Repository navigation
Description
Title
TRUST-MRI: Uncertainty-aware MRI Segmentation on Imputed data
Leaders
Hadrien Padilla, Siddhesh Salfale, Ninad Aithal, Charlotte Livingston, Axel Cera
Collaborators
No response
Project Description
TRUST-MRI is a trust-aware, multimodal medical imaging framework designed to improve MRI reconstruction, missing-modality imputation, and automated annotation. The framework integrates complementary MRI sequences, such as T1-weighted, T2-weighted, FLAIR, and diffusion imaging when available, to recover missing information and generate segmentation masks while explicitly estimating the reliability of its predictions.
Unlike conventional pipelines that produce reconstructed images or annotations without adequately characterizing their limitations, TRUST-MRI incorporates uncertainty estimation, cross-modal consistency, and adaptive refinement. High-uncertainty regions trigger targeted reconstruction, segmentation refinement, or expert review, helping identify potentially misleading synthetic images and unreliable annotations.
The framework will be evaluated on multimodal MRI datasets using reconstruction quality, segmentation accuracy, uncertainty calibration, error-detection performance, and annotation efficiency. Comparisons against existing imputation and segmentation methods will assess whether trust-guided refinement improves downstream analysis under realistic missing-modality conditions.
Objective: Develop a modular, trust-aware multimodal MRI pipeline that combines missing-modality imputation, automated annotation, uncertainty estimation, and adaptive correction.
Potential contribution: A unified reliability-guided framework that connects MRI imputation and annotation, identifies regions where synthetic or predicted information may be unreliable, and prioritizes them for further analysis or human review.
Link to project repository/sources
No response
Goals for Brainhack Montreal
- Develop a preliminary TRUST-MRI prototype for trust-aware multimodal MRI imputation and automated annotation.
- Integrate uncertainty estimation and cross-modal consistency to identify unreliable reconstructions and segmentation predictions.
- Implement uncertainty-guided refinement to prioritize regions requiring further processing or expert review.
- Evaluate reconstruction quality, annotation accuracy, and reliability using existing multimodal MRI datasets.
- Release a reproducible, open-source pipeline to support further development of trustworthy AI tools for neuroimaging.
Skills
MRI Data Processing: Python, NumPy, NiBabel, and TorchIO for loading, preprocessing, and aligning multimodal MRI sequences. Beginner-friendly: dataset exploration, visualization, and basic preprocessing.
Deep Learning and Imputation: PyTorch and pretrained medical imaging models for missing-modality reconstruction. Beginner-friendly: running pretrained models and comparing reconstructed images with ground truth.
Segmentation and Annotation: MONAI, nnU-Net, SAM/MedSAM, and 3D Slicer for automated segmentation and annotation refinement. Beginner-friendly: reviewing segmentation masks, correcting annotations, and visualizing results.
Uncertainty Estimation: Python and statistical methods to estimate prediction variance, cross-modal disagreement, and voxel-level reliability. Beginner-friendly: implementing simple variance maps and uncertainty visualizations.
Evaluation and Visualization: Computing SSIM, MAE, Dice score, and uncertainty–error correlations; creating figures and comparison tables. Beginner-friendly: benchmarking existing models and documenting results.
Documentation and Collaboration: Git/GitHub, experiment tracking, literature review, and scientific writing. Beginner-friendly: documenting workflows, reviewing related papers, testing code, and preparing tutorials.
Onboarding documentation / Expectations
No response
What will participants learn?
No response
Data to use
No response
Credit to collaborators
No response
Image
No response
Type
pipeline development
Project Maturity Status
0 - project idea
Topic
data_visualisation, deep_learning
Tools
Deep_learning
Programming language
Python
Modalities
fMRI
Git skills
0 - no git skills
Anything else?
No response
Activity
Update project data for Issue #7