Evaluation de l’apport du scanner double énergie pour la définition des volumes cibles et organes à risque en radiothérapie

Conventional single-energy CT (SECT) is currently the reference imaging modality for radiotherapy (RT) treatment planning. However, its limited soft tissue contrast often requires the use of additional modalities such as PET or MRI, which introduce registration uncertainties despite their clinical added value. In this context, dual-energy CT (DECT) stands out by providing enhanced contrast images, notably virtual monoenergetic images (VMI). These images, well-supported in the diagnostic imaging literature, improve the visualization of tumors and soft tissues, especially in the 40-60 keV range compared to SECT.A survey conducted by an ESTRO working group in 2021 revealed that many RT departments have access to DECT scanners but do not routinely use their advanced features. In the absence of clear guidelines and dedicated studies, it becomes essential to evaluate the relevance of DECT reconstructions for segmentation and assess whether they could, in some cases, replace additional imaging modalities.Segmentation, which is the definition of organs at risk (OAR) and target volumes (TV), is a major source of uncertainties in RT planning. When performed manually by clinicians, it is subject to high intra- and inter-observer variability, partly due to the limited contrast of CT images. Automatic segmentation tools, already clinically implemented for OARs and lymph nodes, help reduce this variability. Automatic segmentation of TV is still under development and represents a key area of research.This thesis is structured around two objectives to explore the contribution of DECT to segmentation. The first objective was to assess the performance of a clinically deployed commercial automatic segmentation software on various DECT reconstructions. To this end, data from 74 head and neck cancer (HNC) patients were used, including polyenergetic images at 80 kV with and without the DirectDensity® (DD) kernel, 120 kV-equivalent images, and 40 keV VMI. Clinical segmentations were used as ground truth (GT) and compared to automatic contours performed on these four image types. Performance was evaluated using the Dice Similarity Coefficient (DSC), the 95th percentile Hausdorff Distance (95HD), and the Mean Surface Distance (MSD) across 13 OAR structures. The 80 kV DD image yielded the highest DSC scores, significantly outperforming the other reconstructions for all organs (p < 0.05). These results highlighted that the performances of the auto-contouring system can not be generalized across all DECT images, and emphasized the need to identify specific organs that would benefit from these acquisitions to better tailor training datasets.The second objective focused on building a database of head and neck cancer (HNC) patients with non operated tumors to train a model for automatic tumor segmentation from DECT-derived series. Two approaches were pursued: first, the implementation of DECT as the planning image in the clinical protocol; and second, the development of a proof-of-concept DECT acquisition of virtual patients using Monte-Carlo simulation, in order to artificially expand the training datasets of automatic segmentation algorithms. A source model incorporating filtration was defined based on measured spectra and compared to a model based on manufacturer data. This model was validated through half-value layer (HVL) and CTDI measurements and combined with a scanner detector model to generate realistic projections. Finally, the XCAT anthropomorphic phantom library was adapted to include realistic tumor volumes based on patient-derived contours (GTV-T).This work highlights the need to readjust the RT segmentation process to the integration of spectral images, particularly VMI, with the rise of spectral CT scanners. It also emphasizes the potential of MC simulation to assess new imaging systems and to create virtual patient libraries.