CEMMTAUR : 

CT synthesis from Multicentric and Multisquence MRI daTA

with qUality assessment for image-guided Radiotherapy

CominLabs 2022

The aim of radiotherapy (RT) is to deliver a prescribed dose to the tumor while sparing organs at risk of toxicity (OAR). CT (Computer Tomographic)-scans are nowadays the reference imaging for dose planning in radiotherapy, because they provide the electronic density of tissues required for dose calculation. However, CT has limitations related to its poor contrast in the soft tissues causing inaccuracies, mainly in the delineation of the tumor and the organs-at-risk (OAR) as well as in the patient repositioning at each fraction. Conversely, thanks to its high contrast level, MRI is the reference modality for soft tissue imaging and thus, for manual delineation of most tumors and OAR. Moreover, contrary to CT, MRI is non-irradiant.

However, MR-images do not provide electronic density information. To perform dose calculation for planning RT from MRI, several methods have been recently developed, allowing the generation of a pseudo-CT or synthetic CT (sCT). In general, DLMs have shown better performance compared with the bulk density, atlas-based and patch-based approaches, leading to lower reconstruction and dosimetric errors.

A major limitation of DLMs to generate sCT in daily practice is their dependency on the training and validation cohorts which are specific to a center or CT/MRI device, impeding the generalization of the sCT approach. Moreover, the data used in the above studies for sCT training were usually paired, that is, consisting of MR/CT image pairs corresponding to the same patient with images exhibiting pixel‐to‐pixel correspondences. However, it is not easy to get paired data from patients scanned by both CT and MR scanners (and a fortiori several MRI sequences), which can significantly delay data collection. Collecting a sufficiently large training set with accurately registered paired MR and CT volumes and manual delineations of all images is unpractical. In this project, we seek to take advantage of MR or CT training volumes from patients who were scanned for different purposes and who have not necessarily been imaged using both modalities, but whose partial evidence contributes to modeling the expected anatomical variability. Moreover, since there are plenty of unpaired medical images, the available datasets can be easily enlarged. Being able to use unpaired MR and CT training data would relax current limiting constraints of deep learning-based CT synthesis systems, but requires the design of adapted learning techniques.

 

In this context, our project focuses on developing an MRI-to-sCT algorithm for accurate planning with the following objectives in mind:

• Segmentation (delineation) of target volume and OAR across different MRI (T1 and T2, with contrast dye).
• Multimodal non-rigid registration approach (MRI/CT) used for paired training or targeting.
• Improving tumor targeting by exploiting multisequence MRI data.
• Proposing unsupervised (unpaired) learning methods for multi-centre use.
• Standardizing the sCT evaluation.
• Designing of quality assurance methods.

 

This work will focus on pelvic, and brain cancer localizations. Medical images (CT, T1/T2 MRI) of these pathologies will be provided by the Centre Eugène Marquis (CEM).

This project involves two partners collaborating for the first time.

– The LTSI (Laboratoire Traitement du Signal et de l’Image, INSERM U1099, Université de Rennes 1) brings a long history in computer-aided approaches for radiotherapy, including recent approaches for sCT synthesis.

– and the LS2N (Laboratoire des Sciences du Numérique de Nantes, UMR CNRS 6004, Ecole Centrale de Nantes (FR)) with expertise in domain adaptation methods for segmentation, uncertainty measures and multicenter data handling.