Advancing Stroke Classification: A Novel Deep Learning Approach with Confocal Microwave Imaging for Brain Stroke Detection

Traditional methods of stroke diagnosis, such as computed tomography (CT) scans and MRI, face limitations in accessibility, sensitivity, interpretation, invasiveness, and speed. These limitations can delay timely treatment and hinder accurate diagnosis, potentially worsening patient outcomes. Emerging technologies like optical coherence tomography, near-infrared spectroscopy, and biomarkers hold promise for rapid, non-invasive, and portable stroke assessment. Artificial intelligence and machine learning techniques can enhance scan interpretation and personalize treatment plans, while boosting the time required for diagnosis and increase its accuracy. Combining modalities and integrating translational research will accelerate the development and implementation of improved diagnostic approaches for stroke [1-5]. Recently, microwave imaging systems (MIS) is considered as a portable brain scan system [1-5]. The brain MIS objective is detection of cancerous tumors, ischemic or hemorrhage caused by brain injuries, and brain activity surveillance [1, 2]. For this system several main factors significantly affect the imaging performance, e.g., the antenna dimension and its radiation characteristics, image reconstruction methods, post-processing techniques. Several imaging methods used in medical imaging systems were proposed [1-3]. Generally, imaging methods are classified in two main quantitative and qualitative categories. Quantitative methods such as tomography, dielectric constant and region of interest (ROI) are extracted based on recursive methods. The images reconstructed by these methods are time consuming, but has a good spatial resolution. In contrary, qualitative methods, e.g., radar-based methods, are founded on reflection signal delays. As these methods are faster than qualitative methods, these are adopted as real time methods, more suitable for hospital adoption. Machine learning techniques, which are used in MIS, are highly capable of segmentation, clustering and classification [6-15]. A neural network with microwave imaging for a complex data collection forward learning machine is compounded [8]. Additionally, Rekanos [9] have proposed a radial based neural network for proliferated brain position and size estimation into skeleton tissue with microwave imaging. Li et al. [10] have adopted a deep neural network for enhancing the created image. Their deep neural network had been trained to obtain a much better image from created microwave images using the back-projection method as input. Efforts have been made to address the nonlinear electromagnetic inverse problem through iterative techniques. Recently, studies have explored the use of deep learning methods to enhance 2D microwave imaging for breast examinations [11]. Radar-based approaches have also been employed by researchers to investigate machine learning applications in diagnosing breast lesions [12]. Wan et al. [13] introduced an innovative classification technique for automatic diagnosis derived from reconstructed images using microwave tomography. This approach can aid in identifying cancerous tumors within breast tissue. In microwave imaging tomography systems, image reconstruction is based on the characteristics of the dielectric environment. Another classification is founded directly on signal characteristics, without dielectric characteristic reconstruction. E.g., Nanni et al. [15] introduced a novel approach to categorizing intracervical hemorrhage in relation to ischemic stroke. Where the benefits of adopting deep learning for diagnosis and classification of two brain strokes types in the accurate full head phantom model, in multi-static MIS are investigated. The primary objective of this study is to develop and validate a novel deep learning-based classification method to distinguish between ischemic and hemorrhagic strokes using multi-static MIS. Thus, a circular shaped matrix of butterfly antennas is placed around the unhealthy head, so that 0.5 to 6 GHz are covered. Reflected signals are collected by different multi-static channels, then these are used as inputs to the confocal imaging reconstructing algorithm for mapping extracted information in the 2D image. Due to the importance of clustering in distinguishing hemorrhage from ischemic stroke, a new hybrid post-processing method is applied during ROI image creation. Therefore, feature extraction using conventional neural network (CNN) is implemented. Then, support vector machine (SVM) clustering is adopted. One of the challenges is to setup the test setup consisting of the artifact, beam forming and post-processing methods in a MIS. This research aims to address limitations in current diagnostic methods by leveraging innovative machine learning techniques and enhancing the efficacy of MIS for rapid, non-invasive, and precise stroke diagnosis. The proposed imaging system results indicate that using the proposed deep learning method can distinguish the two types of stokes mentioned in radar-based images exhibit significant ambiguity.

