MCA-DN: Multi-path convolution leveraged attention deep network for salvageable tissue detection in ischemic stroke...

Area of Science:

• Medical Imaging
• Computational Intelligence
• Artificial Intelligence in Medicine

Background:

• Accurate and timely treatment of ischemic stroke is critical for reducing disability and death.
• Identifying blood flow restriction in stroke using MRI and computational intelligence aids diagnosis.

Purpose of the Study:

• To propose a novel multi-path convolution leveraged attention based deep network (MCA-DN) for stroke lesion identification and localization.
• To enhance the accuracy of stroke diagnosis by focusing on relevant voxels and prioritizing multi-parametric MRI sequences.

Main Methods:

• Development of MCA-DN, a deep network utilizing multi-path convolution and attention mechanisms.
• The network learns to focus on voxels with enhanced activations and selectively processes multi-parametric MRI sequences.

Main Results:

• MCA-DN achieved high performance in stroke segmentation with a Dice similarity coefficient of 77.3%, sensitivity of 82.8%, and specificity of 98.8%.
• The proposed method outperformed five state-of-the-art methods on multiple datasets, including ISLES-2015 and ISLES-2017.

Conclusions:

• The MCA-DN shows significant potential in assisting patient-specific stroke treatment planning.
• Its competitive performance in estimating reperfusion benefits highlights its value in clinical decision-making.