Related Experiment Video
Updated: Aug 14, 2025
Automated Midline Shift and Intracranial Pressure Estimation based on Brain CT Images
Published on: April 13, 2013
Development and External Validation of a Deep Learning Algorithm to Identify and Localize Subarachnoid Hemorrhage on
Antonios Thanellas1, Heikki Peura1, Mikko Lavinto1
1From the Department of Information Management (A.T.), Helsinki University Hospital, Helsinki, Finland; Department of Neurosurgery, University of Helsinki and Helsinki University Hospital (H.P., M.K.), Helsinki, Finland; CGI (M.L., T.R.), Helsinki, Finland; Machine Intelligence in Clinical Neuroscience (MICN) Laboratory, Department of Neurosurgery (M.V., V.E.S., S.W., L.R.), Clinical Neuroscience Center, University Hospital Zurich, University of Zurich, Zurich, Switzerland; Department of Neuroradiology (C.S.), Clinical Neuroscience Center, University Hospital Zurich, University of Zurich, Zurich, Switzerland.
A new deep learning algorithm accurately identifies subarachnoid hemorrhage (SAH) on head CT scans, demonstrating high sensitivity in external validation. This tool shows promise for improving SAH diagnosis in medical imaging.
Area of Science:
- Radiology
- Artificial Intelligence
- Medical Imaging
Background:
- Limited external validation and public release of deep learning algorithms in medical imaging.
- Hypothesis that a deep learning algorithm can be trained to identify and localize subarachnoid hemorrhage (SAH) on head CT scans.
Purpose of the Study:
- To train and validate a deep learning algorithm for SAH detection on head CT scans.
- To assess the algorithm's performance on external and real-world data.
Main Methods:
- Training an open-source U-Net convolutional neural network using manually segmented SAH head CT scans.
- External validation using data from two foreign countries and a dataset of consecutive emergency head CT scans.
- Assessing performance using patient- and slice-level sensitivity and specificity metrics.
Main Results:
- High sensitivity (99.3%) in identifying SAH cases in an external validation set of 1,379 cases.
- High slice-level sensitivity (87.4%) and specificity (95.3%) on 49,064 axial head CT slices.
- 100.0% sensitivity and 75.3% specificity in identifying all 8 SAH cases in a consecutive emergency dataset.
Conclusions:
- A shared, high-performing deep learning algorithm effectively identifies SAH cases with high sensitivity.
- The study presents novel approaches for developing, training, testing, and reporting deep learning algorithms in medical imaging.
