Improvement of automatic ischemic stroke lesion segmentation in CT perfusion maps using a learned deep neural network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Acute ischemic stroke is a major cause of death and disability globally.
• Accurate segmentation of stroke lesions is crucial for diagnosis and treatment.
• Current methods like CTP map thresholding lack precision (DSC ~50%).

Purpose of the Study:

• To develop a more accurate deep learning technique for ischemic stroke lesion segmentation.
• To improve upon existing machine learning-based segmentation methods.
• To enhance the segmentation of lesions with varying scales and appearances.

Main Methods:

• A novel deep learning model, MultiRes U-Net, was developed based on the U-Net architecture.
• The model was enhanced by incorporating contra-lateral and Tmax images to enrich CTP input data.
• The method was evaluated on the ISLES challenge 2018 dataset.

Main Results:

• The MultiRes U-Net achieved a Dice Similarity Coefficient (DSC) of 68%, a Jaccard score of 57.13%, and a mean absolute volume error of 22.62 ml.
• The proposed method demonstrated improved segmentation accuracy compared to state-of-the-art techniques.
• The enhanced input data improved the robustness of lesion segmentation.

Conclusions:

• The MultiRes U-Net offers a significant advancement in the automated segmentation of acute ischemic stroke lesions.
• This deep learning approach shows promise for clinical application in stroke management.
• Further improvements in segmentation accuracy can enhance patient outcomes.