Context
Deep learning and foundation models have achieved strong performance in medical imaging, but their end-to-end nature often limits the clinical interpretability of their reasoning and clinicians’ ability to intervene in the decision process.
Concept Bottleneck Models (CBMs) address this limitation by mediating predictions through interpretable clinical concepts, such as anatomical abnormalities or imaging biomarkers. Their intervenable structure allows clinicians to correct erroneous concepts and prioritize interventions based on uncertainty. However, CBMs often rely on incomplete predefined concept sets, require human intervention to resolve uncertain concepts, and rarely leverage corrections to improve future cases.
CARe-CBM (Continual Agentic Reasoning with Concept Bottleneck Models) extends CBMs with uncertainty-aware evidence acquisition and continual learning from validated interventions. When a clinical concept is uncertain or insufficient, an agent can selectively acquire evidence from specialized models, clinical information, medical knowledge, or expert feedback to refine the concept and final prediction. Validated interventions can then be retained to improve reasoning on future cases.
Internship Objectives
The internship has three main objectives:
- Develop a CBM that maps medical images to clinically meaningful concepts.
- Integrate concept-level uncertainty with agentic evidence acquisition.
- Develop closed-loop learning to retain validated clinician interventions and improve future predictions.
The methodology will initially be evaluated on the public OLIVES ophthalmology dataset, which provides optical coherence tomography (OCT) images with expert-annotated ophthalmic biomarkers. These annotations will serve as clinical concepts for evaluating concept prediction, uncertainty-guided interventions, and their impact on final predictions. As an extension, we will explore concept construction with limited annotations using clinical knowledge and vision-language models.
Scientific Environment
The internship will be carried out at IMT Mines Alès within the CARe-CBM project, in collaboration with IMT Nord Europe and CHU Lille. The project is supported by the IMT scientific community Ingénierie et Services de la Santé and builds on scientific synergies with the European TWIN-OPHTALMO project, funded and starting in 2027.
Candidate Profile
We are looking for a Master 2 student motivated by medical image analysis and deep learning. A background in biomedical imaging, computer vision, artificial intelligence, or a related field, together with experience in Python and PyTorch, is required.
Familiarity with supervised learning and model evaluation is expected. Experience with uncertainty quantification, ophthalmic OCT, or interpretable AI is an advantage. Good communication and teamwork skills are important, as the student will collaborate with researchers at IMT Mines Alès and IMT Nord Europe and with clinicians at CHU Lille. The student should be able to read and write scientific English and discuss research results in English.
Practical Information
- Duration: 6 months
- Starting date: between February and April 2027
- Location: IMT Mines Alès, 6 avenue de Clavières, 30319 Alès Cedex, France
- Supervisors: G. Andrade-Miranda, Sébastien Harispe and Pedro Soto Vega (IMT Mines Alès), and Halim Benhabiles (IMT Nord Europe)
- Clinical collaboration: Dr Mathilde de Massary (CHU Lille)
Applications: Candidates should send a CV, a motivation letter, and their latest grade transcripts. Recommendation letters or contact information for former supervisors are optional.
Contacts:
gustavo.andrade-miranda@mines-ales.fr
pedro-juan.soto-vega@mines-ales.fr
sebastien.harispe@mines-ales.fr
halim.benhabiles@imt-nord-europe.fr
Full internship offer: Download the detailed internship offer (PDF)
References
[1] P. W. Koh et al., “Concept Bottleneck Models,” Proceedings of ICML, 2020.
[2] C. Steinmann et al., “Learning to Intervene on Concept Bottlenecks,” Proceedings of ICML, 2024.
[3] G. Andrade-Miranda et al., “AgentQC: Policy-Constrained Agentic Assessment of Task-Aware Reliability in Medical Imaging Datasets,” MICCAI Workshop on Agentic AI for Medicine (AgenticMed), 2026.
[4] G. Prabhushankar et al., “OLIVES Dataset: Ophthalmic Labels for Investigating Visual Eye Semantics,” NeurIPS Datasets and Benchmarks, 2022.
