Combining unsupervised and supervised learning for predicting the final stroke lesion.

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate prediction of final ischemic stroke lesion volume is vital for treatment planning and assessing salvageable brain tissue.
• Current methods face challenges due to lesion variability and the time-sensitive nature of stroke intervention.
• Automated prediction tools are needed to support clinical decision-making in acute ischemic stroke.

Purpose of the Study:

• To develop a fully automatic deep learning method for predicting the final ischemic stroke lesion extent and location at 90 days.
• To incorporate cerebral blood flow dynamics into the prediction model for improved accuracy.
• To assist physicians in treatment planning by providing reliable stroke lesion predictions.

Main Methods:

• A novel two-branch Restricted Boltzmann Machine was employed to extract specialized features from Magnetic Resonance Imaging (MRI) parametric maps.
• These extracted features were combined with standard MRI maps and processed through a Convolutional and Recurrent Neural Network (CNN-RNN) architecture.
• The method utilized both unsupervised and supervised learning techniques for comprehensive stroke lesion prediction.

Main Results:

• The proposed deep learning model was evaluated on the ISLES 2017 dataset.
• The model achieved a Dice score of 0.38, indicating moderate overlap between predicted and actual lesions.
• Quantitative metrics included a Hausdorff Distance of 29.21 mm and an Average Symmetric Surface Distance of 5.52 mm.

Conclusions:

• The developed deep learning approach demonstrates potential for automatic prediction of final ischemic stroke lesions.
• Further refinement is needed to enhance prediction accuracy and clinical utility.
• This method offers a foundation for advanced AI-driven decision support in stroke management.