Apparatus and method for medical imaging

Abstract

A method for medical imaging, the method comprising: (i) receiving scattering data representing single-static or multi-static measurements of electromagnetic signal scattering from tissue of a body part of a subject at multiple different signal frequencies, wherein the electromagnetic signal is emitted from one or more antennas and the corresponding scattered signals are measured by the one or more antennas; (ii) processing the scattering data to calculate electric field power values ​​at each of multiple scattering locations of the tissue within the body part of the subject and for each of multiple frequencies; (iii) for each scattering location, summing the calculated electric field power values ​​at the scattering location for multiple frequencies and multiple antennas to generate an image of the tissue within the body part; and (iv) iteratively updating a model of the tissue within the body part based on a comparison of the model with the generated image until a termination criterion is met, wherein the updated model is output as an image of the tissue within the body part of the subject.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging, and in particular to the use of signal processing and electromagnetic computing techniques to detect the presence and location of abnormalities within tissue and to classify such abnormalities.

[0002] background

[0003] Medical imaging techniques such as ultrasound, computed tomography (CT), magnetic resonance imaging (MRI) and nuclear medicine imaging are extremely powerful techniques for imaging the internal features of the human body, but suffer from many shortcomings that limit their applicability. For example, these techniques require expensive equipment, and are therefore usually unavailable in rural or remote health centers. In fact, according to the World Health Organization (WHO), more than half of the world's population cannot obtain diagnostic imaging. In addition, there is a general need for low-cost and safe imaging systems for detection and continuous monitoring of various diseases. Due to the need to limit exposure to ionizing radiation (such as X-rays), most currently available medical imaging systems cannot be used for frequent monitoring. In addition, the huge and static structure and high cost of MRI and other large medical imaging systems often prevent them from being used to monitor diseases that need to be monitored on a regular and short-term basis, and they are impractical to be used by nursing staff for on-site imaging and assessment purposes.

[0004] In addition, conventional medical imaging tools are usually not suitable for emergency on-site diagnosis. For example, cerebral stroke is one of the main causes of disability and death worldwide. According to the Australian Stroke Foundation, in 2017, 55,831 Australians suffered life-threatening strokes every nine minutes, and if no action is taken, this number will increase to one stroke every four minutes by 2050. Similarly, in the case of brain injury, rapid diagnosis is often essential to saving patients. Severe brain injury includes traumatic and acquired brain injury caused by external forces (such as falls or accidents) or internal events (such as stroke or tumors), respectively. It is well known that patients with brain injury need immediate medical treatment. Since the beginning of brain injury, millions of brain cells die per second, causing permanent damage and death in some cases. Therefore, a fast and portable diagnostic system for rapid on-site diagnosis of such damage is needed.

[0005] Electromagnetic imaging is an attractive technology for medical applications and has the potential to create visual representations of the interior of the human body in a cost-effective and safe manner. From an electromagnetic engineering perspective, the human body is an electromagnetically heterogeneous medium characterized by features and tissues with different dielectric properties. In addition, injured tissue has different values ​​of dielectric properties permittivity and conductivity compared to healthy tissue. When injured tissue with high permittivity values ​​compared to adjacent healthy tissue is exposed to electromagnetic waves at microwave frequencies, most of the waves are reflected back toward the radiation source. Microwave medical imaging systems can be used to transmit electromagnetic waves to an object to be imaged, such as a human head. Microwave signals reflected by damaged tissue in the head (e.g., particularly at the site of hemorrhage or clots in the brain) due to changes in electromagnetic properties are received and measured by the system, and data representing the measured signals can be processed to estimate the location and / or dielectric properties of the abnormality, and generate a two-dimensional or three-dimensional image of the head showing damaged tissue.

[0006] The data processing step plays a vital role in electromagnetic imaging systems. Various imaging techniques have been used to detect medical targets from the measured values ​​of scattered electromagnetic signals. These techniques attempt to estimate the dielectric properties of tissues by solving nonlinear equations (tomography) (nonlinear equations have no unique solution, and the solution may not be continuously dependent on the input data), or use time-domain radar-based techniques to find the location of the target. Due to the time-consuming nature of tomography-based techniques, they are almost exclusively applicable to single-frequency or narrow-band multi-frequency signals, and are therefore not suitable for use in medical emergencies (such as brain injury detection) where rapid diagnosis is required. Alternatively, in radar-based imaging, the scattering profile of the imaging domain is mapped onto a two-dimensional or three-dimensional image. This method is more applicable when using ultra-wideband for high resolution, because the required data processing is simpler and faster than tomography. However, current radar imaging methods such as confocal, space-time beamforming MIST and adaptive beamforming imaging methods utilize processing techniques based on delay superposition (DAS), which are susceptible to the influence of outer layer reflection and inner layer refraction that may lead to false detection. Furthermore, variations in signal penetration through tissue at different frequencies limit the effectiveness of these delay calculations and, therefore, the accuracy of the resulting images. In view of these difficulties, there is a need for fast imaging methods that do not suffer from the non-unique and computationally expensive solutions of tomographic techniques or inaccurate processing due to multiple reflections and refractions of the heterogeneous, multi-layered structure of the human body.

[0007] There is therefore a need to overcome or alleviate one or more difficulties of the prior art or at least provide a useful alternative.

[0008] Overview

[0009] According to some embodiments of the present invention, there is provided a method for medical imaging, the method comprising:

[0010] (i) receiving scatter data representing monostatic or multistatic measurements of electromagnetic signal scatter from tissue of a body portion of a subject at a plurality of different signal frequencies, wherein the electromagnetic signal is transmitted from one or more antennas and corresponding scattered signals are measured by the one or more antennas;

[0011] (ii) processing the scattering data to calculate electric field power values ​​at each of a plurality of scattering locations of tissue within the body portion of the subject and for each of a plurality of frequencies;

[0012] (iii) for each scattering location, summing the calculated electric field power values ​​at the scattering location for the plurality of frequencies and the plurality of antennas to generate an image of tissue within the body part; and

[0013] (iv) iteratively updating the model of tissue within the body part based on a comparison of the model with the generated image until a termination criterion is met, wherein the updated model is output as an image of tissue within the body part of the subject.

[0014] In some embodiments, the measurements are multi-static measurements in which an electromagnetic signal is selectively transmitted from each of a plurality of antennas disposed about the body part and a corresponding scattered signal is measured by each of the plurality of antennas.

[0015] The body part may be a head, and the tissue may include brain tissue of the subject.

[0016] In some embodiments, the method comprises:

[0017] (v) using machine learning to process the subject's biodata to select a base template from a template library as a best match for the subject, wherein the template represents a corresponding model of tissue of the corresponding subject's body part, and the subject's biodata represents at least the subject's age, gender, and weight;

[0018] (vi) processing the selected base template and the measurements of the external dimensions and / or shape of the subject's body part by geometrically transforming the spatial coordinates of the selected template to match the measurements of the subject's body part to generate template data representing a model of the subject's body part tissue.

[0019] In some embodiments, the step of processing the scatter data comprises the following steps:

[0020] (vii) normalizing the scatter data and removing clutter from the scatter data; and

[0021] (viii) Processing the normalized and clutter-removed scattering data to calculate the electric field power values.

[0022] In some embodiments, removing clutter from the scattered data includes determining an average value of the measured electromagnetic signal and subtracting the average value from each signal measurement at each frequency to remove strong reflections and clutter from the scattered data.

[0023] In some embodiments, the method includes calibrating the scatter data by dividing the measured scatter parameters of the body part by the measured scatter parameters of the imaging domain in the absence of the body part and when the imaging domain is filled with a material having dielectric properties matching the medium or an average body part phantom.

[0024] In some embodiments, the method includes classifying abnormal tissue within a body part as hemorrhagic or ischemic by converting a frequency domain signal into a time domain signal and mapping the time domain signal to a corresponding graph, determining the node degree and degree sequence properties of the graph, calculating graph degree mutual information to evaluate the similarity of the graphs, and training a classifier with a training set of graph degree mutual information features and their corresponding class labels, and applying the classifier to a graph calculated for the tissue within the body part of the subject.

[0025] In some embodiments, the method includes comparing signals of corresponding pairs of opposed antennas to identify significant differences between signals in different hemispheres of the subject's brain, the differences being indicative of an abnormality in one of the hemispheres.

[0026] According to some embodiments of the present invention, there is provided a computer-readable storage medium having processor-executable instructions stored thereon, which, when executed by at least one processor of a medical imaging system, causes the at least one processor to perform any one of the above methods.

[0027] According to some embodiments of the present invention, there is provided an apparatus for medical imaging, the apparatus comprising components configured to perform any one of the above methods.

[0028] According to some embodiments of the present invention, there is provided an apparatus for medical imaging, the apparatus comprising:

[0029] (i) an input receiving scattering data representing monostatic or multistatic measurements of electromagnetic signal scattering from tissue of a body portion of a subject at a plurality of different signal frequencies, wherein the electromagnetic signal is transmitted from one or more antennas and the corresponding scattered signals are measured by the one or more antennas;

[0030] (ii) an image generating component, the image generating component being configured to:

[0031] - processing the scattering data to calculate electric field power values ​​at each of a plurality of scattering locations of tissue within the body part of the subject and for each of a plurality of frequencies;

[0032] - for each scattering location, summing the calculated electric field power values ​​at that scattering location for a plurality of frequencies and a plurality of antennas to generate an image of tissue within the body part; and

[0033] - iteratively updating the model of tissue within the body part based on a comparison of the model with the generated image until a termination criterion is met, wherein the updated model is output as an image of tissue within the body part of the subject.

[0034] In some embodiments, the measurements are multi-static measurements in which an electromagnetic signal is selectively transmitted from each of a plurality of antennas disposed about the body part and a corresponding scattered signal is measured by each of the plurality of antennas.

[0035] In some embodiments, the body part is the head and the tissue includes brain tissue of the subject.

[0036] In some embodiments, the apparatus includes a template generator to:

[0037] Processing the subject's biographical data using machine learning to select a base template from a library of templates as a best match for the subject, wherein the template represents a corresponding model of tissue of a body part of the corresponding subject, and the subject's biographical data represents at least the subject's age, gender, and weight; and

[0038] The selected base template and measurements of external dimensions and / or shape of the subject's body part are processed by geometrically transforming the spatial coordinates of the selected template to match the measurements of the subject's body part to generate template data representing a model of the subject's body part tissue.