In the present section, it is shown that MIS based on radar can be adopted for stroke detection using confocal imaging algorithm and the results are analyzed. It is clear that detecting hemorrhage related to stroke is of utmost importance and is necessarily considered in real time. Thus, after simulation and setting up the designed imaging system, all reflected signals are stored in MATLAB for image reconstruction. This process consists of three main parts; pre-processing, processing, and post-processing. The main objective at the pre-processing stage is to calibrate the reflected signals and eliminate clutter. The processing stage coherently integrates the data based on delay-and-sum (DAS) algorithm. Finally, at post-processing the deep learning algorithm is adopted to determine the stroke type.

As stated, the first stage for image creation for all imaging algorithms is to calibrate the obtained signals or skin artifact removal.

Next, two calibration methods, ideal subtraction and averaging methods, are presented.

3.1 Ideal subtraction method

Ideal subtraction method is a difference calibration method used for removing background by modeling healthy and unhealthy tissues. Then, the reflected signals are subtracted with regard to each other and the calibrated signals will be obtained as Eq. (1) [19]:

$s_m=x_m^{\text {with}}-x_m^{\text {without}}$     (1)

3.2 Averaging method

The averaging method removes the skin artifact by obtaining the average signal emanating from all channels, then this signal is deducted from each channel. For instance, not considering skin artifact for the signal in the m^th channel yields [19]:

$s_m(n)=x_m(n)-\frac{1}{M} \sum_{i=1}^M x_i(n)$     (2)

where, $s_m$ is the signal without the skin artifact, M is the number of signals, and n is the sample counter.

To create an image based on the obtained signals from the first stage, the confocal image reconstruction algorithm DAS and an improved beamformer called as delay-multiply-and-sum (DMAS) is used [20]. Both beamformers work with a slight difference based on the reflected signal displacement in real time to create a correlated signal.

3.3 DAS beamformer

To implement confocal image reconstruction algorithm with M antennas and considering $S_n$ as the ith reflected signal, the energy at each focal position r can be expressed as Eq. (3):

$I(r)=\int_0^{T_w}\left[\sum_{i=1}^M S_n\left(t-\tau_i(r)\right)\right]^2 d t, r=[x ; y ; z]$     (3)

where, $T_w$ is the window length, $T_s$ distance to sample, and $\tau_i(r)$, the ith discrete time delay, that can be calculated as Eq. (4):

$\tau_i(r)=\frac{2 d_i(r)}{v T_s}$     (4)

For which, $v$ is the average wave propagation speed in the brain medium and $d_i(r)$ is the discrete time distance between the ith displacement antenna, $r_n$, to $r$.

$d_i(r)=\left|r-r_i\right|$     (5)

In a multi-static system $M^2$ signals can be recorded, but for calculating the energy characteristics, only $M(M-1) / 2$ signals are needed.

3.4 DMAS beamformer

Another useful beamformer for confocal image reconstruction algorithm is the DMAS beamformer [21]. It consists of multiplied and summed time modified signals, similar to DAS, which are used to calculate the energy at a focal point. The energy at r inside the brain is defined as follows.

$I(r)=\int_0^{T_w}\left[\sum_{i=1}^{M-1} \sum_{j=n+1}^M S_n\left(t-\tau_i(r)\right) S_j\left(t-\tau_j(r)\right)\right]^2 d t$     (6)

For which, $M$ is the number of antennas used for multi-static imaging. In confocal image reconstruction method, the image pixel intensity (brightness) at the nth cell region and direction, $\theta$, are related as $F_i(n)$.

$F_i(n)={ }_{n=1}^N f_i X_i(n) e^{j \phi_i}$     (7)

In Eq. (7), $X_i(n)$ is the received signal from the ith antenna and N is the total number of receiving antennas. Further, to consider medium effects, propagation damping and loss, $f_i$ is added as a weight parameter. To compensate the phase difference due to different paths, $\varphi_i$ is used [22]. In the present work, it is assumed that the wave is propagating in a spherical front inside the brain medium.

Table 2. Comparison of selected calibration and beamformer methods with quality factor metrics

A comparison between pre-processing and processing algorithms by quality factor metrics is made. First, the signals are passed from the ideal subtraction method and then the averaging method, mentioned above. Later, the DAS and DMAS algorithms are used to determine the energy levels. A white Gaussian noise with signal to noise ratio (SNR) 10dB is added to the reflected signals. The evaluation regions are two displaced 20 mm diameter spherical strokes. As shown in Figure 1, one subject is a hemorrhage stroke and the other a vein clogging or an ischemic. Both are located at the same height, i.e. z = 0. The hemorrhage is located at x = 10, y = 60 and the ischemic at x = -30, y = -45. The post-processing parameters for the hemorrhage of interest are presented in Table 2. To compare algorithm performances, the following criteria are considered.

