Failure-Source-Aware and Actionable Uncertainty for Safe Medical Imaging
Safe-Med collaborative internship · GIPSA-lab, Grenoble × Bio-AI Lab, UiT, Tromsø
Core question:
Can we move from “How uncertain is this prediction?” to “What is the most plausible reason for the uncertainty, and what should be done next?”
Problem context
Deep-learning systems can be accurate on average and still fail on individual medical images. The cause of failure matters: an examination may be degraded by motion or noise, the image may be intrinsically ambiguous even for expert readers, or the case may differ from the training distribution.
Current uncertainty-quantification methods usually return a confidence or risk score that is useful for error detection, calibration, selective prediction, or out-of-distribution (OOD) detection. However, the same level of predictive uncertainty can arise from different failure sources, which may require different responses.
Goal and objectives
The internship will investigate whether uncertainty can be made failure-source-aware and actionable.
The objective is to test whether observable uncertainty signals contain enough information to distinguish clinically meaningful failure factors beyond generic predictive risk.
The internship will aim to:
- Build a benchmark that separates acquisition degradation, reader disagreement / intrinsic ambiguity, and distribution shift.
- Compare standard uncertainty signals: softmax confidence/entropy, calibration scores, deep ensembles, Monte-Carlo dropout, test-time augmentation, and representation/OOD scores.
- Evaluate failure-source attribution at matched predictive risk: cases with similar error risk should still be distinguishable by the underlying factor.
- Test whether factor-aware decisions improve simple intervention policies, such as reacquisition/quality control, expert review, or OOD referral, under the same intervention budget.
Existing related work and current limitation
Medical-imaging uncertainty quantification is well developed for predictive confidence, calibration, and error ranking. Deep ensembles and Monte-Carlo dropout are among the most commonly used approaches, while recent benchmarks separately study robustness to image corruption and OOD detection. Multi-reader datasets also provide a way to quantify disagreement.
The limitation is that these lines of work are usually evaluated separately. Predictive UQ asks whether a prediction is reliable; OOD methods ask whether an input differs from the training distribution; quality-control studies detect degraded images; and multi-rater work models annotation variability.
There is still limited evidence that one framework can:
- distinguish these factors when predictive risk is held approximately constant,
- handle mixed failure factors, and
- show that the distinction changes the downstream action.
Open question and working hypothesis
Open question:
Does the pattern of uncertainty signals contain information about the source of likely failure that is not already explained by the scalar probability of being wrong?
Working hypothesis:
Two cases can have comparable predictive uncertainty but different uncertainty fingerprints.
Acquisition corruption may affect augmentation sensitivity and image-quality signals, reader ambiguity may be reflected in prediction dispersion on genuinely difficult cases, and distribution shift may be more visible in representation- or density-based scores.
Combining these signals may support factor attribution and better routing decisions without replacing the underlying predictive model.
Datasets and why they are used
| Dataset | Failure factor | Usefulness |
|---|---|---|
| ROOD-MRI | Acquisition degradation | Controlled MRI corruptions and distribution shifts with graded severity; useful for paired tests on the same anatomy under different acquisition conditions. |
| LIDC-IDRI | Reader disagreement / ambiguity | Thoracic CT scans reviewed by four experienced radiologists without forced consensus; provides natural case-level disagreement rather than synthetic label noise. |
| OpenMIBOOD | Distribution shift | Standardized medical-imaging benchmarks spanning covariate shift, near-OOD, and far-OOD cases, with a broad set of post-hoc OOD methods. |
| CAMUS (external validation) | Clinical image quality | 500 echocardiography examinations with realistic acquisition variability and expert image-quality categories; useful for testing transfer from controlled corruption to naturally poor-quality images. |
Indicative tasks
The selected student will:
- Reproduce strong predictive-UQ and OOD baselines and define a common evaluation protocol.
- Construct controlled and natural failure-factor test sets, including severity levels and mixed-factor cases.
- Measure error detection, calibration, and factor attribution jointly, including matched-risk analyses to control for generic case difficulty.
- Study which signals are specific, redundant, or confounded by dataset identity, task, or model architecture.
- If simple combinations are insufficient, investigate a compact factor-attribution model using uncertainty signals or learned representations.
- Evaluate actionability through fixed-budget routing experiments and document failure cases and limitations.
Student profile
We are looking for a Master 2 / final-year engineering student with a strong interest in machine learning for medical imaging and reliable AI.
The candidate should have:
- Good Python and PyTorch skills and be comfortable training and evaluating CNN/ViT-based models.
