Deep Learning-Based Automatic Segmentation of Ischemic Stroke Lesions in CT Perfusion Imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Ischemic stroke is a primary cause of global disability, necessitating accurate lesion segmentation for effective patient management.
• Current automatic segmentation methods for ischemic stroke in Computed Tomography Perfusion (CTP) maps show limited accuracy, with Dice Similarity Coefficient (DSC) values around 68%.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for enhanced automatic segmentation of ischemic stroke lesions in CTP images.
• To improve upon the accuracy limitations of existing segmentation techniques.

Main Methods:

• A deep learning model inspired by biological systems and brain information processing was designed.
• The network architecture incorporates five graph convolutional layers for feature extraction and classification from CTP images.
• Model performance was validated using the ISLES 2018 database.

Main Results:

• The proposed model achieved a DSC of 75.41% and a Jaccard Index of 74.52%, outperforming traditional methods.
• The model demonstrated robust performance in noisy conditions, maintaining over 60% accuracy at SNR = -4.
• Significant improvements in ischemic stroke lesion segmentation accuracy were observed.

Conclusions:

• Biomimetic-inspired deep learning networks show significant potential for advancing automatic ischemic stroke segmentation.
• The developed model offers a promising solution for more accurate diagnosis and treatment planning in ischemic stroke cases.
• The approach provides a robust method for lesion segmentation, even in challenging, noisy imaging environments.