[0039] Also described herein is a method for imaging tissue of a body part of a subject, the method comprising:

[0040] (i) using machine learning to process the biographical data of the subject to select a template from a template library as a best match for the subject, wherein the template represents a corresponding model of the tissue of the body part of the corresponding subject, and the biographical data of the subject represents at least the age, gender and weight of the subject;

[0041] (ii) processing the selected template and the measurements of the external dimensions and / or shape of the subject's body part by geometrically transforming the spatial coordinates of the selected template to match the measurements of the subject's body part to generate template data representing a model of the subject's body part tissue;

[0042] (iii) receiving scattering data representing multiple static measurements of microwave scattering from tissue of a body portion of a subject at a plurality of different signal frequencies;

[0043] (iv) normalizing the scattering data and removing clutter from the scattering data;

[0044] (v) processing the normalized and clutter-removed scatter data to calculate an electric field power density at each of a plurality of scattering locations of tissue of the body portion of the subject and for each of a plurality of frequencies;

[0045] (vi) for each scattering location, summing the electric field power densities calculated at the scattering location at a plurality of frequencies to generate an image of tissue of the body portion of the subject; and

[0046] (vii) iteratively updating the model of tissue of the subject's body part based on a comparison of the model with the generated image until a termination criterion is met, wherein the updated model is output as an image of at least the tissue of the subject's body part.

[0047] In some embodiments, the subject's body part is the subject's head.

[0048] Also described herein is a method for imaging brain tissue of a subject, the method comprising:

[0049] (i) using machine learning to process the biographical data of the subject to select a template from a template library as a best match for the subject, wherein the template represents a corresponding model of the head tissue of the corresponding subject, and the biographical data of the subject represents at least the age, gender and weight of the subject;

[0050] (ii) processing the measurements of the external dimensions and / or shape of the selected template and the subject's head by geometrically transforming the spatial coordinates of the selected template to match the measurements of the subject's head to generate template data representing a model of the subject's head tissue;

[0051] (iii) receiving scattering data representing multiple static measurements of microwave scattering from brain tissue of the subject at a plurality of different signal frequencies;

[0052] (iv) normalizing the scattering data and removing clutter from the scattering data;

[0053] (v) processing the normalized and clutter-removed scatter data to calculate an electric field power density at each of a plurality of scattering locations of brain tissue of the subject and for each of a plurality of frequencies;

[0054] (vi) for each scattering location, summing the electric field power densities calculated at the scattering location over a plurality of frequencies to generate an image of brain tissue of the subject; and

[0055] (vii) iteratively updating the model of tissue of the subject's head based on a comparison of the model with the generated image until a termination criterion is met, wherein the updated model is output as an image of at least the subject's brain tissue.

[0056] Also described herein is a method for medical imaging of a subject, the method comprising:

[0057] (i) receiving scattering data representing multiple static measurements of electromagnetic signal scattering from tissue of a subject at a plurality of different signal frequencies;

[0058] (ii) processing the scattering data to calculate an electric field power density at each of a plurality of scattering locations of the subject's tissue and for each of a plurality of frequencies;

[0059] (iii) for each scattering location, summing the electric field power densities calculated at the scattering location over a plurality of frequencies to generate an image of the subject's tissue; and

[0060] (iv) iteratively updating the model based on a comparison of the model of the subject's tissue with the generated image until a termination criterion is met, wherein the updated model is output as the image of the subject's tissue.

[0061] In some embodiments, the tissue comprises brain tissue of the subject. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Some embodiments of the invention are described hereinafter by way of example only with reference to the accompanying drawings, in which:

[0064] Figure 1 is a schematic diagram of an electromagnetic medical imaging system according to an embodiment of the present invention;

[0065] Figure 2 yes Figure 1 A block diagram of the analysis components of the system;

[0066] Figure 3 is a flow chart of an electromagnetic medical imaging method according to an embodiment of the present invention;

[0067] Figure 3a yes Figure 3A block diagram of a calibration model of a calibration process of an imaging method;

[0068] Figure 3b is a schematic diagram of an antenna array filled with a coupling medium used in a calibration process;

[0069] Figure 4a and 4b yes Figure 3 A flow chart of a template generation process of an imaging method;

[0070] Figure 5 It is shown in Figure 3 A flowchart of generating a fused image from a patient-specific head template in an imaging method;

[0071] Figure 6 is a flow chart of the MRI to EM conversion process of the imaging method;

[0072] Figure 7 yes Figure 3 A flowchart of an image generation process of an imaging method;

[0073] Figure 8 yes Figure 2 A functional block diagram of a beam imaging component of an analysis component;

[0074] Fig. 9 yes Figure 7 A flow chart of the beam imaging process of the image generation process;

[0075] Fig.10 is a schematic diagram showing the transmission of a microwave signal from a transmitting antenna into an imaging domain, the scattering of the microwave signal in the imaging domain, and the detection of the scattered signal by a receiving antenna;

[0076] Figures 11 to 19 The different parts or sub-processes of the classification process of imaging methods are shown as follows:

[0077] Fig.11 is a flow chart of the data processing and analysis steps of the classification process prior to classification and localization;

[0078] Fig.12 yes Fig.11 Flow chart of the classification after the steps;

[0079] Fig.13 yes Fig.11 Flowchart of the subsequent positioning steps;

[0080] Fig.14 is a graph of the time series signal generated by the system;

[0081] Fig.15 is a graph of the frequency domain signal generated by the system;

[0082] Fig.16 a and 16b are the graphs of microwave signal and its complex network graph respectively;

[0083] Fig.17 is a schematic diagram showing the differential maximum weight of the signal across the stroke target;

[0084] Fig.18 It is a flow chart of the positioning step of the classification process;

[0085] Fig.19 Includes images of the hemorrhage model of the Duke Stroke brain model;

[0086] Fig. 20 The image fusion process of the imaging method is shown; and

[0087] Fig.21 is a schematic cross-sectional plan view of a subject's head surrounded by antennas, showing the process steps for identifying which hemisphere of the subject's brain contains a feature of interest (e.g., a blood clot): two candidate intersection points (L0, R0) are in the left and right hemispheres, respectively; the 5th antenna that crosses L0 and R2 antennas to R1 is an adjacent antenna far away from L0 and R0; on the other hand, the 13th antenna that almost crosses R0 to L1 is spaced apart from the 5th antenna, and the L2 antenna is far away from L0 and R0.

[0088] Detailed Description

[0089] The electromagnetic imaging systems and methods described herein are capable of rapidly generating images of internal structures within a subject in a safe manner. Although the described systems and methods have been developed for the primary purpose of detecting and characterizing internal injuries and / or diseases in human subjects (particularly within the human brain) and are described herein in this context, it should be understood that many aspects of the systems and methods may be applied to other parts of the human body, other living organisms, and even non-living or inanimate objects.

[0090] like Figure 1 As shown, a system for medical imaging and, in particular, for detecting brain injury includes an antenna assembly or platform 102 including an antenna array, a vector network analyzer (VNA) or transceiver 104, and an analysis component 106. In the depicted embodiment, Figure 1As shown, the antenna assembly or platform 102 is placed on the head of a human subject whose brain is to be imaged. The antenna platform 102 includes an array of antennas arranged around the subject so that each antenna of the array can be selectively excited to radiate a signal of an electromagnetic wave or microwave frequency into the subject's head and through the subject's head to be scattered, and the corresponding scattered signal is detected by all antennas of the array, including the antenna that transmits the corresponding signal. For ease of reference, the entire process of sequentially causing each antenna of the array to transmit a corresponding microwave signal and using all antennas to receive the corresponding scattered signal is referred to as "scanning".

[0091] As will be apparent to one skilled in the art, a vector network analyzer (VNA) 104 excites the antennas as described above, and records the corresponding signals from the antennas as data (referred to herein as "scattering" data) representing the amplitude and phase of the scattered microwaves in the form of what are referred to in the art as "scattering parameters" or "S parameters." The VNA 104 sends this data to an analysis component 106 for processing to generate images of the internal features of the imaged object (e.g., brain clots, hemorrhage sites, and other features), and to classify these features (e.g., as brain clots or hemorrhage sites). In the described embodiment, a VNA having a large dynamic range of greater than 100 dB and a noise floor of less than -100 dBm is used to excite the antennas to transmit electromagnetic signals over the entire frequency band of 0.5 to 4 GHz, and to receive scattered signals from those antennas.

[0092] Although the analysis component 106 (eg Figure 1 ) is in the form of a computer, but this need not be the case in other embodiments. Figure 2 As shown, the analysis component 106 is as follows Figure 2 The 64-bit Intel architecture computer system shown, and the processes performed by the analysis component 106 are implemented as programming instructions of one or more software modules 202, which are stored on a non-volatile (e.g., hard disk or solid state drive) storage device 204 associated with the computer system. However, it will be apparent that, for example, at least some portions of these processes may alternatively be implemented as configuration data for one or more dedicated hardware components, such as an application specific integrated circuit (ASIC) and / or one or more field programmable gate arrays (FPGA).

[0093] The analysis component 106 includes standard computer components, including a random access memory (RAM) 206, at least one processor 208, and external interfaces 210, 212, 214, all interconnected by a bus 216. The external interfaces include a universal serial bus (USB) interface 210 (at least one of which is connected to a keyboard 218 and a pointing device such as a mouse 219), a network interface connector (NIC) 212 that connects the analysis component 106 to a communication network such as the Internet 220, and a display adapter 214 that is connected to a display device such as an LCD panel display 222.

[0094] The analysis component 106 also includes a plurality of standard software modules 226-230, including an operating system 224, such as Linux or Microsoft Windows. In the depicted embodiment, the analysis component 106 also includes web server software 226 such as Apache available at http: / / www.apache.org, scripting language support 228 such as PHP available at http: / / www.php.net or Microsoft ASP, and Structured Query Language (SQL) support 230 such as MySQL available from http: / / www.mysql.com, which allows data to be stored in and retrieved from a SQL database 232.