- A solid understanding of supervised and semi-supervised learning, probability/statistics, and experimental evaluation.
- Ability to work with medical-image datasets and standard preprocessing pipelines. Experience with MONAI is a plus but is not mandatory.
- Strong interest in uncertainty quantification, calibration, OOD detection, robustness, or representation learning.
- Careful scientific practice, including reproducible experiments, clear baselines, ablations, statistical analysis, and concise technical writing.
- Fluency in English, which will be the main working language of the internship and the international collaboration.
Knowledge of French is not mandatory, although it can be useful for day-to-day interactions within the team. GIPSA-lab provides an international research environment in which English is commonly used.
Eligibility condition
Applicants must currently reside in the European Union and must be able to undertake the internship in France without requiring a visa.
Project environment, supervision and international mobility
SafeMed is a collaborative research project between GIPSA-lab in Grenoble, France, and the Bio-AI Lab at UiT – The Arctic University of Norway in Tromsø.
SafeMed is separate from the ANR JCJC project EQUR-XSM, but its scientific theme follows the same broader research direction: reliable, explainable, and uncertainty-aware AI for medical imaging.
| Host environment | GIPSA-lab, Grenoble · integration within the PICSA team |
| Internship period | Start possible in February 2027 · duration: 6 months |
| International mobility | A short visit to UiT in Tromsø is planned after approximately month 4 to present preliminary results, discuss the methodology, receive feedback, and refine the final phase of the internship. |
| Collaboration | Bio-AI Lab · Department of Computer Science · UiT – The Arctic University of Norway |
Grenoble supervisor
Dawood AL CHANTI
Associate Professor at Grenoble INP – UGA and researcher at GIPSA-lab, where he is a member of the PICSA team within the Data Science Department.
At Grenoble INP, he is involved in the SICOM and Biomedical Engineering (BIOMED) programs.
His research focuses on Explainable AI, uncertainty quantification, out-of-distribution (OOD) detection, and medical image analysis, with current research activities largely centered on reliable and trustworthy AI for medical imaging.
Research website:
https://dawoodchanti.github.io/
UiT supervisor
Prof. Dilip K. Prasad
Professor in the Department of Computer Science at UiT – The Arctic University of Norway and co-founder of the Bio-AI Lab.
His research interests include interpretable AI, scalable and sustainable AI, and applications of AI to life sciences.
UiT profile:
https://en.uit.no/ansatte/person?p_dimension_id=88131&p_document_id=615677
Bio-AI Lab:
https://www.bioailab.org/
Internship compensation
In accordance with French regulations, the internship will be compensated at approximately €660 per month.
Expected outcome
The expected result is a rigorous benchmark and first evidence for or against the hypothesis that failure-source information can be recovered beyond predictive risk.
A successful internship should clarify:
- when factor-aware uncertainty is possible,
- when it is confounded, and
- whether it improves practical routing decisions.
The results of the internship are expected to be valorized through a submission to an international conference.
If the work is accepted, the student will be supported to attend the conference and present the results. The associated registration and travel costs will be covered by the project.
Application
Interested candidates should send:
- CV
- Motivation letter
to:
Dawood AL CHANTI
📧 dawood.al-chanti@grenoble-inp.fr
Email subject: Safe-Med Stage Application
Please note that applicants must currently reside within the European Union and be able to undertake the internship in France without requiring a visa.
Selected references and resources
- Uncertainty Estimation in Medical Image Classification: Systematic Review (2022)
https://pmc.ncbi.nlm.nih.gov/articles/PMC9382553/ - Boone et al. — ROOD-MRI: robustness to OOD and corrupted MRI data
https://arxiv.org/abs/2203.06060 - Gutbrod et al. — OpenMIBOOD: Open Medical Imaging Benchmarks for Out-Of-Distribution Detection, CVPR 2025
https://openaccess.thecvf.com/content/CVPR2025/html/Gutbrod_OpenMIBOOD_Open_Medical_Imaging_Benchmarks_for_Out-Of-Distribution_Detection_CVPR_2025_paper.html - LIDC-IDRI — The Cancer Imaging Archive
https://www.cancerimagingarchive.net/collection/lidc-idri/ - CAMUS dataset — CREATIS
https://www.creatis.insa-lyon.fr/Challenge/camus/databases.html - VinDr-CXR multi-radiologist dataset (possible alternative/additional ambiguity dataset)
https://www.physionet.org/content/vindr-cxr/1.0.0/
