The present disclosure relates generally to methods for detecting dielectric changes in matter.
Electromagnetic (EM) waves get perturbed by the distribution and dielectric properties of tissues in the body. Microwave Tomography (MWT) technology has been developed to provide a quantitative image of the dielectric profile of an object of interest (OI). MWT techniques typically rely on two main steps: 1) a data collection step and calibration and 2) a data processing step. In the data collection and calibration step the OI is “illuminated” by multiple antennas radiating at the microwave frequencies. Data from experiments with and without the OI yielding the total field and incidence fields, respectively. The incidence field is often subtracted from the total field to yield the scattering field for which the dielectric profile of the OI is inferred in step 2. Currently, inversion algorithms utilized in MWT are nonlinear and attempt to iteratively solve the EM inverse scattering problem, and it is not possible to know, a-priori the sufficient amount of scattering data needed for proper reconstruction. To alleviate this uncertainty, as much scattering data is collected as possible. However, limitations of space in the near field system and mutual coupling between the co-resident antennas located in the MWT chamber and long acquisition time prevent increasing the number of antenna elements and acquiring of more data. The number of antenna elements is usually determined by a compromise between mechanical scanning (moving of the antenna in space) and having a large number of antennas. With this compromise in mind, it is beneficial therefore to increase the amount of information by using different boundary conditions, use of different polarizations, different sampling techniques and frequency sweeps. There are various methodologies used to calibrate the system such that measurements are stable and reconstruction noise is reduced. Once the scattering data are collected and calibrated, the data is fed into a nonlinear inversion algorithm such that the OI can be approximated. The nonlinear algorithms need to handle two important mathematical issues associated with MWT: nonlinearity and ill-posedness. Nonlinearity is handled by utilizing iterative reconstruction algorithms such as Gauss-Newton and modified gradient methods. The ill-posedness is usually handled by incorporating regularization techniques into the inversion algorithms. Many proposed algorithms have been proposed to reconstruct and pseudo-image the OI, however current state of the art MWT is not capable of routine and robust reconstruction of in vivo dielectric properties in subjects.
This prior work has focused on attempting to reconstruct a dielectric image of the patient from the limited microwave scattering. There remains a need for a process and system to utilize multiple sources, some of which may be more readily available or cheaper to procure, to provide information relating to a patient's condition. Additionally, there is a need for learning from large populations to assist with the detection of dielectric changes in matter and subjects.
Embodiments described herein relate generally to a cheap, non-ionizing, portable microwave imaging system that utilizes big data and machine learning to detect the presence of disease. There remains a need for low cost, rapidly deployable, and readily scalable solutions for the evaluation of acute neurologic disorders in remote and pre-clinical settings, and thus we propose a radically different, ultrafast, plug-and-play diagnostic technology utilizing low energy radiofrequency waves to probe dielectric changes (e.g., macroscopic dielectric changes, microscopic dielectric changes, etc.) in the brain. Dielectrography, which leverages directly measurable changes in the dielectric properties of tissues to identify disease-specific “signatures” of neurologic disease. These signatures can be learned from electromagnetic scattering measurements acquired from an array of wide-band antennas operating in the conventional cellular spectrum (0.5-6 GHz), and ground truth for training dielectrographic measurements can be obtained from expert neuroimaging review (CT/MRI) occurring contemporaneously in the routine clinical setting. Physical changes in tissue electromagnetic properties (EP) caused by ischemic and hemorrhagic stroke can cause measurable changes in electromagnetic (EM) wave scattering and breakage in scattering symmetry that can be learned in a meaningful way to foster intelligent diagnoses.
Advances in supervised learning techniques enable identification of dielectrography signatures in the presence of neurological disease supporting development of a novel diagnostic instrument. The systems and methods of the present disclosure circumvent conventional underdetermined, ill-posed and non-linear challenges associated with conventional EP tomography approaches in favor of a mathematically overdetermined, biologically and clinically relevant technology.
The open source, low-cost modular design is readily scalable and ideal for large-scale dissemination. Big-data methodologies can be used to better train deep learning architectures, providing a population-wide representation of disease signatures. The fast and safe mode of operation, and high portability uniquely position dielectrography for use in conventionally unattainable high-acuity and pre-clinical environments for which urgent detection is imperative.
At least one aspect of the present disclosure is directed to a system of detecting macroscopic dielectric changes in matter. The system includes a data collection array to collect microwave scattering data. The system includes a machine learning device. The machine learning device is configured to receive the microwave scattering data, patient information, data corresponding to at least one of a presence of disease, absence of disease, or one or more disease features (e.g., size, location, texture, etc.). The machine learning device is configured to analyze the microwave scattering data, patient information, and imaging modality data. The machine learning device is configured to output at least one of a predicted disease state and predicted features based on the analyzed microwave scattering data, patient information, and imaging modality data. The machine learning device is configured to compare the imaging modality data corresponding to the at least one of the presence of disease, absence of disease, or one or more disease features to the at least one of the predicted disease state and predicted features. The machine learning device is configured to use the comparison as an input into the machine learning device.