[0095] Web server 226 , scripting language module 228 , and SQL module 230 together provide analysis component 106 with a general capability to allow Internet 220 users having standard computing devices equipped with standard web browser software to access analysis component 106 , particularly to provide data to and receive data from database 232 .

[0096] However, those skilled in the art will appreciate that the specific functionality provided by the analysis component 106 to such users is provided by scripts accessible to the web server 226 (including one or more software modules 202 that implement the process) and also any other supporting scripts and data 234 (including markup language (e.g., HTML, XML) scripts, PHP (or ASP) and / or CGI scripts, image files, style sheets, and the like).

[0097] like Figure 3As shown, the system described herein performs a medical imaging method, which is also referred to herein as a method for brain imaging. The method begins by generating calibration data for the antenna platform 102 in step 302. The image reconstruction and classification process performed by the system relies on high-fidelity data collected from the antenna. However, in any real situation, the system uses radio frequency (RF) cables with different lengths and curved profiles connected to different adapters and antenna baluns. In addition, due to: 1) manufacturing variations / errors, and 2) unstable dielectric properties of the medium in which the antenna is embedded, the antenna itself is always slightly different. Such uncertain changes among different RF channels (including cables, adapters, and antennas) within the system introduce signal differences in amplitude and phase, and ultimately produce measurement errors that reduce detection, resolution, and reliability of imaging reconstruction / classification.

[0098] To mitigate the effects of these variations and distortions, the system includes a frequency domain calibration component and a corresponding data calibration process to eliminate or at least reduce measurement errors. In practice, the calibration process is ideally performed just before each scan of the subject, but may optionally be performed less frequently, such as only at the beginning of each day.

[0099] Assuming that the VNA 104 is well calibrated at all its N ports, the calibration model of the system uses a specific transmission line error model and antenna uncertainty matrix, such as Figure 3a The transmission line error model on the left represents the coaxial cables and cascade adapters from each VNA port to the corresponding antenna element. Because the coaxial lines (and adapters) have good transmission quality with negligible insertion loss, the largest differences among these cables are mainly caused by their bending profiles and effective physical length errors (ΔL i ) is generated. Assuming that the expected length of all cables is L0, the effective length of the cables (including adapters) in channel i can be expressed as L0+ΔL i (ΔL i <<L0). At the same time, considering that the antenna elements and their baluns used in the same channel have a certain degree of phase delay and gain uncertainty, as shown by the frequency-dependent gain G i (f) Directional point source cascaded frequency-dependent phase shifter Because the directional point source in this model is co-located with the antenna phase center, it does not contribute any phase shift alone.

[0100] In this model, each channel has three error parameters, the first error is caused by the coaxial cable and adapter (ΔL), and the other two errors are Due to the antenna elements and their baluns. Before implementing the image reconstruction / classification process described below, the following calibration process determines these error parameters for the purpose of correcting the measured S-parameters.

[0101] Step #1: For each RF channel, place a reflective object in direct contact with the antenna element to render its radiating aperture ineffective and create a short-circuit reflection. (For aperture-based antennas, such as open-ended waveguides, a short-circuit configuration can be achieved with an ideal E boundary condition overlapping the designated radiating aperture; for resonance-based antennas, such as dipoles or bowtie antennas, a short-circuit configuration can be created by bonding a via directly connected to its feed segment, making the antenna non-resonant).

[0102] Then, record the S from the N-port VNA 104 ii DataS 11 ,S 22 ,…,S NN .in this case,

[0103]

[0104] make We have

[0105] Step #2: Insert a homogeneous and low-loss coupling medium into the imaging domain, such as Figure 3b The shape of the coupling medium should match the shape of the antenna array, and its permittivity value should be close to the average value of the imaged object, which in the described embodiment is a human head. Then, an N-port VNA 104 is used to record S ij (f)(i≠j) data. In this case, we have:

[0106] S ij (f) = αG i (f)G j (f)U ij (f)

[0107] S jk (f) = αG j (f)G k (f)U jk (f)

[0108]

[0109] Where U(f) represents the coupling coefficient of the antenna elements in a specific propagation model, which is independent of the characteristics of the individual antennas. It is a different concept from the mutual coupling of antennas. The value of the coupling coefficient ratio (|U ij (f) / U jk(f)|) can be determined analytically and by assuming that the gains of all point sources are the same (G1(f) = G2(f) = ... = G N (f)) is calculated in an ideal simulation environment, in which case:

[0110]

[0111] in are the coefficients determined from simulations.

[0112] The calibration process mentioned above determines the phase and amplitude differences between the different RF channels in the system. If channel 1 is named the reference channel, the following transform function is used to correct the measured data used by the post-processing algorithm.

[0113]

[0114] After the calibration data for the antenna array is generated at step 302, the system can be used to scan an object to be imaged. To facilitate this process, a template generation process 400 as shown in FIG4 is performed to generate an input template 416 of the object through machine learning.

[0115] Template generation (400)

[0116] The template generation process 400 estimates internal information about an object based on external and known characteristics of the object and structures learned from a large number of samples of MRI and / or CT based information. The purpose of this is to provide a map of the spatial distribution of different head tissues in the described embodiment. As shown in Figure 4, the template generation process 400 begins in step 402 by acquiring data representing the surface of the head and biographical data of the subject whose head is to be scanned by the system. In particular, the age, gender, and weight (in addition to height, if available) of the subject are determined by the user and entered into the system. In step 406, the system generates subject head data representing the external shape of the subject's head. In the described embodiment, this is accomplished by using the 3D camera 108, such as Figure 1 As shown, to generate a three-dimensional (3D) surface of the subject's head. A wig cap or the like can be used to hold the subject's hair down. Alternatively, the antenna array itself can be used to measure the outer boundaries of the subject's head to estimate the subject's head data.

[0117] At step 408, the system uses the subject's biographical data and the 3D scanned surface of the subject's head to provide a base template. Use one of the following three options to select:

[0118] 1. Use a previous scan of an actual subject's head as a template;

[0119] 2. Use the template library and find the head closest to the current subject through nearest neighbor (based on the patient biographical data and the head surface); or

[0120] 3. Use machine learning to directly predict the most likely head (based on patient resume data and head surface).

[0121] In option 1, a previous brain scan of the subject is used. The scan data is converted into a suitable electromagnetic format by performing tissue segmentation and then tissue property assignment (assigning values ​​of permittivity and conductivity to the different tissues identified by tissue segmentation).

[0122] In option 2, the system automatically selects a corresponding head template from a database of head templates, which is either stored locally on the system or accessed remotely via the Internet, so that the stored head template whose biographical data and surface most closely matches the subject's biographical data is selected as the subject's initial head template. In the described embodiment, this is done using standard machine learning methods known to those skilled in the art, such as regression or nearest neighbor classification.

[0123] In option 3, the developed machine learning model takes the head surface and resume data as input and directly learns and predicts the template head without a template library. This works by training a machine learning model such as a deep neural network to learn the values ​​that the head can take and how these values ​​change based on the input. This can be achieved through a feed-forward deep neural network or through a generative model such as a generative adversarial network (GAN) or a variational autoencoder (VAE).

[0124] Template preparation

[0125] In the described embodiment, the head template is generated from MRI data, as this provides a large data source for modeling as well as high resolution tissue distribution. In MRI, different intensity values ​​correspond to different tissues based on their magnetic properties. This needs to be converted into a usable format for electromagnetic imaging, and the MRI to EM conversion 404 is performed as follows:

[0126] 1) First, MRI / CT data is segmented into different tissues. The direction and intensity of the scan are first standardized (e.g., by bias field correction), and then the brain is segmented into brain regions and non-brain regions. Next, based on intensity values ​​and using multi-tissue segmentation algorithms such as fuzzy C-means clustering or FAST, brain regions are segmented into three tissue types (gray, white, and CSF (cerebrospinal fluid)), and similarly non-brain regions are segmented into skull regions and skin regions. Deep neural network methods, such as 3D U-Net, can also be used, and multiple modes can be utilized to improve segmentation.

[0127] 2) Assign tissue values ​​to different tissues based on a table of dielectric properties at the expected frequency of imaging.

[0128] Each template represents a two-dimensional or three-dimensional model of a human head, defining the spatial and dielectric properties (permittivity and conductivity values) of the various tissues of the head, including the skin, skull, cerebrospinal fluid (CSF), white matter and gray matter of the brain. This effectively forms a map of pixels (2D) or voxels (3D) with values ​​corresponding to the dielectric properties (permittivity and conductivity) of the corresponding tissues.

[0129] After determining the three-dimensional shape of the subject's head and the corresponding base head template, the selected base head template is transformed or "warped" so that its outer surface matches the measured shape of the subject's head in step 408. In the described embodiment, this is first accomplished by warping landmarks between the base head template and the real patient's head surface according to the following surface-based registration process.

[0130] First, the location of the patient's anatomical landmarks or "fiducials": the nasion (NAS) and the left and right preauricular points (LPA and RPA, respectively) are determined. This can be done using a digitizer or by manually placing markers at these three locations and using computer vision-based marker detection, or they can be manually marked using the system's computer interface. These anatomical landmarks or "fiducials" enable the scanned surface to be transformed into standardized coordinates, where positive X, Y, and Z will pass through the NAS, LPA, and top of the head, respectively.

[0131] Landmark localization: Instead of following the popular configuration of 10-20 or 10-5 used in EEG, a denser set of landmarks is generated, where landmark points are extracted based on fiducials and the 3D head surface. Several planes are created from the fiducials, and the intersections between the planes and the head surface give the landmark positions. The number of landmark points chosen depends on the desired density and accuracy, and generally a larger number of points allows for higher accuracy. This process of landmark localization is performed automatically.

[0132] Warping is the process of mapping from the measured surface represented via landmarks to an example of the head in order to predict the anatomical structure of the measured head. In the described embodiment, this is done using thin plate spline (TPS) warping, although in other embodiments elastic warping methods can be used as an alternative. The resulting data is then resampled to a uniform grid and converted into a suitable format for the imaging algorithm (e.g., into a CAD model).

[0133] In order to facilitate the use of the template with data generated from an actual subject, consistent coordinates are required between the template and the patient (i.e., the voxels need to be at the same location in 3D space so that they can be directly compared). In the described embodiment, this is achieved by matching the antenna sensor positions to the fiducials (using the 3D camera 108).

