Keywords AI, CT-Scan, foundation models, deep learning, Data fusion, personalized medicine
Radiotherapy is a cornerstone of cancer treatment, enabling the selective destruction of tumor cells while controlling disease progression. Despite its proven efficacy across many cancer types, it can also inadvertently damage surrounding healthy tissues, leading to adverse effects that may significantly affect patient’s quality of life [1]. Consequently, accurately predicting treatment-related toxicities is essential to anticipate potential complications and tailor therapeutic planning to each individual patient.
Several studies have explored the use of artificial intelligence (AI) to predict radiotherapy-induced toxicities [2]. Most of these studies rely on conventional techniques that primarily analyze radiomic features. In a recent study, Elhaminia et al. [3] demonstrated that multimodal CNN models combining CT-scan images with clinical and dosimetric patient data can improve prediction performance. However, the models proposed in the literature are typically trained on small datasets, often involving a limited number of patients and focusing on a single cancer type, which greatly restricts their generalization capability. Moreover, none of these models have yet leveraged foundation models, even though such models represent a promising area, as demonstrated in one of our recent works on predicting treatment response in lymphoma [4].
Subject
The main objective of this internship is to develop predictive models of radiotherapy-induced toxicities that are robust, accurate, and transferable across different clinical contexts. More specifically, the work aims to leverage foundation models to extract high-level representations capable of capturing complex and discriminative features from imaging and clinical data. The internship will involve designing inter-modal and inter-model fusion strategies to integrate representations from different latent spaces by modality (CT, clinical, and dosimetric data) and by foundation model in order to fully exploit their complementarity. The main missions of this internship are:
- Conduct a literature review on Foundation models involving CT-Scans
- Design and develop a foundation model based predictive model of radiotherapy toxicities
- Perform experimentation and validation on several public datasets
Duration and supervision
The internship will last 5 months and a continuation with a PhD is potentially possible upon securing funding.
The trainee will work on-site at IMT Nord Europe in collaboration with URJC Madrid and will be supervised by the following team:
- Halim Benhabiles, IMT Nord Europe
- Norberto Malpica, URJC Madrid
Month salary is around 650€
Required skills
- Training level: Master 2 or Engineer 5th year
- Solid scientific background on machine learning techniques and image processing
- Strong capability of coding using Python, PyTorch and/or TensorFlow
- Good communication to collaborate with team members
- Good scientific writing to produce article with the aim of publishing in a relevant conference.
Application send your CV and your academic transcripts before December 15, 2025 to Halim Benhabiles halim.benhabiles@imt-nord-europe.fr
References
[1] Verginadis, I. I., Citrin, D. E., Ky, B., Feigenberg, S. J., Georgakilas, A. G., Hill-Kayser, C. E., … & Lin, A. (2025). Radiotherapy toxicities: mechanisms, management, and future directions. The Lancet, 405(10475), 338-352.
[2] Isaksson, L. J., Pepa, M., Zaffaroni, M., Marvaso, G., Alterio, D., Volpe, S., … & Jereczek-Fossa, B. A. (2020). Machine learning-based models for prediction of toxicity outcomes in radiotherapy. Frontiers in oncology, 10, 790.
[3] Elhaminia, B., Gilbert, A., Scarsbrook, A., Lilley, J., Appelt, A., & Gooya, A. (2025). Deep learning combining imaging, dose and clinical data for predicting bowel toxicity after pelvic radiotherapy. Physics and Imaging in Radiation Oncology, 33, 100710.
[4] Guetarni, B., Windal, F., Benhabiles, H., Chaibi, M., Dubois, R., Leteurtre, E., & Collard, D. (2024). Histopathology Image Embedding Based on Foundation Models Features Aggregation for DLBCL Patient Treatment Response Prediction. MICCAI workshop (pp. 150-159)