In some embodiments, the imaging modality data comprises at least one of MR data, CT data, and PET data. In some embodiments, the machine learning device is a deep neural network machine learning device. In some embodiments, the data collection array comprises at least one of wide-band antennas and field programmable gate arrays. In some embodiments, the predicted disease state is macroscopic arrangements in dielectric properties (e.g., hemorrhage). In some embodiments, the patient information comprises at least one of patient medical history, age, dimensions of body, and sex.
At least one aspect of the present disclosure is directed to a system of reconstructing a dielectric image. The system includes a data collection array to collect microwave scattering data. The system includes a machine learning device. The machine learning device is configured to receive the microwave scattering data. The machine learning device is configured to analyze the microwave scattering data. The machine learning device is configured to output a generated image based on the analyzed microwave scattering data. The machine learning device is configured to identify at least one of the presence of disease, absence of disease, or one or more disease features from the generated image.
In some embodiments, the machine learning device is trained and validated based on patient information and imaging modality data. In some embodiments, the machine learning device is a deep neural network machine learning device. In some embodiments, the data collection array comprises at least one of wide-band antennas and field programmable gate arrays. In some embodiments, the machine learning device is configured to select a reconstruction technique based on a population size. In some embodiments, the system is integrated into a helmet. In some embodiments, the machine learning device is configured to output the predicted disease state based on one or more broadband microwave frequencies. In some embodiments, the machine learning device includes a classification network and a discrimination network.
At least one aspect of the present disclosure is directed to a method of detecting macroscopic dielectric changes in matter (e.g., structures, subjects, etc.). The method includes obtaining microwave scattering data from a data collection array. The method includes receiving, by a machine learning device, the microwave scattering data, patient information, and imaging modality data corresponding to at least one of a presence of disease, absence of disease, or one or more disease features. The method includes analyzing, by the machine learning device, the microwave scattering data, patient information, and imaging modality data. The method includes outputting, by the machine learning device, at least one of a predicted disease state and predicted features based on the analyzed microwave scattering data, patient information, and imaging modality data. The method includes comparing, by the machine learning device, the imaging modality data corresponding to the at least one of the presence of disease, absence of disease, or one or more disease features to the at least one of the predicted disease state and predicted feature. The method includes using, by the machine learning device, the comparison as an input into the machine learning device.
In some embodiments, the method includes using correlations between imaging modality data and microwave tomography to output the predicted disease state. In some embodiments, the microwave scattering data comprises at least one of S-parameter data and scattering S-parameter data. In some embodiments, the imaging modality data comprises at least one of MR data, CT data, and PET data. In some embodiments, the machine learning device is a deep neural network machine learning device. In some embodiments, the data collection array comprises at least one of wide-band antennas and field programmable gate arrays. In some embodiments, the predicted disease state is a presence of macroscopic arrangements in dielectric properties (e.g., hemorrhage). In some embodiments, the patient information comprises at least one of patient medical history, age, dimensions of body, and sex. In some embodiments, the method includes outputting the predicted disease state based on one or more broadband microwave frequencies.
It should be appreciated that all combinations of the foregoing concepts and additional concepts discussed in greater detail below (provided such concepts are not mutually inconsistent) are contemplated as being part of the subject matter disclosed herein. In particular, all combinations of claimed subject matter appearing at the end of this disclosure are contemplated as being part of the subject matter disclosed herein.
The foregoing and other features of the present disclosure will become more fully apparent from the following description and appended claims, taken in conjunction with the accompanying drawings. Understanding that these drawings depict only several implementations in accordance with the disclosure and are therefore, not to be considered limiting of its scope, the disclosure will be described with additional specificity and detail through use of the accompanying drawings.
Reference is made to the accompanying drawings throughout the following detailed description. In the drawings, similar symbols typically identify similar components, unless context dictates otherwise. The illustrative implementations described in the detailed description, drawings, and claims are not meant to be limiting. Other implementations may be utilized, and other changes may be made, without departing from the spirit or scope of the subject matter presented here. It will be readily understood that the aspects of the present disclosure, as generally described herein, and illustrated in the figures, can be arranged, substituted, combined, and designed in a wide variety of different configurations, all of which are explicitly contemplated and made part of this disclosure.
