Hemorrhagic Brain Strokes Detection Using Recurrent Neural Networks-Based Microwave Imaging Technique

This research aims to use a microwave imaging system, leveraging deep learning, to detect hemorrhagic strokes by extracting effective features from backscattered signals of the head. This method is promising for early diagnosis and avoiding surgical intervention, being low-cost, portable, nonionized, real-time imaging, comfortable and harmless. Microwave imaging detects hemorrhagic strokes by exploiting the contrast in dielectric properties between healthy and unhealthy tissues. The proposed method was simulated using Computer Simulation Technology (CST) software, featuring two opposite antipodal Vivaldi antennas, positioned 20 mm from the head model with a gain greater than 7 dBi in the 2.55 GHz band. The antennas act as transmitters and receivers. The head model was created using a realistic human voxel model, with hemorrhagic strokes simulated by using the electrical properties of blood. The proposed model was experimentally verified and fabricated using chemical materials. Data collected from the model included four time-domain and eight frequency-domain features. Three-layer recurrent neural networks (RNNs) were trained using time features, frequency features, and a combination of both. This approach was successful, achieving 100% accuracy in detecting the presence or absence of strokes for simulated and experimental data and 90% accuracy of a 5 mm stroke with improved discrimination of stroke size and localization when using combined time-frequency features for experimental data while the lowest error in xy plane is 13.69 mm. The results are highly encouraging, supporting the development of portable equipment for brain stroke detection.