3.5 Quality factor

To express obtained image quality or, in other words, evaluating the imaging performance quality factors are used. Generally, signal to mean ratio (SMR) and signal to clutter ratio (SCR) parameters are used.

SMR is the ratio of the highest reflected energy in the tumor $\left(r e c_T\right)$ to the average model energy ( $r e c_{T. \text {ave}}$ ), expressed as Eq. (8) [23].

$S M R=10 \log \frac{\max \left(r e c_T\right)}{r e c_{\text {T.ave }}}$     (8)

SCR is the ratio of the highest reflected energy in the tumor $\left(r e c_T\right)$ to the maximum energy at other locations $\left(\mathrm{rec}_T^{\prime}\right)$, expressed as Eq. (9) [23].

$S C R=10 \log \frac{\max \left(r e c_T\right)}{r e c_T^{\prime}}$     (9)

With the above definitions, the tumor is correctly assessed when SMR and SCR values are larger than 10 dB and 0 dB, respectively.

With the results presented in the above table, the subtraction calibration method is selected due to simplicity, high speed, and creation of suitable images for the post-processing stage. Finally, to increase the difference between the subject and background or, with other words, reducing clutter effects, and also removing multiple reflections and background clutter residuals the averaging method is adopted. During the processing stage, initially DAS beamformer and then DMAS is adopted to improve the outcome. Although these beamformers are simple, they are very powerful at concentrating energy at the focal points. As the received signal clarity from the calibration stage is suitable, using the mentioned beamformers, will offer high quality images to the post-processing stage.

Based on the aforementioned framework, the first stage is to load the test image. Then, by using CNN to extract a series of features, the processing speed and classification are improved. Based on these inputs, the main challenge is to determine the weight of the convolution layers. To simulate bleeding targets, blood-like substances have been used. While ischemic targets are modeled with materials having an electrical permeability 15% lower than blood.

This classification method extracts a feature vector from each image, on which the classification is made. Specific images are simulated based on different scenarios. The ResNet-50 model has also been used for CNN from deep learning toolbox in MATLAB [28]. After loading the reconstructed images in MATLAB, they are labeled according to the corresponding image types, the number of these categories are shown in Table 3 along with the number of image samples in each category.

Table 3. Input image labels

Different stroke types are defined for each classification group to assess the stroke type for each input image. The program is automated; the test image is taken as an input, then the neural network algorithm is used through deep learning to extract each image feature. The neural network proposed in this study uses the "fc1000" class capable of detecting edges and spots to extract features. In the designed CNN configuration, the first convolution layer is weighted according to the input image. The extracted features are used as inputs to the SVM classification program. Some of these features are used for training (70%) and others for testing (30%) the classification program. A multi-class SVM program is used for classification.

To evaluate the proposed method, a confusion matrix was created (presented as Table 4) based on the three labels of Table 3. The normalized predicted and actual results are presented as the rows and columns of Table 4, respectively. As the predicted and actual labels have a 1 to 1 correspondence, it can be concluded that the proposed method shows a good detection performance, as the system has a low detection error among different classes. There is only a small discrepancy for the third category, which can be solved by adding more related training samples. The proposed algorithm is able to optimally separate normal tissue and stroke. There is an error in separating the ischemic and hemorrhagic strokes, which is due to the small difference between the electrical conductivity of the two stroke types. To reduce error (improve separation accuracy), a large number of samples per category is needed.

Table 4. Confusion matrix

Table 5. Proposed method comparison with other similar published articles

To evaluate the classifier performance, a comparison has been made with other recently published articles, presented in Table 5. It can be seen that the proposed method has comparable performance as compared to other methods. The proposed method is not only more accurate, but the overall parameter performances, such as processing time (which plays an essential role), small number of samples needed, simpler feature vector, and most importantly, the sample preparation, leading to the development of a specialized database, also excel. It should be noted that if the number of images is combined with other imaging methods and the number of samples increases, the detection accuracy can be higher than 97% [16]. This accuracy can also be improved by adding to the number of input images, which can be hard to acquire. Therefore, it can be concluded that the proposed method is suitable to classify brain image objects using microwave technology. The algorithm was performed using the MATLAB R2019b tool on a 3.60 GHz IntelR Core™ i7 processor running Windows 10 Enterprise 64-bit operating system, having a 7856MB NVIDIA graphics processing unit.