EM waves get perturbed by the body. Changes in the dielectric property distribution of tissue can be altered with macroscopic changes due to disease. These diseases leave patterns or “signatures” that can be learned. Described herein are systems and processes that may utilize a multitude of information sources, combined, to detect disease. In one embodiment the process can utilize correlations between imaging data (e.g., MRI) and MWT to provide a pseudo anatomical image and can be used to learn the presence of disease and its features. Furthermore, in some embodiments, a neural net is trained to learn the best reconstruction technique, based on a sufficiently-large population size. Conventional tomography may utilize information generated from EM waves to reconstruct a three-dimensional image of electrical conductivity and relative permittivity. Studies have shown correlations between changes in the electromagnetic properties of tissues and various diseases. Electrical property (EP) tomographic imaging can be categorized as follows: 1) EP mapping leveraging magnetic field disturbances in the MRI environment. These techniques may rely on MM machines to reconstruct the electrical conductivity and permittivity of tissue, but are impractical in mobile settings due to their weight, size, and magnetic field strengths, and are cost prohibitive (˜$1M/Tesla) in many environments; 2) EP mapping using DC currents using a large number of surface detectors. The use of DC currents to estimate EP may more practical for the mobile setting, however, reconstructing EP using this methodology is ill-posed as the induced currents are localized to the surfaces of organs. Due to the inability to interrogate deep seated tissues, the diagnostic value of these techniques is limited; 3) Microwave (MW) imaging has been proposed as a pragmatic (e.g., low-cost, mobile, etc.) solution for probing electromagnetic properties employing multi-antenna MW-emitting devices at various frequencies. Microwave imaging, however, remains unable to deliver a practical diagnostic tool due to technical challenges, mainly the inherent non-linearity and ill-posedness of the inverse EP mapping reconstruction problem. Finite difference time domain (FDTD) electromagnetic field simulations software has been used to assist with the MW imaging reconstruction problem by iteratively solving Maxwell's equations on a large dataset of anatomical models, however, challenges to modeling the architectural complexity of the human brain limit its generalizability and clinical applicability. Furthermore, there remains a lack of big-data solutions to foster development of a practical and intelligent diagnostic tool equipped for learning from large population sizes.
In one embodiment, information is received from incident and scattering S-parameter measurements acquired from a multitude of antennas at different frequencies and general data such as patient history, dimensions of the body part being imaged, and more. The received information is used to help learn patterns of disease. Learning is conducted with the assistance from imaging modalities such as, but not limited to, MR, CT, PET, etc., which provide the ground truth regarding the presence of the disease as well as spatial information of the disease. Imaging modalities can include imaging modality data. For example, imagining modality data can include MR data, CT data, or PET data. Using a large dataset of patients and diseases, optimal detection of disease is possible. In some embodiments, the system includes a hardware acquisition constituted from an array of wide-band antennas and FPGA-based acquisition system, as well as a deep neural network machine learning system responsible for ascertaining the absence or presence of a disease state using a classification-type approach. As a proof of principle, in silico EM simulations were conducted to illustrate the detection of hemorrhage using the combination of the microwave imager, big data, and deep neural network.
Microwave scattering perturbations from ultra wide-band antenna arrays can be used to obtain dielectric signatures of disease. Deep neural networks (DNNs) can be used for: 1) stroke detection (e.g., classification network), and 2) characterization of the hemorrhage location and size (e.g., discrimination network). Dielectric signatures can be learned on a simulated cohort of 666 hemorrhagic stroke and control subjects using 2D stochastic head models. The classification network can yield a stratified K-fold stroke detection accuracy of 94:60%±2:86% with an AUC of 0.996, while the discrimination network can result in a mean squared error of <0.004 cm and <0.02 cm, for the stroke localization and size estimation, respectively. Microwave wide-band scattering information can be used for intelligent diagnostics.
The Ongoing need for low cost, rapidly deployable, and scalable solutions for preclinical stroke diagnosis can motivate the development of novel stroke detection methods. The systems and methods of the present disclosure present a framework for hemorrhagic stroke detection using learned scattering signatures from ultra wide-band antenna arrays between 0.5 and 6 GHz. Microwave scattering perturbations can be used to learn disease signatures. Dielectric signatures of strokes can be obtained using big-data, on a population level. Deep neural networks in combination with scattering information from microwave sources can be used to classify and discriminate the presence of stroke. Results for stochastic in silico, multi-tissue head models representing the dielectric disturbances occurring in the setting of hemorrhagic stroke and derived from electromagnetic field simulations can be shown. The results can introduce an approach for intelligent diagnostics, circumventing conventional image-formation.
