Related Experiment Video
Updated: Dec 11, 2025
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
Integrating uncertainty in deep neural networks for MRI based stroke analysis
Lisa Herzog1, Elvis Murina2, Oliver Dürr3
1University of Zurich, Epidemiology, Biostatistics and Prevention Institute (EBPI), Hirschengraben 84, 8001 Zurich, Switzerland; Zurich University of Applied Sciences, Institute of Data Analysis and Process Design (IDP), Rosenstrasse 3, 8400 Winterthur, Switzerland.
This study introduces a Bayesian Convolutional Neural Network (CNN) for diagnosing ischemic stroke from MR images, quantifying prediction reliability. This approach improves accuracy and identifies critical cases needing further medical review.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neurology
Background:
- Deep Learning (DL) methods for medical image analysis often lack uncertainty quantification.
- Reliability assessment is crucial for clinical decision-making in automated image analysis, especially in stroke diagnosis.
Purpose of the Study:
- To develop a framework for diagnosing ischemic stroke patients using Bayesian uncertainty in Convolutional Neural Networks (CNNs).
- To provide image-level predictions with associated uncertainty measures for Magnetic Resonance (MR) images.
- To evaluate patient-level diagnostic aggregation methods that incorporate image-level uncertainty.
Main Methods:
- Implementation of a Bayesian Convolutional Neural Network (CNN) for stroke lesion detection on 2D MR images.
- Development and evaluation of aggregation methods to combine image-level predictions for patient-level diagnoses.
- Integration of uncertainty information into aggregation models to report patient-level model uncertainty.
Main Results:
- The Bayesian CNN achieved 95.33% accuracy at the image-level, a 2% improvement over non-Bayesian methods.
- The best patient aggregation method reached 95.89% accuracy.
- Integrating uncertainty measures flagged false classifications, enabling identification of critical patient diagnoses for closer medical examination.
Conclusions:
- Bayesian approaches enhance image-level prediction and uncertainty estimation in stroke diagnosis.
- Uncertainty quantification aids in detecting unreliable patient-level aggregations, improving clinical workflow.
- The proposed framework supports more reliable automated diagnosis of ischemic stroke.
