LVONet: automatic classification model for large vessel occlusion based on the difference information between left...

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Stroke, a leading cause of death, often results from intracranial vessel occlusion.
• Accurate classification of these occlusions is crucial for effective treatment, but visual similarities pose challenges.
• Current automated methods struggle with variations in vessel appearance and occlusion degrees.

Purpose of the Study:

• To develop an advanced automatic classification model for large vessel occlusion (LVO).
• To leverage hemispheric symmetry for improved detection of intracranial vessel occlusions.
• To enhance the accuracy of automated LVO classification in challenging clinical scenarios.

Main Methods:

• A dual-branch attention module was employed to capture long-range dependencies and focus on vessel-specific features.
• A differential information classification module was designed to fuse inter-hemispheric vessel feature differences.
• A novel cooperative learning loss function was introduced to optimize feature discrimination.

Main Results:

• The proposed model achieved high performance on an intracranial LVO dataset.
• Achieved classification sensitivity of 93.73%, precision of 83.33%, accuracy of 89.91%, and Macro-F1 score of 87.13%.
• Outperformed existing state-of-the-art deep learning models in LVO classification.

Conclusions:

• The developed model effectively identifies occluded vessel regions by adaptively focusing on relevant features.
• The approach demonstrates robustness in handling high inter-class similarity and intra-class variability.
• This method significantly improves the performance of large vessel occlusion classification for stroke assessment.