[0134] The machine learning model of "Option 3" is iteratively trained to predict heads, and there is an associated error or cost function that needs to be optimized. This cost is the difference between the predicted template of the machine learning model and the real patient head. A good model will predict a model with a low cost, i.e. it predicts a template head that is very similar to the real head. This model can then be applied to new, unseen heads that were not in the training dataset.

[0135] The generation of an accurate head template as described above allows the imaging and classification processes described herein to generate clearer and higher quality images of the brain and significantly improve the accuracy of stroke diagnosis.

[0136] Back to Figure 3 In the flowchart of FIG. 306 , the subject's head is scanned by the system. That is, the analysis component 106 instructs the VNA 104 to sequentially excite the antennas of the antenna array inside the antenna platform (head-mounted device) 102 so that they generate electromagnetic waves in the form of microwave signals and emit the electromagnetic waves into and through the subject's head so that they can be scattered by features in the subject's head, and the scattered signals detected by each antenna of the antenna array inside the antenna platform 102 and the corresponding signals are sent to the VNA 104 for processing to generate corresponding S parameters.

[0137] The detection capability, image resolution, and accuracy of the systems and processes described herein depend on the performance of the antenna array, and the compactness and portability of the system depend on the size and weight of the antenna array and the antenna assembly 102, of which the antenna array is a part. The antenna array is an array of individual antenna elements that have broadband and unidirectional electromagnetic properties, have a high front-to-back ratio (FBR), while also being compact and having a low profile. These electromagnetic properties ensure a high penetration depth of EM waves into tissues of the human head, resulting in a high signal-to-noise ratio (SNR), which enhances the useful reflected signals from features of interest (e.g., stroke) inside the brain. The antenna assembly 102 and its imaging antenna array described below meet these requirements.

[0138] The antenna array may be provided in any of several forms, including as a flexible array or a matched loaded waveguide array.

[0139] In order to avoid strong clutter effects in the received microwave signal, which may cause artifacts in the final image, mutual coupling of antennas and head skin / hair reflections are minimized or at least reduced by providing the antenna in the form of a dielectric loaded conical or wavy waveguide antenna.

[0140] The antennas have appropriate flexible coupling pads between their ends and the imaged head. In addition, they should be partially or fully surrounded by a lossy decoupling medium. For optimal matching between the antenna and the head, the coupling medium inside the antenna and the pad between the antenna and the head are formed of a mixed material (epoxy resin plus aluminum powder or polyester resin plus aluminum powder for the coupling medium and cyano resin CR-V plus ceramic powder for the flexible pad) to have appropriate dielectric properties (dielectric constant of 40-50 and loss tangent less than 0.1 at 0.7 GHz), while the decoupling medium surrounding the antenna is composed of a mixed material (cyano resin CR-S or cyano resin CR-U with graphite powder and carbon powder) in order to have a high dielectric constant greater than 60 and a high loss tangent greater than 1 at 0.7 GHz. The entire antenna array should be surrounded by an absorber to protect the system from the surrounding environment and absorb any stray scattered fields from the imaged head. The antenna array can be fixed or movable. In the latter case, the imaging field will be variable in size and in this case suitable tools such as position sensors should be used to determine the exact position of the antenna as required for the data processing and imaging process.

[0141] For successful image reconstruction, the relative positions of the antennas arranged around the subject's head need to be known to determine the corresponding time delays of the scattered microwave signals. When the head-mounted device is worn by the subject, the antenna array generally becomes conformal to the subject's head, but the antennas can still have different orientations and distances relative to the subject's head. In view of this, the system uses the 3D contour of the subject's head obtained from step 406 of the template generation process and the known antenna array configuration to calculate the spatial position and orientation of each antenna, in particular, the excitation port of each imaging antenna, and this constitutes the prior information provided to the s-parameter processing described herein.

[0142] The process of determining the relative distance and orientation of the antenna array includes using the reference points found in step 406, particularly the center point, i.e., Nasion, to align the head mounted device to the correct position before the system is operated. Because the information of the antenna configuration in the head mounted device is known, by using the 3D head surface acquired in step 406, the relative distance of each antenna and its orientation toward the subject's head is determined.

[0143] Image Generation 306 / 308

[0144] After the head has been scanned at step 306, the resulting scatter data is processed to generate one or more two-dimensional or three-dimensional images of the patient's head using an image generation process 308 that combines the benefits of radar beamforming and tomography and that the inventors therefore refer to as "beam imaging." In particular, the use of this new process reduces the time required to process the scatter data and generate images, which is an important benefit in emergency situations.

[0145] As an overview, the image generation or "beam imaging" process 308 performed by the system utilizes the measured multi-static scattering parameters of the imaging antennas surrounding the imaged object to generate a two-dimensional or three-dimensional image. It also utilizes the measured scattering parameters of the antennas to calibrate the signals measured from the subject's head when the imaging domain is either filled with a matching medium or contains a phantom simulating a normal human head.

[0146] Mean trace subtraction, where the mean measured value (over all signals) is subtracted from each signal at each frequency sample, is used to remove strong reflections and clutter from the scatter data. As a priori information, the effective dielectric properties of the wave propagation paths inside the imaged field as seen by each antenna relative to each voxel of the imaged field are estimated using shortest path finding techniques and an estimated permittivity model (template) of healthy subjects.

[0147] The processed data is then virtually back-propagated to the imaged domain by taking into account the estimated dielectric properties of the object to estimate the power intensity of the tissue inside the domain. A spatial distribution of scattered power intensity within the imaged domain is created at each frequency step (also referred to herein as a "map" or "profile" for convenience) using the estimated dielectric properties from the healthy template and a simplified Green's function. The spatial distribution of power intensity for different frequencies is then superimposed to generate an image that reveals the abnormality (if any) in the imaged domain. In order to determine the dielectric properties of the abnormality, an iterative calculation process is used to update the dielectric permittivity value of the healthy template, and the updated properties are used to recalculate the power intensity. The comparison of power intensities for each pair of consecutive iteration steps and the updating of the permittivity map continue until a suitable accuracy criterion is met (e.g., when the difference in values ​​between consecutive iterations is less than a threshold).

[0148] a) Calibration 702

[0149] The image generation or beam imaging process 308 is now described in more detail. Figure 7As shown, the image generation process 308 begins at step 702 by calibrating the measured scattering parameters of the subject's head using the measured scattering parameters of the imaging domain; that is, in the absence of the subject's head and in the case where the imaging domain is filled with a material having dielectric properties that match the dielectric medium or a common head model. The calibration process also mitigates the effects of any mismatch between antennas due to, for example, antenna manufacturing and assembly variations or errors.

[0150] By transforming the scattering parameter S from the subject's head in the imaging field mn Divided by the scattering parameter from the background matching medium (or calibration object, such as a simulation head) To obtain calibrated data as follows:

[0151]

[0152] Where m and n are the receiver and transmitter indices, and Na represents the number of imaging antennas.

[0153] In some embodiments, the logarithm of the data is used, as follows:

[0154]

[0155] Using the logarithm of the scattering parameter allows the amplitude and phase difference of the data to be compared rather than its real and imaginary parts. In this way, phase change information useful for the retrieval of the dielectric properties of the subject's head is preserved. In addition, the scattering data from different antennas becomes symmetric and suitable for clutter removal in the next step of the imaging process.

[0156] b) Clutter removal 704

[0157] The contrast between the dielectric properties of the outer layers of the subject's head (or other medical object to be imaged) causes a large portion of the waves to be reflected back to the transmitting antenna, and these reflections are typically strong enough to mask reflections from deeper tissues. Therefore, reflections from the outer layers should be removed or at least significantly reduced to allow detection of features of interest within the imaged object. For example, reflections can be reduced by any suitable method, including standard methods known to those skilled in the art, such as mean trace subtraction, spatial filtering, subspace projection, differential methods, or hybrid clutter suppression. In the described embodiment, a combination of mean subtraction and differential methods is used in step 704 to remove strong reflections or "clutter" from the calibrated data, as follows.

[0158] If the distance between each antenna and the outer layer of the object (i.e. the skin of the subject's head) is constant, then the contribution of the signal caused by external reflections will be similar for all antennas, and these reflections can be separated by removing the constant value from the received signal. An average subtraction process of subtracting the calculated average value from the received signal can effectively mitigate reflections in this situation. However, using waveguide antennas, it is generally impractical to provide a uniform distance between the antenna and the skin of the subject's head. Nevertheless, the clutter removal process described below can be applied to signals from antennas with similar distances to the skin, such as 0-10 mm. In particular, the antenna signals are classified based on their distance to the skin (or to the outer layer if the object to be imaged is not a human head), and the average of all received signals in each category Ci is subtracted from the corresponding signal of the antenna at each frequency step:

[0159]

[0160] in, represents the clutter-removed signal, and Nc is the number of classes. In the multi-static configuration of the described embodiment, the clutter removal process is applied solely to the calibrated reflected signal and transmission signal As defined in equation (3) above. The clutter removal process 704 estimates reflections by averaging signals from antennas with similar distances to the subject's head rather than by inverse scattering. In addition, applying average subtraction to each frequency step enables the process to take into account the frequency dependence of electromagnetic wave penetration depth.

[0161] c) Beamforming 706

[0162] By eliminating boundary reflections, the imaged domain can be considered as a homogeneous medium. In step 706, the clutter-free and calibrated scattering parameters are passed through a frequency domain beamformer, such as Figure 8 As shown in Figure 1, the received signal is imaged according to the position. In the beam imaging process, Fig. 9 As shown, the scattered power intensity in the imaged region is calculated by solving Maxwell's equation, and the total power is estimated by summing the calculated power intensities over all frequency samples and antenna positions. As a result of the significant permittivity contrast between healthy and unhealthy tissue, regions of high energy levels in the resulting image correspond to unhealthy tissue.

[0163] exist Fig.10A schematic representation of the imaging domain is shown in , where an electromagnetic wave propagates from a transmitter n to an object with unknown dielectric properties (conductivity ε and permittivity σ). The scattered fields from different point scatterers inside the imaged domain are then measured by an antenna m outside the boundary of the imaging domain.

