A multibranch fusion network (MBF-Net) for intracranial large-vessel occlusion classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Large-vessel occlusion (LVO) in acute ischemic stroke (AIS) leads to severe outcomes.
• Accurate diagnosis via computed tomography angiography (CTA) is critical for treatment.
• Existing AI models struggle with imbalanced medical data and the long-tail effect in LVO classification.

Purpose of the Study:

• Develop a multibranch fusion network (MBF-Net) for intracranial LVO subtype classification.
• Address data imbalance and improve diagnostic accuracy in LVO detection.
• Enhance the extraction of spatial and semantic information from CTA images.

Main Methods:

• Proposed MBF-Net with multiple branches to mitigate the long-tail effect.
• Introduced branch hierarchical deep aggregation (BHDA) and semantic information enhancement (SFE) modules.
• Incorporated an auxiliary attention guidance module (AAGM) for training focus and interpretability.

Main Results:

• MBF-Net achieved superior performance compared to state-of-the-art methods.
• Demonstrated high sensitivity (94.44%), precision (89.29%), and F1-score (83.37%) in LVO classification.
• Effectively handled class imbalance, improving overall diagnostic accuracy.

Conclusions:

• MBF-Net offers an advanced, accurate, and interpretable solution for LVO classification in medical imaging.
• The model's robust performance on imbalanced datasets highlights its clinical potential.
• MBF-Net can aid in the rapid and precise diagnosis of LVO, improving patient outcomes.