To show a proof of principle that is feasible computationally, a 2D axial model of the human head was modelled in Comsol Multiphysics 5.3 finite element modeling (FEM) solver (Burlington, MA, USA). The human head was placed in the center of an 8-antenna structure, each radiating at twelve different frequencies between 1 GHz and 6 GHz. Using the Livelink interface to MatLab (The MathWorks, Inc., Natick, Massachusetts, United States) the geometry of the 2D head and brain was altered, a slight random rotation of the head was introduced and the size/location of a hemorrhagic stroke was introduced. The following parameters were set as follows and a normal distribution between the min and the max values was used using the rand function in Matlab:
After each of these parameters was randomized in Matlab, conductivity and permittivity dielectric properties of: brain, white matter, gray matter, skull, blood and air can be assigned according to the literature values at each microwave frequency to properly simulate the EM wave propagation inside a realistic 2D head. After the tissue properties were assigned, 12 simulations can be conducted at different frequencies between 1 and 6 GHz. These twelve simulations can be conducted for each antenna in the array yielding 96 different simulations per simulated head. After the simulations converged, the complex S-parameter matrix at each frequency can be obtained yielding an 8×8×12 complex measurement matrix for each subject. Overall, ˜2250 subjects can be simulated, and the presence or absence of hemorrhage can be recorded using a value of 0 or 1, respectively.
Once all the simulations converged. The measurement matrix can be flattened, and the data can be divided into a training set of 2046 subjects and 205 for the validation set HO % set for training). The problem can be redesigned as a binary classification problem for detection of the presence or absence of hemorrhage.
Detection accuracy on the validation set can be ˜83%. Further improvement can be implemented by changing the neural network architecture, optimization of hyperparameters, incorporation of patient information such as age, medical history, dimensions of body, sex, etc. Age related change in dielectric properties of tissues were not included in these simulations and can possibly improve the real-life detection accuracy.
In one embodiment, network A, the detection of disease is defined as a classification problem. As shown in
In a second embodiment, network B shown in
The helmet 800 can facilitate rapid, nearly real-time (<10 seconds) acquisition and may be powered from a conventional AC outlet. The helmet 800 is safe for routine use as it has a maximum output power of a mobile phone making it suitable for pre-clinical, high-acuity and biologically harmless longitudinal assessment and monitoring. The helmet 800 can be equipped with a fully modular, 70-antenna configuration leveraging advances in radio-frequency (RF) electronics and acquisition strategies allowing use of ultra-wide-band (UWB) detector design and maintenance of high sensitivity to dielectric disturbances at varying depths. In some embodiments, dielectrography does not rely on ionizing radiation, strong magnetic field, or user dependence. This user-independence and relief from the demands of technical and medical expertise can vastly broaden the reach of diagnostic medical technology, increasing accessibility and practicality of early, advanced diagnosis in underserved areas.
The observed changes in the S-parameters for three different strokes (represented as ovals A, B and C) are summarized in
In some embodiments, the system uses multi-layered deep neural networks with a softmax layer. One example of the advantage of using AI in a supervised learning approach is that patterns of disease can be learned from the imaging data, and neural nets are quite well suited for that. Reinforcement learning techniques or conventional deep neural nets can be used to solve the classification problem that may be used to diagnose a certain medical condition. Encoder-decoder network architectures can be used to solve the tomographic problem where an image is reconstructed. There is a pressing need for safe, low cost, rapidly deployable, and readily scalable solutions for the evaluation of acute neurologic disorders in remote and pre-clinical settings, and thus radically different, ultrafast (e.g., <10 seconds per scan), plug-and-play diagnostic technology is proposed herein, utilizing low energy radiofrequency waves to probe macroscopic dielectric changes in the brain. This technique is referred to as dielectrography, leveraging known and directly measurable changes in the dielectric properties of tissues to identify disease-specific “signatures” of neurologic disease. These “signatures” can be learned using AI from electromagnetic scattering measurements acquired from an array of wide-band antennas operating in the conventional cellular spectrum (700 MHz-6 GHz), and ground truth for training dielectrographic measurements can be obtained from expert neuroimaging review (CT/MRI) occurring contemporaneously in the routine clinical setting.
In one embodiment the system is very portable, thus being capable of being deployed in ambulances, military bases, and third world countries that cannot afford state of the art imaging machines such as MRIs and CT. The Maxwell detector can be “fine-tuned” to detect a wide array of diseases provided there is a gross-change in the dielectric properties of tissues.
In one embodiment, the system can be used to assess hydrocephalus in adults and children, traumatic brain injury identification and progression assessment, and other conditions were there are macroscopic changes to the dielectric tissue distribution that are different than in the underlying population.
