Multi-modal implicit neural representations for MRI-guided radiotherapy

Niveau : M2/TFE

Période : Début Janvier-Mars 2025 (durée 6 mois)

Radiotherapy is a standard cancer treatment method, which has to balance delivering a prescribed radiation dose to a tumour while sparing organs at risk (OARs). Computer Tomography (CT) scans are the reference imaging for dose planning in radiotherapy because they provide tissue density information required for dose calculation. However, CT’s poor contrast in soft tissues hinders physicians’ accurate delineation of the OAR and precise patient repositioning. Conversely, thanks to its high contrast level, Magnetic Resonance Imaging (MRI) is the reference modality for soft tissue imaging and, thus, better for manual delineation of most tumour volumes and OAR. However, MR images do not provide tissue density information and therefore cannot be used for dose calculation. In this context, the long-term goal of the CEMMTAUR project is to develop new methods and tools for MRI-guided radiotherapy, focusing on brain and prostate cancer.

Mission: The goal of this internship is to assist the automatic segmentation and registration of OARs in multi-modal CT/MR images, towards enabling dose calculation in either modality. To this end, the intern will

• rely on two annotated datasets from local hospitals with delimited tumours and organs at risk in two anatomies (brain and prostate).
• Develop innovative deep learning approaches for segmentation and registration across modalities.

The developed method will rely on the new Implicit Neural Field Representation concept. Neural fields have been shown capable of efficiently representing and reconstructing images and shapes [1], and have been very recently extended for the segmentation [2] and registration of medical images [3]. Initial experiments on CEMMTAUR’s brain database for registration using the method in [4] were very promising. In this new internship we will focus on the generalisation of the approach to the simultaneous registration of several patients, as well as on the multi-modal aspect [3,5]

[1] Amirali Molaei, Amirhossein Aminimehr, Armin Tavakoli, Amirhossein Kazerouni, Bobby Azad, Reza Azad, Dorit Merhof. Implicit Neural Representation in Medical Imaging: A Comparative Survey [github]

[2] Zhang, Y., et al.. Swipe: Efficient and robust medical image segmentation with implicit patch embeddings. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 315-326). 2023

[3] Sun S, Han K, You C, Tang H, Kong D, Naushad J, Yan X, Ma H, Khosravi P, Duncan JS, Xie X. Medical image registration via neural fields. Med Image Anal. 2024 Oct;97:103249. doi: 10.1016/j.media.2024.103249. Epub 2024 Jun 27.

[4] Byra, M., Poon, C., Rachmadi, M.F. et al. Exploring the performance of implicit neural representations for brain image registration. Sci Rep 13, 17334 (2023). [github]

[5] Sideri-Lampretsa, V., McGinnis, J., Qiu, H., Paschali, M., Simson, W., & Rueckert, D. (2024). SINR: Spline-enhanced implicit neural representation for multi-modal registration. In Medical Imaging with Deep Learning.

Diana Mateus (diana.mateus@ls2n.fr), Mathilde Monvoisin