Convolutional autoencoder-based deep learning for intracerebral hemorrhage classification using brain CT images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Intracerebral haemorrhage (ICH) is a severe stroke type with high mortality.
• Accurate diagnosis via non-contrast computed tomography (NCCT) is vital for surgical decisions.
• Challenges include limited expert access and inter-observer variability in diagnosis.

Purpose of the Study:

• To develop and evaluate a hybrid deep learning model for automated ICH diagnosis using NCCT images.
• To compare the model's performance against Principal Component Analysis (PCA) based methods.
• To assess the model's ability to highlight ICH regions for clinical relevance.

Main Methods:

• A hybrid model combining Convolutional Autoencoder (CAE) for feature extraction and Dense Neural Network (DNN) for classification was developed.
• Tenfold cross-validation and holdout methods were used for robust training and generalization.
• Performance was benchmarked against a PCA-DNN model using a dataset of 3293 NCCT images from 108 patients.

Main Results:

• The CAE-DNN model achieved 99.84% accuracy, 99.69% sensitivity, 100% specificity, 100% precision, and 99.84% F1-score.
• The developed model significantly outperformed the PCA-DNN comparator and existing literature results.
• Saliency maps from the CAE-DNN model effectively highlighted ICH regions, correlating with expert annotations.

Conclusions:

• The hybrid CAE-DNN model offers a highly accurate and reliable method for automated ICH detection and localization from NCCT scans.
• This AI-driven approach has the potential to improve diagnostic efficiency and patient triage in clinical settings.
• The model's ability to localize ICH provides valuable insights for treatment prioritization.