Deep Learning for Segmenting Ischemic Stroke Infarction in Non-contrast CT Scans by Utilizing Asymmetry.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Non-contrast computed tomography (NCCT) is crucial for acute ischemic stroke (AIS) treatment decisions.
• NCCT's limitations in contrast and signal-to-noise ratio challenge accurate radiologist diagnosis and automated lesion segmentation.
• Accurate segmentation of AIS lesions on NCCT remains a significant hurdle in clinical practice.

Purpose of the Study:

• To develop an advanced segmentation method for ischemic lesions in NCCT scans.
• To enhance the identification and segmentation of ischemic regions by integrating symmetry-based learning with the nnUNet model.
• To improve the diagnostic accuracy and clinical utility of automated AIS lesion segmentation.

Main Methods:

• A novel approach combining a Generative Module (2.5D ResUNet) and an Upstream Segmentation Module within the 3D nnUNet framework.
• Integration of symmetry-based learning principles to improve lesion identification.
• Utilized the AISD dataset (397 NCCT scans) for training, validation, and internal testing, with additional external datasets for validation and performance assessment.
• Incorporated an intensity-based lesion probability function and specific input channels to boost sensitivity and specificity.

Main Results:

• Achieved a Dice Similarity Coefficient (DSC) of 0.6720 and Hausdorff Distance (HD95) of 35.28 on the internal test dataset.
• Demonstrated satisfactory segmentation on the external test dataset with DSC of 0.4891 and HD95 of 46.06.
• Achieved a superior Area Under the Curve (AUC) of 0.991 for classification on the external test set compared to the baseline nnUNet (0.947).

Conclusions:

• Introduced a novel, symmetry-integrated segmentation technique for ischemic lesions in NCCT scans.
• The method shows significant potential for enhancing the accuracy of automated segmentation and classification of AIS lesions.
• This approach could lead to improved clinical decision-making and patient outcomes in stroke care.