[0164] In the beam imaging process 706, the electric field Ep at the location of the point scatterer p is calculated by estimating the electric field at the location of the scatterer. For this purpose, the target response at each receiving antenna is correlated with the incident field and synthetically back-propagated to the imaged domain. The electric field at the location of the scatterer is then calculated as follows:

[0165]

[0166] where G is the 2D or 3D Green's function of the point scatterer, r = |rnp| + |rpm| is the distance from the transmitting antenna to the point scatterer and then to the receiving antenna, and γ is the propagation constant of the medium through which the wave travels and is calculated by:

[0167]

[0168] in ω=2πf is the angular frequency, μ is the magnetic permeability, and σ and ε are the conductivity and permittivity of the wave propagation path, respectively.

[0169] The electric field Ep of a point scatterer is different from that in conventional electromagnetic estimation (E m =E scat +E inc ) is usually separated from the incident field by the scattered field E scat However, in the procedure described in this paper, Ep represents the sum of the scattered field and the incident field at the location of the point scatterer:

[0170] E p =E scat (r p )+E inc (r p ) (6)

[0171] as well as

[0172] E m =∫ dv E p dv (7)

[0173] The electric field at the far-field distance r from the scatterer (r>2D2 / λ, where D is the diameter of the point scatterer and λ is the effective wavelength inside the imaged domain) behaves as a spherical wave, and the uniform plane wave equations govern. Thus, the Maxwell equations corresponding to equation (4) are:

[0174]

[0175] The differential equation corresponding to the radial part of the ansatz, G(r), is the Bessel differential equation:

[0176]

[0177] where the complex solution is given by:

[0178]

[0179] which are Hankel functions of order i of the first and second kind, respectively, and J and N are Bessel functions of the first and second kind, respectively. The Hankel functions are singular due to the presence of singularities of the Bessel functions of the second kind (also called Neumann functions). describes a radially symmetric cylindrical wave travelling away from the scatterer (scattering field), while the Hankel function of the second kind represents the wave traveling towards the scatterer (incident field). Considering the scattered field as a reflected wave due to the mismatch between the dielectric properties of the scatterer and the wave propagation path, both the scattered field and the incident field exist at the location of the point scatterer with the same frequency and amplitude, and therefore the singular Neumann functions cancel in superposition. Therefore, the solution is given by:

[0180]

[0181] After clutter removal, the first-order (i=1) Bessel function is the dominant mode and provides the best solution. Therefore, the point scatterer field can be calculated by assigning the Green function from equation (12) into equation (4):

[0182] E p =2S mn J1(γr)e -γr (13)

[0183] The Bessel function J1(γr) can be replaced by its polynomial approximation:

[0184]

[0185] When the imaging domain is filled with a matching medium, the point scatterer field It can also be calculated by the following formula:

[0186]

[0187] where γb is the background propagation constant, and is the measured scattering parameter of the background.

[0188] In the calculation of the field of a point scatterer, the scattered field is considered as a reflected wave superimposed on the forward wave in the background medium. Therefore, the reflection coefficient associated with the scatterer can be defined as:

[0189]

[0190] where * denotes the conjugate operation. Therefore, the scattered field at the location of the point scatterer due to the wave transmitted by the nth antenna and received by the mth antenna is given by:

[0191]

[0192] In a multi-static imaging configuration (where the Na antenna acts as both a transmitter and a receiver), the total power density ("density" because it is calculated per pixel area or voxel volume) produced by all antennas at the location of each scatterer is calculated to estimate the location of any anomalies. For this purpose, the total power density is calculated by summing the estimated electric fields inside the imaged object from the different transmitters and receivers around the object, as follows:

[0193]

[0194] The incident field En is usually chosen to be 1 (V / m), so it is neglected in the calculations in equation (18). To compensate for the lack of discrete observation points, different scattering profiles for different frequencies are superimposed to generate an electromagnetic power image, as follows:

[0195]

[0196] Where Na and Nf are the number of antennas and frequency samples respectively.

[0197] The average electric field at a point scatterer can also be calculated as:

[0198]

[0199] In this case, the mean field The real part, imaginary part and / or phase of can be used as the final image.

[0200] For completeness, the process can also be used in a monostatic measurement configuration by changing the triple summation operator to a double summation in equations (19) and (20), as follows:

[0201]

[0202] The described process can also be used to generate a 3D image by using a 3D antenna array configuration around the imaged object. In this case, the point scatterer position rp represents a 3D vector in the volume of the imaged object. The 3D antenna array can be created by adding more rings of antennas above and / or below the 2D imaging plane or by moving the antennas in the z direction to scan the entire object. In either arrangement, interpolation techniques are used to generate a 3D image or a multi-layer image of the object.

[0203] Although the above process does not provide a detailed image of the dielectric properties, it does show areas of contrast in the dielectric properties, which is sufficient for rapid detection of important scatterers and their approximate locations. If desired, the process can be extended to generate detailed images as described below.

[0204] An important aspect of the beam imaging process 706 involves the calculation of the propagation constants used in equations (19) and (20). The average dielectric properties of the imaging domain can be used in the propagation model. However, in order to create more accurate images of heterogeneous objects such as the human body, a more accurate estimated spatial distribution of the dielectric properties in the imaging domain is needed.

[0205] Due to multiple reflections and refractions of electromagnetic waves in a dispersed and heterogeneous environment, the waves pass through different tissues and travel along different paths depending on the position of the transmitter-receiver relative to the imaged object. In a homogeneous environment, the path can be calculated as a direct (straight) path (shortest path) from each transmitter to a point scatterer and from a point scatterer to each receiver. However, in a heterogeneous medium, the electromagnetic wave travels along the path with the minimum propagation time or minimum delay. Therefore, in step 906, the wave propagation path can be found by finding the path with the minimum propagation time (t=γr / ω) inside the imaged object, providing the required values ​​of the propagation constant and distance in equations (19) and (20). For this purpose, a template including the anatomical and dielectric properties of the imaged object is considered to be beneficial. The template can be an actual healthy model of the imaged object extracted from images of other imaging modes (such as MRI and / or CT equipment), or an approximate model using the imaged 3D head profile 412.

[0206] The shortest path finding method is then used to find the wave propagation path inside the template. In the described embodiment of the process, the Dijkstra algorithm is used. By defining the propagation time (tuv) of the wave propagating from voxel v to its neighboring voxel u as the cost function, the shortest path finding algorithm finds the path with the minimum cost (propagation time) through the model. As a result, the position-specific propagation constant of the imaged domain seen by each transmit-receive antenna pair γmn(rp) is estimated. The propagation distance parameter γr used in equations (19) and (20) is then calculated by summing the γrs of all voxels along the propagation path from transmitter n to receiver m, as follows:

[0207]

[0208] where γuv and ruv are the propagation constant and the distance from voxel v to its neighboring voxel u, respectively. Therefore, the main equation for the imaging process is:

[0209]

[0210] Another aspect involves updating a model of the brain tissue of a (e.g., healthy) subject (initially generated from or as an initial template) to estimate the dielectric properties of abnormalities within the imaged object. The image created by equation (22) or (23) shows the location of important scatterers (abnormalities). Therefore, the healthy model used in the initial step can be simply updated by changing the characteristic values ​​of the voxels corresponding to the abnormal location. In this case, any suitable technique such as finding local maxima, thresholding, differentiation (subtracting subsequent images), clustering, and the like can be used to determine the location of the abnormality. In some embodiments, a thresholding process is utilized in which a histogram of the image generated by equation (22) or (23) is generated and the standard deviation of the histogram is considered to be a threshold T. The location of the abnormality is then determined based on the following conditions:

[0211] If I(voxel)>T, then voxel∈“unhealthy tissue”, otherwise voxel∈“healthy tissue”

[0212] After finding the location of the unhealthy tissue, the dielectric model is updated by replacing the permittivity and conductivity values ​​of the voxels corresponding to the location of the abnormality in the model by the property values ​​of the target. Optionally, an optimization technique can be used to incrementally change the values ​​for healthy tissue according to the following equation:

[0213] ε new =ε healthy +Δε,σ new =σ healthy +Δσ (28)

[0214] The shortest path finding process is then performed on the updated template to calculate new location-specific propagation constants. The new propagation constants are then substituted into equation (22) or (23) to estimate the new scattering profile of the object. This iterative process is repeated until the updated values ​​have not changed significantly relative to their values ​​in the previous iteration, e.g. Fig. 9 As shown, the indication error has reached a small value.

[0215] An optional stopping criterion for the iterative loop is to compare the calculated and measured electromagnetic energy in the antennas. When antenna n transmits an electric field to the imaged object, the electric field is scattered in the imaged domain and received by all antennas. Therefore, the total electric field received by the antennas around the object is the sum of the scattered electric fields from different scatterers in the domain. Therefore, the total energy measured by all antennas can be compared with the calculated scattered energy to find the energy mismatch ΔS between the utilized template and the actual object being measured, as follows:

[0216]

[0217] where Np is the number of scatter points in the imaged domain. An iterative process can then be utilized to minimize the energy mismatch ΔS through optimization.

[0218] As a summary, the beamforming process can be represented by the following steps, as shown in Fig. 9 As shown in the flow chart:

[0219] Step 906: The input template 416 of the healthy subject (in the context of the described embodiment, the subject's head) defines the initial values ​​of the updateable model 902 of the subject; this model 902 and the position 904 of the antenna relative to the subject's head are processed to calculate the position-specific propagation constants of the subject by a shortest path finding method.

[0220] Step 908: The calibrated and clutter-removed scatter parameters of the object are back-propagated into the imaging domain to estimate the electric field at each pixel (or voxel if 3D) within the imaging domain

[0221] Step 910: The electric field power density at each pixel (or voxel) (summed over all antennas) is calculated and summed over each transmit frequency to generate a power density map or image of the object (i.e., representing the spatial distribution of scattered electric field power within the object)

[0222] Step 912: Determine the pixel / voxel locations of any anomalies (significant scatterers) (e.g., by using a thresholding technique)

[0223] Step 914: Duplicate model 902 and update the copy of the model by replacing the values ​​of the pixels / voxels corresponding to the anomaly locations

[0224] Step 916: Perform a test to determine if there are any significant differences between the two models; if the test determines that there are significant differences between the models, the process branches to step 918 below.

