Job offer

Radiotherapy is used in more than half of cancer patients. To be effective and safe, it must be delivered with very high precision: the tumour must receive the prescribed dose while limiting irradiation of nearby organs at risk. This requirement is especially critical in head and neck cancers, where many anatomical structures are close to each other and sensitive to dose deviations. In head and neck radiotherapy clinical trials, this quality is monitored by experts who manually review treatment plans. This work, known as "radiotherapy quality assurance," is time-consuming, costly, and can be partly subjective: different experts may reach different conclusions for the same case. As a result, some plans may show deviations-minor or major-from trial protocols, which makes it harder to ensure consistent treatment across centres. This PhD project aims to develop artificial intelligence tools capable of automatically analysing radiotherapy plans for head and neck cancers. Using medical images, delineated structures, and dose distributions, the algorithms will learn to detect potential errors, deviations from protocol criteria, and higher-risk situations that should be prioritised for review. Performance will be assessed by comparison with experts' quality-assurance decisions and, when possible, by analysing associations between the predicted indicators and available clinical outcomes (for example reported toxicities and tumour control). These aspects will be detailed in the "context and objectives" section. This approach will be applied to patients treated with photon therapy and to an external hadron therapy cohort (real-world proton treatments and in silico carbon ion treatments), which are even more demanding. Ultimately, this approach will make it possible to check treatment quality faster and more consistently for patients enrolled in clinical trials, without slowing down their care. It will help improve patient safety, the reliability of clinical trials, and the overall effectiveness of radiotherapy treatments against cancer.

Detailed project, context, scientific questions, methodological program and perspectives Scientific questions and challenges Quality assurance in radiotherapy (RTQA) currently relies on expert-based review of treatment plans. While this approach has been foundational, it presents several critical limitations: -First, limited scalability: expert review cannot be applied exhaustively in large multicentric trials. In practice, only a subset of cases is evaluated, which prevents detection of systematic deviations, temporal drifts, or operator-dependent variability. -Second, subjectivity: inter-observer variability leads to inconsistent assessments, even for identical treatment plans, limiting reproducibility and standardization. -Third, weak linkage with clinical outcomes: RTQA is primarily protocol-driven and infers potential treatment failure based on deviations from predefined constraints, without direct validation against actual patient outcomes such as toxicity or tumor control. Although RTQA must provide prospective feedback before treatment initiation, its predictive value could be significantly enhanced by incorporating outcome-driven learning from prior patients. -Fourth, loss of spatial information: conventional RTQA metrics, such as dose-volume histograms (DVH), reduce complex three-dimensional dose distributions into summary indicators, thereby ignoring spatial heterogeneity and local dose-anatomy interactions that are critical determinants of both efficacy and toxicity. Altogether, these limitations prevent RTQA from becoming a standardized, predictive, and patient-centered process.

Scientific strategy This project aims to develop an artificial intelligence-based framework for radiotherapy quality assurance, explicitly linking treatment plans to clinical outcomes. The core hypothesis is that RTQA should not only assess compliance with protocol constraints, but also predict clinically relevant endpoints, such as toxicity and tumor control. To address this, the project will leverage hierarchical Multiple-Instance Learning (MIL) combined with deep learning, enabling the modeling of complex, multimodal, and spatially structured data at the patient level.

This work is conducted within the framework of the national Streamline-QART program, which provides access to large-scale, curated radiotherapy datasets.

Methodological program The project is structured into four complementary work packages:

1. MIL-Attention on tabular data Clinical variables, DVH-derived metrics, organ-at-risk constraints, and protocol parameters will be modeled using Multiple-Instance Learning with attention mechanisms (MIL-Att).

Each patient will be represented as a set of instances (e.g., organs or DVH components), allowing the model to: identify key contributors to RTQA failure quantify their relative importance link these factors to clinical outcomes

2. MIL-Attention on imaging data (CT + RTDOSE) Three-dimensional CT images and corresponding dose distributions will be decomposed into spatial regions. A spatial MIL-Attention model will learn: which anatomical regions associated with specific dose patterns are predictive of toxicity or tumor control.

This approach preserves spatial information and enables the generation of localized risk maps.

