Stroke-GFCN: ischemic stroke lesion prediction with a fully convolutional graph network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate interpretation of medical images is crucial for acute brain stroke diagnosis and timely surgical intervention.
• Current automatic segmentation methods for brain stroke lesions often lack clinical reliability.
• Computer Tomography (CT) perfusion parameters offer rich data for stroke lesion analysis.

Purpose of the Study:

• To investigate the segmentation of brain stroke lesions using a novel geometric deep learning model.
• To leverage multi-modal CT perfusion parameters for improved lesion detection.
• To evaluate the model's performance against state-of-the-art methods.

Main Methods:

• Proposed a geometric deep learning model utilizing spline convolutions and graph-based operations.
• Employed unpooling/pooling operators on graphs within a fully convolutional network architecture.
• Conducted experiments evaluating architecture hyperparameters and comparing with existing segmentation techniques.

Main Results:

• Deeper network layers achieved higher Dice coefficient scores (DCS), reaching up to 0.3654.
• The proportional unpooling method adapted better to lesion boundaries, reducing Hausdorff distance.
• The model performed comparably to state-of-the-art methods without advanced training optimizations.

Conclusions:

• The proposed end-to-end trainable fully convolutional graph network effectively predicts ischemic stroke brain lesions from CT perfusion data.
• Geometric deep learning is feasible for complex segmentation tasks, with the proposed model outperforming others.
• The model demonstrates adaptability to irregular lesion boundaries, showing promise for clinical application.