Mapping electrical properties (EP) is an area of active research as changes in the EP of tissues correlate with macroscopic disease (e.g., strokes, cancer, etc.). Fast, mobile and low-cost techniques often employ scattering measurements from microwave (MW) sources to map EP of tissues. MW imaging however fails to deliver due to non-linearity and ill-posedness of the inverse reconstruction problem. In on embodiment, these challenges are mitigated by: 1) Rather than producing EP images from underdetermined data, the problem is transformed to an overdetermined classification problem easier to solve; 2) Neural networks and stochastic gradient decent-like techniques are used to overcome non-linear challenges; 3) Big data is used for learning disease “signatures” in a large cohort of patients; and 4) Scattering measurements are conducted over a wide-range of frequencies, which are sensitive to dielectric disturbances at different depths and improve acuity to discern disease states.
Unique attributes of the proposed design include its small form factor, low power, low-cost, portability, lack of ionizing radiation, acquisition speed, and intelligent diagnostic capabilities making it ideally suited for use in rapid response and high acuity contexts, and for both initial and longitudinal diagnostic evaluation.
Microwave tomography (MWT) has historically faced two main challenges: 1) non-linearity of the reconstruction, handled unsuccessfully by iterative reconstruction techniques (e.g., Gauss-Newton), and 2) ill-posedness (e.g., the number of measurements is significantly smaller than the number of unknowns). These conditions defy robust solutions through conventional algorithms; however, the maturation of AI, assimilation of big data and the resources to learn from large populations, and advances in RF electronics supporting fast acquisitions and capacity to conduct broadband frequency excitations can be exploited to enable high performance in vivo MWT. While the conventional MWT problem is underdetermined and conventional regularization (e.g., L1) techniques are employed for image reconstruction, here the problem was relaxed to disease classification, which is computationally better conditioned and biologically relevant. Together with utilization of AI and big-data, the untapped potential of dielectrography can reach full realization. In silico verification of this approach has been confirmed with highly promising results.
In early silico proof-of-concept investigations, a 6-layer deep neural net was utilized to determine the presence or absence of disease. In these trials, the last layer of the neural network had a sigmoid activation function to support a binary classification and a binary cross-entropy loss function was used to drive learning. Alternative approaches may include Decision Trees, Naïve Bayes Classifier Algorithm, etc. that can enhance performance of the classifier.
In one embodiment, the AI model can utilize electromagnetic scattering measurements and patient history, which can be abstracted from anonymized electronic medical records archived for all patients in the system, as model inputs. At the same time, imaging diagnosis, namely the presence of neurological disease (attained from MRI/CT) can be used to train a deep neural network classifier to detect the presence or absence of disease. In exploratory in situ simulations, a deep neural net with 6 fully connected layers correctly detected stroke with ˜85% accuracy on a limited training/test set (including merely 12 frequency measurements). Conventional classification metrics can be used to ensure generalization. Once generalized, the system design and network coefficients can be made openly available (open source) for the benefit of the community. The application can be further expanded to include a broader gamut of neurologic disease, once the framework is implemented and matured.
A total of 666 head models can be generated in simulation using COMSOL Livelink, enabling scripting within the Matlab (Mathworks, Natick, Massachusetts USA) environment. A 2D synthetic, trans-axial slice through the brain, can be modeled based on the 3D Specific Anthropomorphic Mannequin (SAM) model to produce a simulated central slice through the brain at the level of the lateral ventricles. Stochastic modifications to the brain dimensions can be applied to each head model that include variation in the width and length of the head and brain independently, and width and length of the lateral ventricles. The head models can be placed centrally within the antenna array, with random reorientation of the head prescribed relative to the array, to represent varying head orientations occurring when the head is fixated to a plastic stereotactic frame permitted to pivot relative to the antenna array, yet remain immobile with respect to translation. Variations in head anatomy and rotation are summarized in Table 2.
A total of five different tissues can be included in the simulations: skull, blood, cerebrospinal fluid (CSF), average brain, and air. For each frequency of interest between 0.5 and 6 GHz, the dielectric properties can extracted. Random-valued head anatomies can be fed to the DNN representing subject-wise variation of the head, thus presenting the DNN with highly varying conditions that disrupt the S-parameter matrices to facilitate learning. Fifty percent of the heads can be selected at random and intracranial blood can be introduced in random locations, simulating intraparenchymal hemorrhagic strokes. Sizes of the strokes can be independently varied in the x- and y-directions with a 2±0.5 cm (mean±standard deviation) in diameter (ranging between 0.5 and 4 cm, in the x- and y-directions, independently) (Table 2). The stroke x- and y-locations can be selected from a uniform random distribution, with the outer bounds of the distribution changing depending on the size of the brain. A total of 106,560 simulations can be conducted using different head anatomies, frequency sweeps, and antennas driven. Execution of these simulations can be performed in >720 continuous hours of simulation time.
