ISP-Net: Fusing features to predict ischemic stroke infarct core on CT perfusion maps.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Acute ischemic stroke is a leading cause of death, necessitating accurate infarct core delineation for effective treatment.
• Challenges in stroke infarct core estimation include anatomical variability and complex perfusion parameter relationships.

Purpose of the Study:

• To develop a deep learning model, Ischemic Stroke Prediction Network (ISP-Net), for predicting infarct core after thrombolysis treatment.
• To improve the accuracy and efficiency of infarct core estimation in CT perfusion (CTP) imaging.

Main Methods:

• Developed ISP-Net, an encoder-decoder semantic model utilizing fused CTP features (CBF, CBV, MTT, Tmax) and five-path convolutions.
• Introduced a multi-scale atrous convolution (MSAC) block for enhanced feature extraction.
• Evaluated ISP-Net on a retrospective, multi-center dataset with gold standard infarct cores from follow-up NCCT or DWI scans.

Main Results:

• Achieved a mean Dice Similarity Coefficient (DSC) of 0.801, precision of 81.3%, sensitivity of 79.5%, specificity of 99.5%, and AUC of 0.721 in cross-validation.
• Demonstrated superior performance compared to advanced deep learning methods like Deeplab V3, U-Net++, CE-Net, X-Net, and Non-local U-Net.
• Showed no significant difference in prediction error between follow-up NCCT and DWI scans.

Conclusions:

• ISP-Net offers a fast and accurate method for estimating stroke infarct cores.
• The model's predictions can assist physicians in selecting appropriate thrombolysis or thrombectomy therapies.