Acute ischemic stroke lesion core segmentation in CT perfusion images using fully convolutional neural networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Computed Tomography (CT) is crucial for stroke diagnosis, but automated analysis of acute ischemia is challenging.
• Subtle ischemic changes in CT images hinder accurate quantification by automated methods.
• Existing automated tools struggle with precise lesion segmentation in acute stroke.

Purpose of the Study:

• To develop and evaluate an automated deep learning tool for segmenting acute stroke lesion cores using CT and CT perfusion imaging.
• To improve the accuracy and efficiency of automated stroke lesion detection and quantification.
• To provide a viable alternative to time-consuming Magnetic Resonance Imaging (MRI) for initial stroke assessment.

Main Methods:

• Developed an improved deep learning model for lesion segmentation, incorporating regularized network training, symmetric modality augmentation, and uncertainty filtering.
• Utilized the Ischemic Stroke Lesion Segmentation (ISLES) 2018 challenge dataset for training (94 cases) and testing (62 cases).
• Evaluated contributions via cross-validation and compared performance against state-of-the-art methods on a blind test set via the ISLES 2018 leaderboard.

Main Results:

• The deep learning tool achieved a competitive Dice similarity coefficient of 49% on the ISLES 2018 testing leaderboard.
• The method demonstrated robust performance among top-ranked approaches in the challenge.
• Individual contributions to the model's performance were quantitatively assessed through cross-validation.

Conclusions:

• The automated deep learning tool provides an efficient method for estimating acute stroke lesion core size and location from CT imaging.
• This tool can aid in clinical decision-making by offering rapid assessment without requiring MRI.
• The publicly available tool facilitates further research in automated stroke analysis.