Microwave technology for detecting traumatic intracranial bleedings: tests on phantom of subdural hematoma and numerical simulations

Traumatic brain injury (TBI) is termed “the silent epidemic.” It strikes up to 2 % of the population each year and is the leading cause of death and severe disability among young people [3, 10]. Furthermore, it affects many elderly, who are more vulnerable due to widespread use of anticoagulant medications and lower brain plasticity [31, 33]. Road traffic injury, sports injuries, assaults and combat operations commonly produce TBI, and elderly are often exposed to accidental falls [3, 6, 10, 33]. Intracranial bleedings constitute the most important complication of TBI [3]. Larger bleedings must be evacuated promptly to save the lives of these patients and mitigate injury [3, 34]. Unfortunately, many patients are treated too late [4, 16, 34]. Patients without clear symptoms, such as decreased level of consciousness, are often undertriaged in the prehospital phase [13]. Delay of treatment due to patients initially being transported to a non-trauma center, to be later transferred to a trauma center, causes substantial mortality [13, 16]. Therefore, there is a need for methods for the prehospital setting to predict the need for lifesaving interventions from clinical observations, as demonstrated by Liu et al. [18] using vital sign measurements and a machine-learning algorithm, and to directly detect occult injury such as intracranial bleedings.

The clinical standard for detecting intracranial bleedings is computed tomography (CT) [10]. The main disadvantage with CT is that it is not well suited for field use. Although mobile units with CT have been developed for examining stroke patients [35], it would be advantageous to incorporate devices for detecting intracranial bleedings in ordinary road and air ambulances. One candidate is a recently developed handheld instrument using near-infrared spectroscopy, the Infrascanner [17, 28, 30]. Its principle for detection is to compare the amount of light absorption for the left and right brain hemispheres, by four point measurements on opposite sides of the skull. Strong asymmetry is likely due to a hematoma on the side with largest absorption, since extravascular blood contains tenfold the concentration of hemoglobin compared to intravascular blood [17]. In a multicenter study including 365 patients, of whom 96 were confirmed to have intracranial hemorrhage, detection sensitivity of 88 % and specificity of 91 % were shown [28]. This result applied for bleedings larger than 3.5 mL, closer than 2.5 cm from the surface of the brain. If all bleedings were included, the sensitivity decreased to 69 % [28]. The disadvantages with the Infrascanner are that it cannot give detailed information about location and size of hematoma [28]. It cannot detect bleedings deep inside the brain, such as deep-sited intraparenchymal or intraventricular hemorrhage, due to the limited penetration depth of approximately 2.5 cm [28]. Another possible future candidate is electrical impedance tomography. Oh et al. [24] showed using epidural electrodes placed on exposed cortex of anesthetized rats that somatosensory evoked responses could be recorded with an SNR of >50 at 225 Hz. That study could pave the way for imaging fast neural activity in the brain, which could be used for detecting and monitoring TBI [24]. However, a shortcoming of electrical impedance tomography is that the cranium has electrically insulating properties that limit current penetration and impede deep imaging of the brain without using implanted electrodes [20].

We propose the use of broadband microwave technology (MWT) [27], in conjunction with a diagnostic algorithm, to detect and classify different types of intracranial bleedings. Recently, Persson et al. [26] showed that MWT can distinguish hemorrhagic stroke from ischemic stroke, in the two first proof-of-principle clinical studies. Twenty and 25 patients with acute stroke were included, respectively. The patients were measured with an antenna system worn on the head. The article demonstrated areas under the receiver operating characteristic curves of 0.85 and 0.88, respectively, using a machine-learning algorithm based on singular value decomposition (SVD). At 90 % detection sensitivity for hemorrhagic stroke, the specificity was 65 %, for the second clinical study [26].

MWT can detect bleedings due to the dielectric contrast between blood and brain matter [9, 26]. MWT could suite for field use since cheap, fast and portable systems working in the time domain are emerging [26, 27, 38]. An advantage with MWT is that it employs low-power (\(\sim\)1 mW) microwave signals. MWT has the potential to detect lesions deep inside the brain, since microwaves will penetrate the skull and brain, and transmission signals between antennas on opposite sides of the head can be utilized [26]. Furthermore, a patient could potentially be monitored in real time, to follow-up on the progression of injury. Such monitoring is not appropriate with CT, i.e., due to excessive ionizing radiation. Our long-term goal is to incorporate MWT into road and air ambulances, to increase the precision of triage and decrease the time to treatment for trauma and stroke patients.

