Stroke Classification With Microwave Signals Using Explainable Wavelet Convolutional Neural Network

Area of Science:

• Medical Imaging
• Signal Processing
• Artificial Intelligence

Background:

• Stroke is a major cause of death and disability globally.
• Microwave imaging offers a portable solution for medical diagnostics.
• Accurate stroke classification from microwave signals remains a challenge.

Purpose of the Study:

• To develop an accurate stroke classification method using microwave imaging.
• To link identified microwave signal features back to the original data for interpretability.
• To propose a wavelet convolutional neural network (CNN) for enhanced stroke detection.

Main Methods:

• A wavelet convolutional neural network (CNN) was proposed, integrating multiresolution analysis with CNNs.
• A game theoretic approach was employed for model explanation and feature identification.
• The algorithm was validated using simulated data and experimental data from head phantoms, incorporating noise and manufacturing tolerances.

Main Results:

• Classification accuracy ranged from 81.7% in 3D simulations to 95.7% in lab experiments.
• The model successfully identified key features for discriminating between ischemic and hemorrhagic strokes.
• Wavelet coefficients within 0.95-1.45 GHz and time slots of 1.3-1.7 ns were found to be significant.

Conclusions:

• The proposed wavelet CNN provides an effective and accurate method for stroke classification via microwave imaging.
• The approach enhances the interpretability of microwave imaging by linking features to specific stroke types.
• This technology holds promise for portable, non-invasive stroke diagnosis.