[0225] Step 918: The process loops back to recalculate the propagation constant and distance at step 906 using the updated model

[0226] Otherwise, if the test at step 916 determines that there are no significant differences between the current model 902 and the most recently updated model, the latter is output as a beamforming image 920 of the object.

[0227] It will be apparent that the test at step 916 defines the termination or exit criteria for the iterative loop, and thus the beam imaging process 706 is an iterative process.

[0228] The power density map generated at step 910 shows the location of abnormalities in the brain. For example, a hemorrhagic stroke appears clear with high intensity, while a clot appears as an area of ​​minimal intensity compared to its adjacent tissue. The high intensity of a suspicious area is an indicator that it is a hemorrhagic area, while a missing abnormality or a low intensity when symptoms of a stroke persist indicates a blood clot.

[0229] The final beamforming image 920 of the subject's head may represent only the permittivity and / or conductivity of the imaged region, or it may be updated to include data defining the location of one or more tissue properties of interest (e.g., water content) and / or one or more associated labels (e.g., whether the tissue is abnormal, etc.).

[0230] d) Optimization

[0231] If desired, the beam imaging process 706 can be accelerated by reducing the number of frequency samples, assuming that this can be done without compromising (or unduly compromising) image quality. To this end, the Nyquist theorem can be applied in the frequency domain to find the maximum frequency step size (and thus the minimum number of frequency samples) for reconstruction sampling. The Nyquist theorem is applied in the frequency domain given the time-limited nature of the received signal.

[0232] According to the Nyquist theorem, in order to theoretically be able to recover the entire data, the sampling step size δf should be less than 1 / 2τ, where τ is the temporal width of the time-limited signal. If the Nyquist criterion is not met (undersampling), part of the data will be missed, causing overlap of the reconstructed signal and resulting in images with incorrect and / or multiple targets. Oversampling (sampling at a rate higher than the Nyquist rate) can improve resolution and avoid overlap, but requires additional time for measurement and for processing. By considering that the temporal width of the received signal is equal to the data acquisition time and based on the signal bandwidth B, the minimum number of samples can be calculated as:

[0233]

[0234] Using a minimal amount of sample during imaging significantly reduces processing time while maintaining accurate imaging.

[0235] The beam imaging process 706 described herein is significantly different from standard microwave tomography, which typically requires solving an ill-posed inverse problem to estimate the dielectric properties of tissue. Compared to time domain methods (e.g., confocal, space-time beamforming MIST, and adaptive beamforming imaging methods), the beam imaging process 706 described herein is less computationally intensive and performs all of its calculations in the frequency domain, making it less susceptible to noise and multiple reflections generated by multi-layer structures.

[0236] The final beam imaging image 920 of the subject's head shows any abnormalities in the brain. For example, a hemorrhagic stroke appears clear with high intensity, while a clot appears as an area of ​​minimal intensity compared to its adjacent tissue. The high intensity of a suspicious area is an indicator that it is an area of ​​hemorrhage, while a missing abnormality or a low intensity when symptoms of a stroke persist indicates a blood clot. The image may represent only the permittivity and / or conductivity of the imaged area, and it may be updated to include data defining the location of one or more tissue properties of interest (e.g., water content) and / or one or more associated labels (e.g., whether the tissue is abnormal, etc.).

[0237] Category 312

[0238] Determining whether a stroke is caused by bleeding or clots and finding its location is important for the diagnosis and treatment of acute stroke, and is also a key issue in post-stroke management. In general, identifying stroke subtypes from electromagnetic head scanning systems is based on image reconstruction algorithms. However, image-based methods are time-consuming and have low accuracy. A novel complex network method applied to an electromagnetic scanning system to identify intracranial hemorrhage (ICH) from ischemic stroke (IS) and find the location of the stroke is described below. The classification is based on using graph degree mutual information (graph degree mutual information, GDMI) to evaluate the differences between ICH and IS groups, where each subject is composed of multi-channel antenna receiving signals. Each signal is converted into a graph to avoid different signal amplitudes. Then, the relationship between each pair of graph degree features is calculated by mutual information and input into a support vector machine to identify IS from ICH. 95% accuracy is achieved when identifying ICH for IS. Regarding positioning, a weighted graph is applied to extract the intensity features from the transmitter to the receiver between antenna pairs. The maximum weighted pair will pass through the target. The execution time for pattern feature extraction and classification or localization is less than one minute, which is suitable for a stroke emergency treatment system.

[0239] 1. Overview of the Method

[0240] Fig.11 is a flow chart of the data processing and analysis steps prior to classification and localization. In this case, the system workflow can be summarized as follows:

[0241] 1. As Fig.17 As shown, a multi-channel input signal is received from an antenna array, where the default number N of channels is 16.

[0242] 2. The scattered signal received from each antenna is a frequency domain signal (amplitude and phase) or a time series, where the number of sample points is greater than 700.

[0243] 3. If the input is not already a time series, the inverse FFT is used to convert these signals into a multi-channel time series.

[0244] 4. Next, the fast weighted horizontal visibility algorithm (FWHVA) described in G. Zhu, Y. Li, and PP Wen, “Epileptic seizure detection in EEGs signals using a fast weighted horizontal visibility algorithm” (Computer methods and programs in biomedicine, Vol. 115, pp. 64-75, 2014) is used to map the multi-channel time series to NxN=16×16=256 graphs.

[0245] 5. Extract degree sequence and weighted intensity sequence.

[0246] 6. Finally, output features: degree and intensity are used for classification and positioning respectively.

[0247] Fig.12 yes Fig.11 Flowchart of the classification process following the process of .

[0248] Fig.13 yes Fig.11 Flowchart of the positioning process after the process.

[0249] 2. Detailed classification method

[0250] To supplement the output image 920 indicating whether the patient has had a stroke, a classification process is performed to assess the type of stroke using S parameters. In addition, templates and fused brain images as described below can be used to improve accuracy. In general, features can be calculated by correlation, coherence or synchronization algorithms and forwarded to a general classifier (e.g., SVM or random forest) or forwarded by a deep learning neural network to identify the subject as ICH, IS or healthy.

[0251] In the described embodiment, a complex network approach is used to assess the stroke type using the measured S parameters at step 312. The classification process 312 is not only able to distinguish between ICH, IS, and healthy people, but also to locate the stroke location.

[0252] Fig.12 is a flow chart of the classification process 312. The scattered field data (S parameters) are collected from N antennas in the frequency domain, resulting in N×N signals. Then, the inverse FFT transforms the N 2 frequency domain signal into N 2 time series signals. Each time series signal (X = x1, x2..., x T) is then mapped to a graph G(V;E) consisting of nodes / vertices (V) and edges (E). In particular, the fast weighted horizontal visibility algorithm (FWHVA) is used to transform the data into a horizontal visibility graph (HVG), where at each time point x i are vertices of the graph, and any two points (x i ;x j ) is represented by edge e i，j ,in:

[0253]

[0254] where e i，j =1 indicates that the edge exists, and zero indicates that no edge exists. Fig.16 (b) shows a graph generated from the scattering parameters from the second antenna of a human head with IS based on equation (1).

[0255] After each graph is constructed, node degrees and degree sequence features are used to characterize the graph. i The degree d(v i ) is from v i The number of connected edges, and the degree sequence (DS) is the sequence of graph degrees connected in the order of node numbers [d(v1), d(v2), ... d(v T )]. For example, in Fig.16 In the graph of (b), degree d(1) = 1 and d(2) = 2, and DS = (1; 2; 3; 5, ...; 2; 1).

[0256] Once the graphs are constructed, the graph mutual information is calculated between each pair of graphs. Mutual information (MI) is often used to measure the mutual dependence of a random variable on another variable. Given two discrete variables X and Y, the mutual information of X and Y is given by:

[0257]

[0258] Where p(x) is the probability density of X, and p(x,y) is the joint probability distribution function of X and Y. In general, if X and Y are dependent, the MI is higher. In the described embodiment, instead of calculating the mutual information between complex graphs, the system uses a simpler and faster method referred to herein as graph degree mutual information (GDMI), which compares the degree sequences of two graphs. Given two degree sequences, DSx, DSy of two HVGs, Gx, Gy, the graph degree mutual information (GDMI) of the two graphs is measured as follows:

[0259] GDMI(Gx,Gy)=MI(DSx,DSy) (3)

[0260] If Gx and Gy are similar, the corresponding GDMI(Gx; Gy) is high; otherwise, if Gx and Gy are different, the value is low.

[0261] After inputting each combination of the graph sequence into equation (3), we obtain N 2 ×N 2 If the degree sequences of two graphs Gx and Gy are the same, then GDMI(Gx, Gy) = 0, so that the diagonal of the GDMI matrix is ​​always zero.

[0262] In order to classify the stroke type (hemorrhage / clot), a support vector machine (SVM) classifier is used. SVMs are successfully used to classify HVG features associated with alcoholics and sleep EEG signals. They can perform both linear space discrimination and nonlinear classification by selecting different kernel functions, which can be linear, polynomial kernels, radial basis functions (RBF) or sigmoidal. The described embodiment uses an SVM algorithm with an RBF kernel. The training set is first used to learn the decision hyperplane between the features of the two stroke types using a multi-class SVM, and optionally learns the health condition, which is then applied to classify the stroke type according to the new measurements.

[0263] The classification process 312 is fast and reliable because it normalizes data for different head sizes. Currently, using a general purpose computer, the calculation speed is less than one minute (wall clock time). As described below, the classification process 312 can also find strokes based on the generated time series and antenna intersections.

[0264] 3. Detailed positioning method

[0265] To perform localization, the inventors noted that a brain containing a stroke has an imbalance in the transmitted / received signals. Assuming that the head is approximately symmetrical across the major and minor axes (i.e., the sagittal plane across the left and right separating the brain and the central coronal plane across the front and back separating the brain), symmetrical antenna pairs across these symmetry lines should have similar signals in a healthy head. If symmetrical antenna pairs measure different signals, this indicates an abnormality, such as a stroke, and this observation can be used to facilitate localization.

[0266] Fig.17 A cross section of a brain surrounded by a 16-element electromagnetic imaging antenna array is shown with antenna signal pairs and a stroke 1702. Since only one signal passes through the stroke, the signal from antenna pair (1-7) will be significantly different from its symmetric pair (1-11). However, the 1-13 and 1-5 antenna pairs should have similar received signals. The maximum weight pair will traverse the stroke target. Given a 16-element antenna array, the two approximate symmetry axes of the brain correspond to signals originating from antennas 1, 5, 9, and 13.

