Related Experiment Video
Updated: Aug 5, 2026
14:08
Automated Midline Shift and Intracranial Pressure Estimation based on Brain CT Images
Published on: April 13, 2013
Towards Explainability in Deep Learning for Detection of Five Major Intracranial Hemorrhage Subtypes on Head CT Using
H Sekkat1,2, A El Haouat3, O El Mouden4,5
1Sciences and Engineering of Biomedicals, Biophysics and Health Laboratory, Higher Institute of Health Sciences, Hassan First University, 26000, Settat, Morocco. sekkat.isss@uhp.ac.ma.
Journal of Imaging Informatics in Medicine
|July 29, 2026
Summary
An explainable AI framework accurately detects intracranial hemorrhage (ICH) on CT scans. This deep learning model enhances rapid diagnosis of neurological emergencies, improving patient outcomes.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neurology
Background:
- Intracranial hemorrhage (ICH) is a critical neurological emergency requiring prompt diagnosis.
- Non-contrast head CT is the standard initial imaging technique for ICH.
- Explainable AI (XAI) offers potential for improving the speed and accuracy of ICH detection and classification.
Purpose of the Study:
- To develop and evaluate an explainable deep learning framework for multi-label ICH detection.
- To utilize DICOM-native, multi-window CT data for enhanced detection capabilities.
- To assess the explainability of the AI model's predictions.
Main Methods:
- A retrospective analysis of the RSNA 2019 Intracranial Hemorrhage Detection Challenge dataset was performed.
- A ResNet34-based deep learning model was trained on DICOM-native, multi-window CT images (brain, subdural, bone).
- Explainability was assessed using Grad-CAM, Integrated Gradients, and other faithfulness analyses.
Main Results:
- The model demonstrated strong and consistent performance in detecting any ICH (ROC AUC 0.973-0.974, PR AUC 0.894-0.899).
- High F1 scores were achieved for common ICH subtypes like intraventricular (0.788) and intraparenchymal (0.767).
- Explainability analyses confirmed a functional association between highlighted regions and model confidence, though rare subtypes showed limited performance.
Conclusions:
- The proposed DICOM-native, explainable AI framework achieves robust intracranial hemorrhage detection.
- The approach shows promise for enhancing rapid diagnosis in time-critical neurological emergencies.
- Further development is needed to address performance limitations for rare ICH subtypes due to class imbalance.