Multi-frequency S-parameter matrices can be exported into python. The dimensions of each complex-valued matrix can be nch×nch×nf, where nch is the number of channels and nf is the number of simulated frequencies. Amplitude and phase of the complex-valued matrices can be reformatted, for each model, to nch2*nf. Training, testing, and validation a the neural net models can be conducted in the Keras environment. Before training, the data across subjects can be split such that 70% of the data are used for training, 15% for validation, and 15% for testing, respectively. Training, testing, and validation data sets can be normalized using a scalar normalizer such that the data was scaled to unit variance. Two distinct DNNs can be used for the stroke diagnosis: the first DNN can be used to classify the presence of hemorrhagic stroke using S-parameter matrix data (“Classification DNN”) and the second DNN can be used to discriminate the stroke location and size inside the head (“Discrimination DNN”).
For both the classification and discrimination networks, a dropout layer with dropout threshold of 7.5% can be introduced between the fully connected layers to reduce over-fitting. Each layer of the classification network can be initialized using a truncated normal distribution with zero mean and standard deviation of 0.75 and the discrimination network can be initialized using a truncated normal distribution with zero mean and standard deviation of 0.05. Model optimization for both networks can be performed using the Adam optimizer. A binary cross-entropy loss function with 20 epochs and a batch size of 100 can be used for the classification network training, while a mean square error (MSE) loss function, 1000 epochs, and a batch size of 100 can be used for discrimination network training. The loss function and accuracy of each trained network as a function of epoch can be plotted. Classification accuracy can be quantified using the stratified K-fold method with 10 splits. A Receiver-Operating Characteristic (ROC) curve can be plotted and the area under the curve (AUC) can be computed. The discrimination network can be evaluated by calculation of the test-set MSE for the errors in predicted x-location and y-location and x-stroke size and y-stroke size. For visualization purposes, results can be compared against the ground truth in randomly-chosen stroke models selected from the test-set. Frequency-dependent stroke dielectric effects and antenna-dependent stroke dielectric effects can be assessed by plotting changes in the magnitude of the S-parameter matrices occurring due to the presence of stroke for 0.5, 1 and 3 GHz and presented dB scale.
The discrimination network can be designed to predict both the size and location of the stroke. Discrimination can be performed via learning patterns from frequency dependent S-parameters from different antenna elements, rather than relying on conventional microwave imaging reconstruction. To better discriminate stroke spatial characteristics, a larger network with more redundancy can be used which can include a 6× larger first layer, compared to the classification network, and a relatively larger number of neurons in the subsequent hidden layers, permitting finer prediction of size and location (
Physical changes in tissue electrical properties (EPs) caused by hemorrhagic stroke can cause measurable changes in EM wave scattering. These changes in the S-parameter matrices can be used to train networks for intelligent diagnosis and discrimination of stroke. Broadband acquisitions between 0.5 and 6 GHz can provide complementary information. Multi-antenna and multi-frequency information shows that higher S-parameter disruptions at higher frequencies can occur due to the higher sensitivity in shallow depths, due to the skin effect, as well as the size of the lesion. These changes can provide information for the DNNs to facilitate learning the presence of stroke as well as the capability to localize and estimate the size of the stroke. The additional contrast arising from a broadband frequency sweep in a simulated cohort can be utilized to facilitate learning and generalization, while circumventing conventional MI reconstruction inverse problems.
Two-dimensional EM simulations can be employed due to computational demands inherent to three-dimensional simulations. More than 100,000 simulations can be conducted to calculate electromagnetic fields and thus S-parameters across all frequency, subject, and antenna permutations. The challenges inherent to engineering realistic, 3D patho-anatomic models for electromagnetic simulation can constitute ongoing motivation for the development of robust experimental methods. Reduction of the dimensionality to two-dimensions to alleviate computational demand can consequently eliminate the z-dimension (e.g., the craniocaudal direction) thus converting the hemorrhage to an affected 2D area of the simulated brain with Maxwell's equations solved in 2D. A transition to three dimensions in realizable antenna array systems could benefit from the utilization of a greater number of antennas, thus capturing 3D spatial detail such as that arising from a 3D hemorrhage volume within the head.
Improvements in the prediction of the presence, location, and size of intracranial hemorrhage can be achieved through machine learning-assisted MI by leveraging the additional contrast arising from broadband microwave frequencies. The machine learning device can output a predicted disease state based on one or more broadband microwave frequencies. The reduction in the number of frequency sweeps and/or coils used can result in acceleration of S-parameter acquisitions to improve patient comfort or time to diagnosis. Further, the selection of a simulated cohort of 666 stroke and control subjects constitutes only an initial approximation of an achievable cohort in contemporary stroke trials and is representative of stroke imaging workflow.