In the future, it may also be possible to image the brain and visualize stroke and TBI lesions [11, 12, 14, 19, 21, 22, 32, 36, 37]. Contemporary systems do not provide detailed anatomical images compared to CT, but may still be valuable for detection of intracranial hemorrhage, see, e.g., [21, 22]. Current imaging algorithms are not sufficiently fast to enable prehospital use without use of a priori information [19]. Mobashsher et al. [21] showed that a portable wideband microwave system with a custom unidirectional antenna has potential to visualize intracranial bleedings. They used a realistic phantom fabricated by a 3D printer. The research group has presented several further studies on improvements of image reconstruction to achieve fast and accurate imaging [11, 12, 36, 37]. Recently, Mahmood et al. [19] showed that a priori information from magnetic resonance imaging and automatic segmentation of tissue types can be used to facilitate rapid imaging of intracranial bleedings.

The advantage of a machine-learning approach as compared to imaging is that a classification algorithm can generate a diagnosis in real time without need for physician/operator image interpretation. The disadvantages are that not as detailed information about the lesion is presented, and that an accurate classifier may require a large set of training data. We have shown that MWT has potential to localize traumatic intracranial bleedings using a diagnostic algorithm [5]. That study focused fully on localization of bleedings. It did not evaluate detection accuracy.

In this study, we assess the potential of MWT for detecting traumatic intracranial bleedings, using a microwave helmet and a classification algorithm based on SVD. We focus on subdural hematoma (SDH), which has very high mortality: 50–85 % [10]. It is the traumatic intracranial bleeding that most commonly requires surgical evacuation [10]. We model SDH by constructing a phantom in a human cranium. Furthermore, we perform numerical simulations to mimic expected inter-variability from microwave measurements on an SDH patient population, by modeling SDH of several different sizes, at different locations, in crania of varying sizes, with a layer of cerebrospinal fluid (CSF) of varying thickness.

The classification accuracy was overall high. Figure 5 shows the classification results for the laboratory measurements. The classification accuracy is 100 % for all classes. Table 2 and Figs. 6, 7 show the classification results for the simulated data. The classification accuracy is 82–96 % for all classes, i.e., when the full dataset was used with fivefold cross-validation. The true positive rate (sensitivity) for detection of bleeding (regardless of size) was 100 %, whereas the true negative rate (specificity) was 96 % (Fig. 7). However, a high accuracy required a relatively large training dataset. The classification accuracy dropped when smaller subsets of the data were used (Fig. 7). For both the laboratory and the simulated data, the bleeding size correlates with the subspace distances, i.e., the smaller the bleeding the closer it is to the No bleeding subspace (Figs. 5, 6). Bleedings of sizes close to each other are more similar to each other than sizes far apart. For the few misclassifications of simulated data, the predicted class is a bleeding with a nearby size (Table 2).

The outcome of the classification of the validation dataset based on training on the original simulation dataset is shown in Table 3 and Fig. 8. The classification accuracy is 99 % (Bleeding versus No bleeding), with only a few misclassifications (Table 3). Figure 8 shows that the size of the bleedings in the unseen dataset correlates with the subspace distance to the No bleeding class of the training data, i.e., smaller bleedings are closer to the No bleeding class.

Analysis of plots of raw reflection \(S_{ii}\) and transmission coefficients \(S_{ij}\) revealed that the effect of SDH is small compared to the effects of, e.g., head size, see Fig. 9. No consistent visual trends due to SDH were identified in plots of magnitude and phase for the measurements on the laboratory phantom. Even though plots of the mean magnitude indicate a difference (Fig. 9), the standard deviation for laboratory data is relatively large and obscures any possible trends due to SDH. For the simulated data, it was clear that head size is the parameter producing the largest variability, whereas differences due to bleeding size and position are relatively small (Fig. 9).

This study shows that MWT is a promising technology for prehospital detection of traumatic intracranial bleedings. The classifier distinguished all observations on three sizes of bleedings and no bleeding for the laboratory SDH model (Fig. 5). Furthermore, using a finite element model, we simulated a large number of bleedings with different sizes, at various positions, in crania of different sizes. The classifier identified SDH with high accuracy (Table 2; Fig. 7 ), despite the relatively high variability from varying head size (Fig. 9b, c). Furthermore, classification accuracy was high on an unseen dataset with different values of head and bleeding sizes (Table 3). This indicates that the model was not overfitted to the original dataset.

