View Proposal

---

Proposer

Heba Elshimy

Title

Multimodal MRI Segmentation of Ischemic Stroke Lesions with Calibrated Uncertainty: An ISLES 2026 Benchmark Study

Goal

To build and evaluate a deep learning pipeline that segments ischemic stroke lesions from multimodal MRI on the ISLES 2026 benchmark, adding per-voxel uncertainty estimation so the model can flag low-confidence regions for clinical review.

Description

Stroke is a leading cause of death and disability, and accurate delineation of the ischemic lesion on MRI underpins treatment decisions and outcome prediction. Manual segmentation is slow and variable, motivating automation. ISLES is the long-running community benchmark for this task, providing curated multimodal MRI with expert reference masks and a centralised, container-based evaluation. This research project tries to address this question: can a strong segmentation baseline be matched or beaten on the ISLES task, and does adding uncertainty estimation improve clinical usefulness without hurting accuracy? Pipeline: (1) preprocess the multimodal MRI inputs; (2) baseline with nnU-Net; (3) add uncertainty (example: Monte Carlo dropout, open to research) to produce a per-voxel confidence map; (4) Optional: submit to the Grand Challenge leaderboard.

Resources

Data (public): the ISLES 2026 challenge dataset

Background

Url

External Link

Difficulty Level

High

Ethical Approval

None

Number Of Students

1

Supervisor

Heba Elshimy

Keywords

machine learning, deep learning, segmentation, multi-modal, ai in healthcare, medical imaging

Degrees