[0267] To perform localization, a weighted horizontal visibility graph (w-HVG) is computed. The process of converting the time series to an unweighted graph but with initial weight terms is as described above, and the resulting graph is denoted as G(V, E, W). At node v i and v j (Equivalently, x i and x j ) between the edges e i，j The weight w i，j is defined as follows:

[0268]

[0269] Weighted graphs are characterized by the concept of strength, where a node v i The strength is defined as:

[0270]

[0271] For a weighted graph with T nodes, its average strength is defined as:

[0272]

[0273] The differential weighted intensity is defined as:

[0274]

[0275] in, is the average strength between antennas 'i' and 'j', and Opp() indicates symmetrically opposed antenna pairs originating from the same antenna i. For example, the opposed antenna pairs include antennas 2-16, 3-15, 4-14, ... when transmitting from antenna 1 or 9, and the opposed antennas include antenna pairs 1-9, 2-8, 3-7, ... when transmitting from antenna 5 or 13. Note that in real experiments, the antenna array may not be well calibrated. Therefore, the bias in equation (7) can be removed as shown in equation (8):

[0276]

[0277] The antenna pairs that pass through the stroke target from the ith antenna can be detected using the following equation (9), where "%" refers to the modulo operator. Note that neighboring antennas are not considered because the information obtained from them is less than that obtained from more distantly spaced antennas.

[0278]

[0279] When the input signal contains the signal from the brain template, the true location of the stroke can be obtained by computational geometry methods. First, the intersection point can be calculated based on equation (10) using the four points (x1, y1), (x2, y2), (x3, y3) and (x4, y4) of the antenna position.

[0280]

[0281]

[0282] However, equation (9) is not sufficient to identify which hemisphere has the target, so a natural way to detect whether the hemisphere is in the left or right hemisphere is by comparing features (such as amplitude, average or intensity) from the left hemisphere signal with features from the right hemisphere signal. However, since the head is not always completely symmetrical, this method may rarely provide a perfect answer. To address this difficulty, the process described in this article uses a pair of opposing antennas 5 (left hemisphere) and 13 (right hemisphere) to detect whether the intersection point should be placed on the left or right. The relevant process steps can be described as follows:

[0283] (a) The left intersection point L0(lx,ly) is calculated based on equations (10) and (11).

[0284] (b) The right intersection point R0(rx,ry) is calculated based on equations (10) and (11).

[0285] (c) Draw a line segment from antenna 5 to antenna R1 on the right side, which is not the 13th antenna. The line segment has a minimum distance to L0 which is separated from the 13th antenna.

[0286] (d) Pick up the neighboring antenna R2 of R1, and the line segment from R2 to antenna 5 is far away from L0 and R0.

[0287] (e) Calculate the differential weighted node w5r from the fifth antenna to R1 and R2 respectively, where w5 r =|S5r1-S5r2|.

[0288] (f) Similarly, for antenna 13, run steps c to e to obtain the differential weight node w13 from the 13th antenna to the two adjacent antennas L1 and L2 respectively. l .

[0289] (g) If w5 r >w13 l , then the intersection point is in the left hemisphere, otherwise, it is in the right hemisphere.

[0290] Fig.21is a schematic cross-sectional plan view of a subject's head surrounded by antennas, showing process steps (a) to (g). The differential weighting node between (5, R1, R2) is significantly different from the differential weighting node between (13, L1, L2), providing two possible candidates (L0 or R0) for the bleeding location. Following the above steps, it can be determined that the bleeding is located in the left hemisphere (at L0 instead of R0).

[0291] Fig.19 is a flow chart of the localization process. The output of the localization process is a set containing antenna pairs and their corresponding assigned weights. Each weight corresponds to the probability of a signal passing through an abnormality (e.g., a stroke), where a larger value indicates that the signal between the antenna pairs passes through the stroke with a high probability. Therefore, the localization process can determine those antenna pairs whose signals pass through the abnormality, and this can be used to enable the clinician to determine the true location within the brain template.

[0292] By combining the signal with the brain template, a geometric approach can be used to calculate the true location of the stroke. Given two independent antenna pairs that contain an anomaly according to the high differential weighted node value DW, the intersection between the antenna pairs is determined to find the stroke / anomaly. That is, the stroke location can be found as the intersection of the line between the antenna pair at position (x1, y1) and (x2, y2) and the line between the antenna pair at position (x3, y3) and (x4, y4). Fig.19 As shown, for an example of 3 antenna pairs, this crossover approach can be extended to additional antenna pairs. The output of the localization is the true position in the brain based on the brain template and the highly weighted set of antenna pairs.

[0293] Local area tomography 314

[0294] As a final verification, using the fused image 310 as the initial image, differential equation tomography can be used for local area tomography at step 314. In particular, the differential equation tomography method described in International Patent Application No. PCT / AU2018 / 050425 entitled "Atomographic imaging system and process" is used to determine the precise dielectric properties of the suspicious area identified by the beam imaging process 308 at different frequencies, and these determined properties are then used to determine the type of abnormality. For example, higher dielectric properties (mainly permittivity) indicate bleeding, while lower values ​​indicate clots. In the described method, the fused image is updated to include more accurate dielectric properties of the suspicious area determined by the differential equation tomography method.

[0295] As described in the patent application, standard integration-based microwave tomography suffers from computational limitations, primarily due to the need for Green's functions. These limitations prevent the rapidity of the image reconstruction process, which is an important requirement for urgent medical needs in stroke detection. In order to cope with these limitations, obvious approximations are usually utilized. These approximations include considering the imaging antenna as a point source (which makes the problem underdetermined), assuming uniformity along one of the coordinate axes of the domain, and requiring a background matching medium. These approximations result in the reconstruction of only two-dimensional images with a certain level of error. In order to alleviate the above-mentioned limitations and the resulting challenges, differential equation-based microwave tomography built on waves and the third Maxwell equations is utilized in this project. The proposed method is independent of Green's functions and their corresponding limitations. The method can quickly reconstruct true three-dimensional images without any need for the above-mentioned approximations by solving linear and determined problems.

[0296] If the classification thus obtained is different from the classification given by the classifier 312, the "proof by contradiction" method known to those skilled in the art is applied to obtain a reliable decision. According to this method, the tomography process is performed again, but this time starting with the classifier output. The convergence rate of the cost function in the second execution of the local area tomography is compared with the initial convergence rate that led to the first different answer. The convergence rate that reaches a lower error level is most likely to be the correct diagnosis. However, in the case where the two methods converge to the same error level, another optimization method is used, such as Newton's method, conjugate gradient method or contrast source method.

[0297] In any case, the decision is validated using a modified integrated tomography process 316, as described below.

[0298] Global Tomography 316

[0299] Finally at step 316, a global tomography process based on the modified integral-based tomography is applied to the initial brain template to verify the dielectric properties throughout the brain, including the suspected region.

[0300] Conventional integration-based tomography methods can be classified into three types, depending on whether they are based on the Born approximation, Newton iteration method (distorted Born approximation), or contrast source inversion (CSI). CSI-based methods are attracting more interest in EM tomography due to the desirable properties of high accuracy and low computational cost. The traditional CSI method includes two optimization steps: the first step optimizes the contrast source, while the second step optimizes the contrast function based on the best contrast source from step 1. The best contrast function provides the permittivity and conductivity information of the imaged region, which is the desired final solution.

[0301] According to many studies of CSI-based methods, the inventors found that the optimization of the contrast source in step 1 is much easier than the optimization of the contrast function in step 2. Therefore, in the process described in this article, only the conjugate gradient method is used to optimize the contrast source. After the optimal contrast source is obtained, the brain template is used to calculate the total electric field using the FDTD / MOM forward solver. Because the contrast source is the product of the contrast function and the total electric field, the calculated electric field can be used to extract the contrast function from the optimized contrast source.

[0302] The formulation and procedures of the method can be summarized as follows:

[0303] 1. Reconstruct the contrast source by using the conjugate gradient method.

[0304]

[0305] where w is the contrast source, G is the Green's function, and S is the received signal (S parameters).

[0306] 2. Use the brain template and the FDTD / MoM solver to calculate the total electric field:

[0307] E tot =f FDTD / MoM (x temp ) (28)

[0308] Where E tot is the electric field, f FDTD / MoM (·) is the FDTD or MoM solver, and χ temp It's a brain template.

[0309] 3. By using w and E tot To extract the contrast function of the brain:

[0310]

[0311] where χ is the contrast function of the brain containing both permittivity and conductivity information.

[0312] The described process offers three main advantages:

[0313] 1. The optimization of the contrast source converges faster and with greater accuracy than the optimization of the contrast function;

[0314] 2. The total electric field calculated using the brain template can generally be used to represent the electric field when abnormal tissue (such as hemorrhage or clot) is present. Therefore, optimization of the electric field is not necessary in the case of stroke.

[0315] 3. The contrast function extracted by using the electric field of the contrast source and the brain template is simple and stable.

[0316] The above methods use a single frequency to reconstruct the electrical properties of the brain. However, due to the dispersed characteristics of biological tissues, multi-frequency methods provide more information about the electrical properties of biological tissues. Therefore, the inventors have also developed a global tomography method based on multiple frequencies. In particular, four methods based on multiple frequencies have been developed, as described below.

[0317] 1. Differential Sum and Average Method (DSMM)

[0318] The first method is to sum the reconstructed conductivity at different frequencies and calculate the average value. Assume that at frequency f n The reconstructed differential contrast of each pixel is expressed as in is calculated as:

[0319] χ diff = x-x temp (30)

[0320] where χ temp is the contrast of the brain template, and then the average of the reconstructed differential contrast at different frequencies is calculated as:

[0321]

[0322] Where N is the number of frequency samples.