3. Multimodal fusion through MIL-Att

Tabular and imaging representations will be integrated into a multimodal MIL framework, enabling joint modeling of: dosimetric parameters protocol deviations spatial dose-anatomy patterns This fusion will allow the model to capture complex interactions across modalities and improve predictive performance.

4. Outcome-driven RTQA Beyond reproducing expert RTQA scores, the models will directly predict: severe toxicities tumor control outcomes

This represents a paradigm shift toward clinically grounded RTQA, where plan quality is evaluated based on its expected impact on patients.

External validation will be performed using an independent hadron therapy cohort (protons and in silico carbon ions) to assess generalizability across modalities. Expected scientific outcomes Development of interpretable AI models for RTQA Identification of key determinants of treatment quality and failure Generation of spatial dose-toxicity and dose-control maps Establishment of a patient-centered RTQA paradigm linking planning to outcomes Perspectives for the team and the laboratory

This project will position the team at the forefront of AI-driven radiotherapy quality assurance and outcome modeling. At the scientific level, it will: -strengthen expertise in multimodal learning and spatial modeling -enable integration with ongoing projects in digital twins and world models -foster collaborations between clinicians, physicists, and data scientists

At the clinical level, it has the potential to: improve treatment personalization reduce toxicity through better plan evaluation standardize RTQA across centers

At the institutional level, it will contribute to: large-scale national and international programs development of translational pipelines from data to clinical decision support Ultimately, this work aims to transform RTQA from a retrospective, protocol-based verification process into a prospective, predictive, and outcome-driven decision tool.

Clinical validation of automated RTQA A key objective of the project is the clinical validation of automated RTQA approaches, by demonstrating their ability to predict clinically relevant outcomes, including toxicity and tumor control, beyond conventional expert-based assessments. This validation will rely on retrospective and prospective datasets, and on external cohorts, to ensure robustness, reproducibility, and generalizability across institutions and treatment modalities.

Perspectives This PhD project will strengthen the local expertise in the field of artificial intelligence applied to radiotherapy. It will: foster the development of innovative methodologies at the interface of clinical research, medical physics, and data science support the production of high-impact international publications contribute to the design and conduct of national and international clinical trials More broadly, the project will position the teams as key contributors to the emerging paradigm of AI-driven, outcome-based radiotherapy quality assurance, with potential extensions toward digital twins and personalized treatment optimization., The PhD position will be openly advertised at European/international level (e.g., EURAXESS, Campus France) for a minimum of one month. Advertisement period: April 1st, 2026 - June 6th, 2026 The position is published subject to final funding confirmation., Final selection The project supervisors (J Thariat, A Corroyer) will propose a ranked shortlist of up to three candidates.

The PSIME Doctoral School will: -review the applications -validate eligibility -confirm or adjust the ranking

Final decision will be made during the Doctoral School selection committee (June 18th, 2026).

Outcome The PhD position will be offered to the top-ranked validated candidate. In case of withdrawal or administrative ineligibility, the position may be offered to the next candidate on the ranked list. Additional comments

We value curiosity, collaboration, and the ability to engage in a multidisciplinary clinical research environment. You will join an ongoing project involving several students, with established collaborations and a supportive research environment.

Master Degree or equivalent

Research Field Computer science

Education Level Master Degree or equivalent

Skills/Qualifications

Educational background Master's degree (or equivalent) in one of the following fields:Medical physics, Data science / Artificial intelligence, Applied mathematics / Statistics, Computer science

Core scientific skills Strong foundation in: Machine learning and/or deep learning; Statistical modeling and data analysis Understanding (or strong interest) in: Medical imaging (CT, imaging-derived data); Radiotherapy concepts (dose, DVH, treatment planning); Ability to work with multimodal and longitudinal datasets Technical skills Programming proficiency in: Python (preferred) or R Experience with:Deep learning frameworks (e.g., PyTorch, TensorFlow); Data preprocessing and large-scale dataset handling Familiarity with:Image processing and/or 3D data analysis; Scientific computing environments Methodological competencies Interest in advanced learning paradigms, including:Multiple-Instance Learning (MIL); Attention-based models; Multimodal data fusion Ability to:Develop, train, and evaluate predictive models; Interpret model outputs and ensure clinical relevance Awareness of:Model validation, generalizability, and robustness issues