Acquisition systems capable of realizing this approach could employ wide-band antenna arrays encircling the patient's head. Each antenna can be connected to controllable RF/microwave switches routing each antenna to receive or transmit signals and connected to a multi-port vector network analyzer (VNA). Alternatively, SDRs can be utilized alongside directional couplers to gather wide-band scattering information. Such systems can be two orders of magnitude less expensive than state of the art imaging systems (e.g. clinical MRI).
Intracerebral hemorrhage was simulated. Blood can provide a stronger dielectric contrast than ischemia, which may be more pathologically variegated at the tissue level, challenging development of realistic simulated conditions. The detection accuracy for hemorrhage >1 cm in diameter (see Table 2 for stroke variation in x- and y-directions) was 94.6%, and therefore the dependable exclusion of hemorrhage can serve as a screening tool in the acute stroke setting to facilitate rapid thrombolysis as per current guidelines in early-presenting patients. This latter point bears emphasis, as intravenous thrombolysis remains the only approved pharmacologic therapy in acute ischemic stroke patients, but with rare exceptions cannot be administered beyond the initial, hyperacute window. Consequently, only 12% of patients receive critical intravenous thrombolysis, despite a 30% improvement in outcomes among those treated.
A method of detecting dielectric changes in matter can include obtaining microwave scattering data from a data collection array. The microwave scattering data can include at least one of incidence S-parameter data and scattering S-parameter data. The S-parameter data can include S-parameter matrices. The S-parameter matrices can include information of reflected and forward waves. The data collection array can include field programmable gate arrays, software defined radio, network analyzers, and other hardware.
The method can include receiving, by a machine learning device, the microwave scattering data, patient information, and imaging modality data corresponding to at least one of a presence of disease, absence of disease, or one or more disease features. The imaging modality data can include at least one of MR data, CT data, and PET data. The machine learning device can include a deep neural network machine learning device. The patient information can include at least one of patient medical history, age, dimensions of body, and sex.
The method can include analyzing, by the machine learning device, the microwave scattering data, patient information, and imaging modality data. The method can include outputting, by the machine learning device, at least one of a predicted disease state and predicted features based on the analyzed microwave scattering data, patient information, and imaging modality data. The at least one of the predicted disease state and predicted features can include a presence of hemorrhage.
The method can include comparing, by the machine learning device, the imaging modality data corresponding to the at least one of the presence of disease, absence of disease, or one or more disease features to at least one of the predicted disease state and predicted features. The method can include using, by the machine learning device, the comparison as an input into the machine learning device.
In some embodiments, the method can include using correlations between imaging modality data and microwave tomography to output the at least one of the predicted disease state and predicted features. In some embodiments, the machine learning device can output the at least one of the predicted disease state and predicted features based on one or more broadband microwave frequencies. In some embodiments, the machine learning device can include a classification network and a discrimination network. In some embodiments, the method can include outputting the at least one of the predicted disease state and predicted features based on one or more broadband microwave frequencies.
Systems and methods for hemorrhagic stroke detection and discrimination by means of learning disease dielectric signatures from S-parameter measurements using wide-band antenna arrays are described. Two DNN architectures were designed and tuned for classification and discrimination in presence of stroke in stochastic in silico, multi-compartment head models representing the tissue dielectric disturbances occurring in the setting of hemorrhagic stroke and derived from electromagnetic field simulations. The systems and methods of the present disclosure offer a potential strategy for rapid stroke detection and management, circumventing conventional image generation.
As shown in
System 100 may also include a display or output device, an input device such as a key-board, mouse, touch screen or other input device, and may be connected to additional systems via a logical network. Many of the embodiments described herein may be practiced in a networked environment using logical connections to one or more remote computers having processors. Logical connections may include a local area network (LAN) and a wide area network (WAN) that are presented here by way of example and not limitation. Such networking environments are commonplace in office-wide or enterprise-wide computer networks, intranets and the Internet and may use a wide variety of different communication protocols. Those skilled in the art can appreciate that such network computing environments can typically encompass many types of computer system configurations, including personal computers, hand-held devices, multi-processor systems, microprocessor-based or programmable consumer electronics, network PCs, minicomputers, mainframe computers, and the like. Embodiments of the disclosure may also be practiced in distributed computing environments where tasks are performed by local and remote processing devices that are linked (either by hardwired links, wireless links, or by a combination of hardwired or wireless links) through a communications network. In a distributed computing environment, program modules may be located in both local and remote memory storage devices.
Various embodiments are described in the general context of method steps, which may be implemented in one embodiment by a program product including computer-executable instructions, such as program code, executed by computers in networked environments. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Computer-executable instructions, associated data structures, and program modules represent examples of program code for executing steps of the methods disclosed herein. The particular sequence of such executable instructions or associated data structures represents examples of corresponding acts for implementing the functions described in such steps.
