Related Experiment Video
Updated: Jun 7, 2025
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
Structural-based uncertainty in deep learning across anatomical scales: Analysis in white matter lesion segmentation
Nataliia Molchanova1, Vatsal Raina2, Andrey Malinin3
1Radiology Department, University of Lausanne and Lausanne University Hospital, Lausanne, Switzerland; MedGIFT, Institute of Informatics, School of Management, HES-SO Valais-Wallis University of Applied Sciences and Arts Western Switzerland, Sierre, Switzerland; CIBM Center for Biomedical Imaging, Lausanne, Switzerland.
Uncertainty quantification (UQ) in deep learning models for white matter lesion (WML) segmentation reliably indicates prediction errors at lesion and patient scales. Novel UQ measures improve trustworthiness assessment in multiple sclerosis (MS) MRI analysis.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neurology
Background:
- Automated deep learning (DL) tools are increasingly used for segmenting white matter lesions (WMLs) in multiple sclerosis (MS) patients' magnetic resonance imaging (MRI) scans.
- Assessing the trustworthiness of these DL tools is crucial for clinical applications.
- Uncertainty quantification (UQ) offers a potential method to evaluate the reliability of segmentation predictions.
Purpose of the Study:
- To explore UQ as an indicator of DL model trustworthiness for WML segmentation in MS.
- To investigate the relationship between uncertainty measures and prediction errors at voxel, lesion, and patient scales.
- To develop and validate novel UQ measures for lesion and patient scales.
Main Methods:
- Developed novel UQ measures for lesion and patient scales based on structural prediction discrepancies.
- Extended error retention curve analysis for evaluating UQ performance at lesion and patient scales.
- Analyzed UQ performance in both in-domain and out-of-domain settings using a multi-centric MRI dataset (444 patients).
Main Results:
- Proposed UQ measures effectively capture model errors at lesion and patient scales.
- These novel measures outperform methods that average voxel-scale uncertainty.
- UQ at different anatomical scales correlates with specific types of segmentation errors.
Conclusions:
- UQ is a valuable indicator of DL model trustworthiness for WML segmentation in MS.
- Lesion and patient-scale UQ measures provide more meaningful insights into model reliability than voxel-scale averages.
- The developed UQ framework enhances the interpretability and reliability of automated WML segmentation.