Transversal skills Strong analytical and problem-solving abilities Capacity to work in an interdisciplinary environment (clinicians, physicists, data scientists) Ability to translate clinical questions into quantitative models Scientific rigor and autonomy Communication and dissemination Good written and oral communication skills in English

Ability to:Contribute to scientific publications; Present results in international conferences; Willingness to engage in scientific outreach (CSTI) activities Personal qualities: Curiosity and motivation for translational research; Interest in improving patient care through innovation; Team spirit and adaptability Initiative and commitment over a long-term project; Additional desirable skills (not mandatory) Prior experience in:Medical data analysis; Radiotherapy or oncology; Clinical research datasets Knowledge of: DICOM formats (RTDOSE, RTSTRUCT); GPU computing environments

Summary: We seek a candidate with strong skills in machine learning, data analysis, and programming (Python), with an interest in medical applications and the ability to work at the interface between clinical research and artificial intelligence. Specific Requirements

Eligibility and academic requirements -Master's degree (or equivalent, 300 ECTS) completed before the start of the PhD -Strong academic record in a relevant field (medical physics, AI, data science, applied mathematics, computer science, or related discipline) -Eligibility for enrollment in a Normandy doctoral school -Technical and project-specific requirements -Demonstrated experience in programming (Python preferred) -Prior exposure to machine learning or deep learning methods -Ability to handle complex datasets, ideally including: medical imaging data; tabular clinical data -Interest in multimodal data integration and spatial modeling -Research environment requirements -Willingness to work in a hospital-based research environment -Ability to interact regularly with:clinicians (radiation oncologists); medical physicists; data scientists -Respect of data protection and ethical regulations related to medical data

Project engagement Commitment to a 36-month full-time PhD project

Ability to contribute to:model development;validation on multicentric datasets;scientific dissemination (publications, conferences) Language requirements: Proficiency in English (written and oral) Basic knowledge of French is an advantage (clinical environment), but not mandatory

Summary : Candidates must hold a Master's degree in a relevant field, have programming and machine learning experience, and demonstrate the ability to work with multimodal medical data in an interdisciplinary clinical research environment.

Languages

Proficiency in English (written and oral) Basic knowledge of French is an advantage (clinical environment), but not mandatory

Research Field Medical sciences

Years of Research Experience 1 - 4

Research Field Computer science

Years of Research Experience 1 - 4

Research Field Mathematics » Statistics

Years of Research Experience 1 - 4, Applicants must hold (or be about to obtain before the start date) a Master's degree (or equivalent, 300 ECTS) in a relevant field such as medical physics, computer science, data science, artificial intelligence, applied mathematics, statistics, or biomedical engineering.

Doctoral enrollment Candidates must be eligible for admission to a doctoral school in Normandy (France).

Research background Applicants should demonstrate prior experience in data analysis, programming, or machine learning, typically through Master's projects, internships, or research experience.

Language proficiency Good command of English (written and oral) is required. Basic knowledge of French is an advantage but not mandatory.

Commitment Candidates must be available for a full-time PhD project (36 months) starting October 2026.

Interdisciplinary environment Ability and willingness to work in a clinical research environment, interacting with clinicians, medical physicists, and data scientists.

Ethics and data protection Willingness to comply with ethical standards and data protection regulations related to medical data. Selection process, academic excellence -relevance of background to the project -technical and research skills -motivation and ability to work in an interdisciplinary environment

Fully funded PhD position (36 months) co-financed by the Normandy Region and Centre François Baclesse Salary aligned with French doctoral standards (including social security and healthcare coverage) Access to a unique multicentric clinical dataset in radiotherapy (imaging, dosimetry, outcomes) Opportunity to work on cutting-edge AI methods (deep learning, multimodal modeling, medical imaging) Integration in a high-level interdisciplinary environment: radiation oncologists medical physicists data scientists Strong involvement in national and international research networks (UNICANCER, GORTEC, clinical trial groups)

Support for: international conferences (ESTRO, MICCAI, AAPM…) high-impact publications Access to computational resources (GPU) and advanced research infrastructure Participation in a translational research program with direct clinical impact

Opportunities to develop skills in: AI for healthcare clinical research scientific communication Contribution to scientific outreach activities (CSTI)

Summary: Fully funded PhD, interdisciplinary clinical-AI environment, access to unique radiotherapy datasets, international collaborations, conference participation, and strong publication opportunities. Eligibility criteria