A Deep Cascade Architecture for Stroke Lesion Segmentation and Synthetic Parametric Map Generation over CT Studies

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Stroke is a leading global cause of death, requiring rapid diagnosis for better patient outcomes.
• Computed Tomography (CT) imaging has limitations in detecting early stroke lesions.
• Accurate and timely identification of stroke lesions is crucial for effective treatment.

Purpose of the Study:

• To develop a novel deep learning model for enhanced stroke lesion retrieval.
• To improve the accuracy of acute stroke lesion identification using multimodal imaging data.
• To refine stroke lesion segmentation through a cascaded training approach with synthetic perfusion maps.

Main Methods:

• Utilized a deep representation autoencoder architecture with additive cross-attention modules.
• Integrated multimodal inputs from CT studies and perfusion parametric maps.
• Employed a cascaded training strategy to generate synthetic perfusion maps for progressive refinement.

Main Results:

• The proposed deep learning method achieved a Dice score of 0.66 and a precision of 0.67 on the ISLES 2018 dataset.
• Outperformed classical techniques in stroke lesion segmentation.
• Demonstrated the effectiveness of multimodal inputs and synthetic data generation.

Conclusions:

• The novel deep representation model shows significant promise for improving stroke lesion detection.
• The multimodal approach enhances the identification of acute ischemic stroke.
• This method supports expert analysis and could lead to faster stroke diagnosis.