Related Experiment Video
Updated: May 15, 2025
10:25
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
47.7K
Impact of spectrum bias on deep learning-based stroke MRI analysis
Christian Hedeager Krag1, Felix Christoph Müller2, Karen Lind Gandrup3
1University Hospital Copenhagen - Herlev and Gentofte, Department of Radiology, Denmark; Department of Clinical Medicine, University of Copenhagen, Denmark; Radiology AI Testcenter (RAIT.dk), Denmark.
European Journal of Radiology
|May 13, 2025
Summary
Excluding uncertain acute ischemic lesions (AIL) in stroke MRI analysis significantly overestimates diagnostic accuracy. Factors like MRI artifacts and lesion characteristics influence uncertainty, requiring careful consideration in validation studies.
Area of Science:
- Medical Imaging
- Neurology
- Artificial Intelligence in Medicine
Background:
- Spectrum bias can affect diagnostic accuracy in medical imaging.
- Accurate identification of acute ischemic lesions (AIL) on MRI is crucial for stroke diagnosis and management.
Purpose of the Study:
- To evaluate spectrum bias in stroke MRI analysis by excluding uncertain AIL cases.
- To identify patient, imaging, and lesion factors associated with uncertain AIL.
Main Methods:
- Retrospective observational study of 989 adult brain MRIs for suspected stroke.
- Uncertain AIL identified by reader disagreement or low certainty grading.
- Deep learning tool assessed for impact on diagnostic odds ratios after excluding uncertain cases.
- Statistical analysis of patient, MRI acquisition, and lesion factors.
Main Results:
- Excluding uncertain AIL (6% of cases) increased the diagnostic odds ratio four-fold (68 to 278).
- A simulated case-control design further amplified this increase (68 to 431).
- MRI artifacts, smaller lesion size, older lesion age, and infratentorial location were linked to uncertain AIL.
Conclusions:
- Excluding uncertain AIL cases leads to a significant overestimation of diagnostic performance.
- MRI artifacts, lesion size, location, and age are key factors contributing to AIL uncertainty.
- These factors must be addressed in future validation studies of AI tools for stroke MRI.
