadrant de thèse |
Vincent Gregoire | PU-PH, Centre Léon Bérard | Co-encadrant de thèse |