[0323] 2. Correlation coefficient method (CCM)

[0324] The described correlation coefficient method can be used to estimate the similarity of two sets of signals. Values ​​of the correlation coefficient close to 1 indicate high similarity to the original signal, while values ​​close to 0 indicate dissimilarity. The method calculates a correlation coefficient map between the reconstructed conductivity distributions of the patient and the template. Regions containing stroke will show high dissimilarity (low correlation coefficient), while regions without stroke will show high similarity (high correlation coefficient). The correlation coefficient of the reconstructed contrast distribution is calculated as:

[0325]

[0326] where s and s temp is the standard deviation of the reconstructed contrast of the patient and template at different frequencies, and μ and μ temp is the average of the reconstructed contrast at different frequencies.

[0327] 3. Differential Entropy Based Method (DEM)

[0328] Entropy is a measure of the uncertainty of a random variable. When the outcomes of an experiment are equally probable (in other words, the probability distribution is uniform), the maximum value of entropy is obtained. Conversely, the minimum entropy value suggests that an event has occurred. Therefore, the entropy value can also be used to measure the similarity between two sets of signals. Low entropy values ​​between the reconstructed contrast of the patient and the corresponding template head suggest stroke areas, while high entropy values ​​suggest normal areas. In the method used by the system described in this article, Rényi entropy is used with an order of 50. The Rényi entropy of the differential contrast is calculated as:

[0329]

[0330] in, is the normalized reconstructed contrast, which is calculated as:

[0331]

[0332] 4. Hybrid Approach

[0333] Based on the results of using the above-mentioned DSMM, CCM and DEM, the inventors found that different methods have their own advantages and disadvantages. In particular, DSMM is able to find the location of the stroke with an expected shape reconstruction, however, when compared with the results from CCM and DEM, the artifact clusters are serious. CCM is able to remove artifact clusters; however, the location of the stroke may move, and therefore accurate positioning cannot be achieved. Finally, DEM is able to accurately find the stroke location while suppressing artifact clusters; however, the shape of the stroke area suffers from distortion. In view of the above, the inventors have developed a hybrid method that combines these three methods so that the advantages of each method can be provided. The simplest of these hybrid methods sums the results obtained from each method to provide a hybrid image according to the following formula:

[0334]

[0335] Image Fusion(310)

[0336] Back to Figure 3 In order to provide a single output image from the system to aid in the diagnosis of stroke or trauma, at step 310, an image fusion process 310 generates a composite image of the subject's head. The fusion process combines complementary features of the head template, beamforming image, and multi-frequency radar / tomography-based image, such as Fig. 20 The fusion enables the head boundaries and high-resolution details from the head template to be combined with the object detection from beam imaging and the dielectric property estimation / refinement from the tomography-based approach.

[0337] Several techniques can be used for image fusion. One approach involves combining images based on wavelet or multiscale features. In fusion based on multiscale transforms, high-frequency components of one image (belonging to useful image details, such as "edges") are extracted by multiscale transforms and then combined with another image (e.g., by substitution, addition, or selection methods) to generate a fused image. Other methods are based on training machine learning models to process features within the image and decide which features are useful features to be retained. This can be achieved by models such as artificial neural networks (ANNs). There are many other methods for image fusion. Each fusion method has its own advantages that are beneficial to the fusion process. For example, methods based on multiscale transforms are suitable for multispectral images and produce better frequency resolution outputs with more detailed features. ANNs are attractive for image fusion due to their ability to fuse images without undergoing complex computational workloads. However, the use of a single technique may be undesirable due to resolution limitations or requirements for high-quality training data, etc. Therefore, although the use of a single method produces promising results, methods that combine the advantages of various techniques are desirable for more effective fusion.

[0338] In the system described herein, a hybrid process for image fusion is preferably used, which involves several different fusion methods to process images, extract important features and remove noise from different head imaging modes (i.e., beam imaging, tomography). The results from different individual fusion techniques are then combined to construct a fused image. The integration process combines the results from various methods and ensures that the final fused image inherits the advantages of each fusion method. It quantitatively and qualitatively improves the quality of the fused image compared to the original image. In addition, in order to further improve the accuracy of the fused image, a patient-specific head template is also used in this fusion strategy. When the input from beam imaging and tomography displays information about the internal tissue irradiated by the electromagnetic signal, the head template provides information about the external head layer based on the actual 3D head surface of the patient. More importantly, the relative distribution of brain tissue is also contained in the template. Therefore, it can be used in combination with the features extracted from beam imaging and tomography to enhance the visual representation of the internal brain tissue in the final fused image.

[0339] Many modifications will be apparent to those skilled in the art without departing from the scope of the invention.

Claims

1. A method for medical imaging, the method comprising: (i) receiving scattering data representing monostatic or multistatic measurements of electromagnetic signal scattering from tissue of a body part of a subject at a plurality of different signal frequencies, wherein electromagnetic signals are transmitted from one or more antennas and corresponding scattered signals are measured by said one or more antennas, initial values ​​of a model of said tissue of said body part of said subject being defined by an input template of a healthy subject, said model and the positions of said one or more antennas relative to said body part of said subject being processed to calculate a propagation constant of said subject and a distance from a transmitting antenna to a scattering location and then to a receiving antenna by a shortest path finding method; (ii) processing the scattering data to calculate an electric field power value at each of a plurality of scattering locations of the tissue within the body portion of the subject and for each of the plurality of different signal frequencies, wherein the electric field power value is calculated based on the propagation constant and the distance; (iii) for each of the plurality of scattering locations, summing the calculated electric field power values ​​for the plurality of different signal frequencies and the one or more antennas at that scattering location to generate an image of the tissue within the body portion; and (iv) iteratively updating the model based on a comparison of the model of the tissue within the body part with the generated image until a termination criterion is met, wherein the updated model is output as an image of the tissue within the body part of the subject.

2. The method of claim 1, wherein the measurements are multi-static measurements, wherein an electromagnetic signal is selectively emitted from each of a plurality of antennas disposed around the body part and a corresponding scattered signal is measured by each of the plurality of antennas.

3. The method of claim 1 or 2, wherein the body part is the head and the tissue comprises brain tissue of the subject.

4. The method according to claim 1 or 2, comprising: (v) using machine learning to process the biographical data of the subject to select a base template from a template library as the best match for the subject, wherein the templates in the template library represent corresponding models of tissues of the body parts of the corresponding subject, and the biographical data of the subject represents at least the age, gender and weight of the subject; (vi) processing the selected base template and the measurements of the external dimensions and / or shape of the subject's body part by geometrically transforming the spatial coordinates of the selected base template to match the measurements of the subject's body part to generate template data representing a model of the tissue of the subject's body part.

5. The method according to claim 1 or 2, wherein the step of processing the scattering data comprises the following steps: (vii) normalizing the scatter data and removing clutter from the scatter data; as well as (viii) Processing the normalized and clutter-removed scattering data to calculate the electric field power value.

6. The method of claim 5, wherein removing clutter from the scattered data comprises determining an average value of the measured electromagnetic signal and subtracting the average value from each signal measurement at each frequency to remove strong reflections and clutter from the scattered data.

7. A method according to any one of claims 1, 2 and 6, comprising calibrating the scattering data by dividing the measured scattering parameters of the body part by the measured scattering parameters of the imaging domain in the absence of the body part and when the imaging domain is filled with a material having dielectric properties matching a medium or a generic body part model.

8. The method according to any one of claims 1, 2 and 6 includes classifying abnormal tissue within the body part as hemorrhagic or ischemic by: converting frequency domain signals into time domain signals and mapping the time domain signals to a corresponding graph, determining the node degree and degree sequence characteristics of the graph, calculating graph degree mutual information to evaluate the similarity of the graphs, and training a classifier with a training set of graph degree mutual information features and their corresponding class labels, and applying the classifier to the graph calculated for the tissue within the body part of the subject.

9. The method of any one of claims 1, 2 and 6, comprising comparing the signals of corresponding pairs of opposing antennas to identify significant differences between the signals in different hemispheres of the subject's brain, the differences indicating an abnormality in one of the hemispheres.

10. A computer-readable storage medium having stored thereon processor-executable instructions which, when executed by at least one processor of a medical imaging system, cause the at least one processor to perform the method according to any one of claims 1 to 9.

11. An apparatus for medical imaging, comprising components configured to perform the method according to any one of claims 1 to 9.

12. An apparatus for medical imaging, comprising: (i) an input receiving scatter data representing monostatic or multistatic measurements of electromagnetic signal scatter from tissue of a body portion of a subject at a plurality of different signal frequencies, wherein the electromagnetic signal is transmitted from one or more antennas and corresponding scattered signals are measured by the one or more antennas; (ii) an image generating component, the image generating component being configured to: - defining the body part of the subject by an input template of a healthy subject The model and the position of the one or more antennas relative to the body part of the subject are processed to obtain an initial value of the model of the tissue within the body part of the subject. A shortest path finding method calculates a propagation constant and a distance of the subject, wherein the propagation constant is the propagation constant of the medium through which the electromagnetic wave passes, and the distance is the distance from the transmitting antenna to the scattering location and then to the receiving antenna; - processing the scattering data to calculate at each of a plurality of scattering locations of the tissue within the body part of the subject and for the plurality of scattering locations an electric field power value for each of the different signal frequencies, wherein the electric field power value is calculated based on the propagation constant and the distance; - for each of the plurality of scattering locations, summing the calculated electric field power values ​​for the plurality of different signal frequencies and the one or more antennas at that scattering location to generate an image of the tissue within the body part; and - iteratively updating the model based on a comparison of the model of the tissue within the body part with the generated image until a termination criterion is met, wherein The updated model serves as the group of the subject within the body part The woven image is output.

13. The apparatus of claim 12, wherein the measurements are multi-static measurements in which an electromagnetic signal is selectively emitted from each of a plurality of antennas disposed about the body part and a corresponding scattered signal is measured by each of the plurality of antennas.

14. The apparatus of claim 12 or 13, wherein the body part is the head and the tissue comprises brain tissue of the subject.

15. The apparatus according to claim 12 or 13, comprising a template generator to: Processing the subject's biographical data using machine learning to select a base template from a template library as a best match for the subject, wherein the templates in the template library represent corresponding models of tissues of the corresponding subject's body parts, and the subject's biographical data represents at least the subject's age, gender, and weight; and The selected base template and measurements of external dimensions and / or shape of the subject's body part are processed by geometrically transforming the spatial coordinates of the selected base template to match the measurements of the subject's body part to generate template data representing a model of tissue of the subject's body part.

Images