Partners

IBISC, South-Francilien Hospital Center (CHSF), Johns Hopkins University
Duration: 5–6 months
Period: February – August 2026
Monthly allowance: ~€670

Title of the Internship

Uncertainty-informed multimodal fusion for thrombus and ischemic lesion segmentation on MRI in acute ischemic stroke

The project leverages SWAN/PHASE, DWI/ADC, and TOF-MRA sequences to synthesize hypoperfusion-relevant information and improve clinical decision support [1–5].

Context

Stroke is the leading cause of acquired disability in adults and the second leading cause of death worldwide [6].
In ischemic stroke, a thrombus blocks a cerebral artery, depriving downstream tissue of oxygen.

Treatment selection (thrombectomy or thrombolysis) requires:

• precise clot localization, and

• reliable estimation of the ischemic area
  from hyperacute multimodal MRI [1–5].

However, signal variability, partial redundancy across modalities, and low-contrast / high-noise regions create segmentation uncertainty, which limits clinical accuracy [7–9].
Recent cross-modal attention approaches demonstrate high thrombus detection rates (≈0.97) but only moderate segmentation performance (Dice ≈0.65), highlighting the need to explicitly incorporate uncertainty into multimodal fusion.

Objectives of the Internship

Design and implement an uncertainty-aware multimodal segmentation system to improve thrombus and ischemia segmentation.

1. Uncertainty map generation

Compute voxel-wise uncertainty maps (entropy, Bayesian variance) to highlight ambiguous regions such as thrombus boundaries:
[
U_{\text{entropy}}(x) = -\sum_c p_c(x) \log p_c(x)
]
Using Bayesian deep learning tools for calibrated uncertainty [7–9].

2. Uncertainty-guided attentive fusion

Inject uncertainty maps as attention masks to:

• reinforce cross-modal fusion in high-uncertainty areas,

• leverage complementary evidence from secondary modalities.

Compatible with modern biomedical backbones such as U-Net [10].

3. Localized diffusion regularization (“focused blurring”)

Apply a Gaussian blur weighted by local uncertainty and train a diffusion model on blurred images to strengthen contextual learning around the thrombus:
[
I_{\text{blurred}}(x) = (I * G_{\sigma(x)})(x), \quad \sigma(x) \propto U(x)
]
Using recent denoising diffusion frameworks [11].

4. Backbones & estimators

• 3D U-Net

• Cross-modal attention networks

• Diffusion models

• Uncertainty estimation: MC-dropout, deep ensembles, Bayesian learning [7–9]

• Information-theoretic analysis (Mutual information & PID) [12,13]

5. Experimental comparison

Compare focused vs. global blur; evaluate Dice, sensitivity, and clinical precision, with modality-specific contributions from SWI/PHASE, DWI/ADC, and TOF-MRA [1–5].

Applications & Expected Impact

Datasets & environment

Multimodal MRI from stroke patients (SWAN/PHASE, DWI/ADC, TOF-MRA).
Experiments on CHSF, MATAR, and ISLES2022 datasets.
Environment: PyTorch/Python, RTX 3090, CHSF–IBISC collaboration.

Clinical utility

Improved segmentation of thrombus and ischemic regions will:

• enable more reliable estimation of penumbral “mismatch,”

• support reperfusion triage when symptom onset is uncertain,

• enhance treatment benefit prediction.

Deliverables

1. Prototype of the uncertainty-guided multimodal segmentation model

2. Quantitative analysis linking uncertainty and mutual information

3. Visual reports of clinically high-uncertainty regions

4. Draft of a manuscript targeting IEEE TMI or Frontiers in Neuroinformatics

Expected outcomes

• Functional prototype of the proposed model

• Statistical link between uncertainty and cross-modal information

• Visualization tools for uncertainty hotspots

• Manuscript preparation

Candidate Profile

We are seeking highly motivated candidates:

1. from mathematics, physics, computer science, or engineering programs

2. with a strong background in linear algebra, analysis, probability, statistics, machine learning, and deep learning

3. with solid programming skills (preferably Python)

Knowledge of medical imaging (especially MRI) is a plus, but not required.
Basic knowledge of optimization is also appreciated.

Practical Information

The intern will be mainly hosted at the UFR Science & Technology (40 rue du Pelvoux, Evry city center).
Some periods may take place at the Corbeil Hospital.
Monthly allowance: ~€670.

Application procedure

Send your:

• motivation letter

• CV

• academic transcripts (from BSc Year 1 onward)

to: Vincent Vigneron / Hichem Maaref

What We Offer

• Hands-on experience with cutting-edge AI techniques for medical imaging

• Work on real-world, high-impact healthcare problems

• Close mentorship from experienced researchers at IBISC

• Opportunities to co-author publications and present at conferences

• Possibility to continue into a PhD

Références

[1] E Mark Haacke, S Mittal, Z Wu, J Neelavalli, and Y-CN Cheng. Susceptibility weighted imaging (swi). Magnetic Resonance in Medicine, 52(3) :612–618, 2009.

[2] Àngel Rovira, Pilar Orellana, and José Alvarez-Sabín. The “susceptibility vessel sign” on t2*-weighted mri in acute ischemic stroke. Stroke, 40(2) :554–557, 2009.

[3] Steven Warach et al. Acute human stroke studied by diffusion-weighted mri. Annals of Neurology, 37(2) :231–241, 1995.

[4] DC Tong et al. Quantitative diffusion mri of acute ischemic stroke : Adc values predict tissue outcome. AJNR American Journal of Neuroradiology, 19(1) :104–110, 1998.

[5] Martin R Prince and Jeffrey Link. 3d contrast in time-of-flight mr angiography. Journal of Magnetic Resonance Imaging, 12(5) :776–783, 2000.

[6] Valery L Feigin et al. Global, regional, and national burden of stroke and its risk factors, 1990–2019. The Lancet Neurology, 20(10) :795–820, 2021.

[7] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In NeurIPS, 2017.

[8] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation : Representing model uncertainty in deep learning. In ICML, pages 1050–1059, 2016.

[9] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In NeurIPS, 2017.

[10] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net : Convolutional networks for biomedical image segmentation. In MICCAI, pages 234–241, 2015.

[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020.

[12] Paul L Williams and Randall D Beer. Nonnegative decomposition of multivariate informa- tion. arXiv :1004.2515, 2010.

[13] Amer Makkeh, Dirk O Theis, and Raul Vicente. Broja-2pid : A robust estimator for bivariate partial information decomposition. Entropy, 23(10) :1274, 2021.