This conceptual study evaluates the hypothesis that it is possible to detect small signal changes caused by intracranial bleedings in a heterogeneous dataset with high variability, due to inter-individual differences between patients. Results for the phantoms constructed in the human cranium indicate that it is possible to detect SDH using the current microwave system in conjunction with the classifier, with consideration taken to the variability caused by taking the helmet on and off (Fig. 5). However, this does not consider the inter-patient variability. It was not feasible to construct physical phantoms of many different head sizes. Therefore, the simulations were performed to evaluate the impact of inter-patient variability. The results indicate that it is possible to differentiate bleeding sizes (Fig. 6), if a large training dataset is available (Fig. 7).

Bleeding size correlated with the subspace distance to the No bleeding class and bleedings of nearby sizes were more similar to each other compared with bleeding sizes far apart (Figs. 5, 6, 8). We have previously shown that there is good potential for estimating the position of a bleeding [5]. These results combined point to that MWT, in addition to reliable SDH detection, can be of further clinical value by estimating size and position of the lesion. In particular, the potential for size estimation can be valuable for monitoring patients, to detect an expanding bleeding before the patient’s condition deteriorate.

The laboratory model of SDH was made in a human cranium. This was important to make the model more realistic, especially since it was expected that the cranium would attenuate the signal substantially [32]. The SDH model was simplified due to practical reasons. It was not feasible to construct an advanced phantom with a realistic anatomy of the human brain. However, the healthy human brain is symmetric with respect to the left and right hemispheres. Our hypothesis is that the altered scattering pattern of the microwave signals, due to the asymmetry introduced by a lesion, is the main reason for that bleedings can be detected by the classifier. In this regard, the model is suitable. A disadvantage with the laboratory phantom used in this study is that although the bleeding compartments were cut out and filled with bleeding solution in a meticulous manner, it cannot be guaranteed that small pockets of air are not present. Furthermore, the dielectric properties could not be matched closely to published values across the full frequency interval (Fig. 1). In future studies, it would be valuable to confirm the results obtained in this study using improved and more realistic models of the brain, such as [21].

The laboratory data showed relatively large variability/noise (Fig. 9). This was to a large degree caused by taking the helmet on and off. It was done to mimic the clinical scenario, where the helmet position will vary. From Fig. 5, we see that the variability from taking the helmet on and off manifests in the subspace distance. The repeated observations in groups of three are close together, while the difference between the groups is considerably larger. The differences between the SDH models were larger than the differences due to helmet position (Fig. 5). The classification accuracy was high, and the trends for the subspace distances for different classes followed the numerical simulations (Figs. 5, 6). Thus, the classifier seems capable of identifying the pattern due to bleeding, despite that this effect was small relative to other sources of variability/noise, and the apparent strong cross talk between antennas (Fig. 9). This was achieved for a relatively small dataset, as compared to the numerical data where a large training dataset was necessary (Fig. 7). A plausible explanation is that for the laboratory phantom the cranium size and bleeding position were fixed and the effect of bleeding size should then be more readily detectable.

For the numerical simulations, a 2D electromagnetic finite element method was employed to achieve a fast and efficient simulation tool, which described the physics for the measurement setup reasonably well. The model was rather simple for practical reasons. Our aim was not to create simulation data that are highly realistic with respect to patient anatomy, but to create a realistic model of the variability due to head size, bleeding position and bleeding size, and CSF, as will be experienced for future measurements on patients.

We consider this to be a suitable approach to simulate patient measurements in the context of a classification test, since the fundamental physics and numerical method are well known and understood. The main limitations of the simulation model are that the anatomy was simplified and that simulations were performed in 2D, which limits the antenna structures that are possible to realize.

Accurate detection may require that the classification algorithm is trained on a relatively large number of patients (Fig. 7). Further development of the classification method, to realize a high classification accuracy for smaller training datasets, may decrease the number of patients needed in future clinical studies.

This study presents a new method for detection of traumatic intracranial bleedings using MWT in conjunction with a diagnostic algorithm. Through laboratory measurements and numerical simulations, we show that the method has potential for accurate detection of clinically significant bleedings. This is promising for the development of a simple-to-use instrument to recognize patients with occult, severe brain injury already in the prehospital setting.