Software and web implementations of the present invention could be accomplished with standard programming techniques with rule based logic and other logic to accomplish the various database searching steps, correlation steps, comparison steps and decision steps. It should also be noted that the words “component” and “module,” as used herein and in the claims, are intended to encompass implementations using one or more lines of software code, and/or hardware implementations, and/or equipment for receiving manual inputs.
As used herein, the singular forms “a”, “an” and “the” include plural referents unless the context clearly dictates otherwise. Thus, for example, the term “a member” is intended to mean a single member or a combination of members, “a material” is intended to mean one or more materials, or a combination thereof.
As used herein, the terms “about” and “approximately” generally mean plus or minus 10% of the stated value. For example, about 0.5 would include 0.45 and 0.55, about 10 would include 9 to 11, about 1000 would include 900 to 1100.
It should be noted that the term “exemplary” as used herein to describe various embodiments is intended to indicate that such embodiments are possible examples, representations, and/or illustrations of possible embodiments (and such term is not intended to connote that such embodiments are necessarily extraordinary or superlative examples).
The terms “coupled,” “connected,” and the like as used herein mean the joining of two members directly or indirectly to one another. Such joining may be stationary (e.g., permanent) or moveable (e.g., removable or releasable). Such joining may be achieved with the two members or the two members and any additional intermediate members being integrally formed as a single unitary body with one another or with the two members or the two members and any additional intermediate members being attached to one another.
It is important to note that the construction and arrangement of the various exemplary embodiments are illustrative only. Although only a few embodiments have been described in detail in this disclosure, those skilled in the art who review this disclosure will readily appreciate that many modifications are possible (e.g., variations in sizes, dimensions, structures, shapes and proportions of the various elements, values of parameters, mounting arrangements, use of materials, colors, orientations, etc.) without materially departing from the novel teachings and advantages of the subject matter described herein. Other substitutions, modifications, changes and omissions may also be made in the design, operating conditions and arrangement of the various exemplary embodiments without departing from the scope of the present invention.
While this specification contains many specific implementation details, these should not be construed as limitations on the scope of any inventions or of what may be claimed, but rather as descriptions of features specific to particular implementations of particular inventions. Certain features described in this specification in the context of separate implementations can also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation can also be implemented in multiple implementations separately or in any suitable subcombination. Moreover, although features may be described above as acting in certain combinations and even initially claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.
This application is a continuation-in-part of International Application No. PCT/US2020/013204 filed Jan. 10, 2020, which claims priority to U.S. Provisional Application No. 62/791,669 filed Jan. 11, 2019. All of the above applications are incorporated herein by reference in their entirety for all purposes.
Number | Name | Date | Kind |
---|---|---|---|
20110319746 | Kochba et al. | Dec 2011 | A1 |
20120190977 | Persson | Jul 2012 | A1 |
20160287131 | McCollough et al. | Oct 2016 | A1 |
20170007148 | Kaditz | Jan 2017 | A1 |
20170160363 | Chen | Jun 2017 | A1 |
20180253531 | Sharma | Sep 2018 | A1 |
20190148021 | Styner | May 2019 | A1 |
20210090212 | Piat | Mar 2021 | A1 |
20230237649 | Dillman | Jul 2023 | A1 |
Entry |
---|
Belle, Ashwin et al., Big Data Analytics in Healthcare, Jan. 5, 2015, Hindawi Publishing Corporation, BioMed Research International vol. 2015, pp. 1-16 (Year: 2015). |
Abubakar, et al., “Imaging of biomedical data using a multiplicative regularized contrast source inversion method,” IEEE Transactions on Microwave Theory and Techniques 50(7), pp. 1761-1771 (2002). |
Agilent Technologies, “Agilent Understanding and Improving Network Analyzer Dynamic Range,” Application Note 1363-1, retrieved from http://anlage.umd.edu/Microwave%20Measurements%20for%20Personal%20Web%20Site/5980-2778EN.pdf, 8 pages (2000). |
Alqadami, et al., “Wearable Electromagnetic Head Imaging System Using Flexible Wideband Antenna Array Based on Polymer Technology for Brain Stroke Diagnosis,” IEEE Transactions on Biomedical Circuits and Systems 13(1), pp. 124-134 (2019). |
Brown, “Electrical impedance tomography (EIT): A review,” Journal of Medical Engineering & Technology 27(3), pp. 97-108 (2003). |
Catapano, et al., “On Simple Methods for Shape Reconstruction of Unknown Scatterers,” IEEE Transactions on Antennas and Propagation 55(5), pp. 1431-1436 (2007). |
Chalouhi, et al., “Review of Cerebral Aneurysm Formation, Growth, and Rupture,” Stroke 44(12), pp. 3613-3622 (2013). |
Colton & Kirsch, “A simple method for solving inverse scattering problems in the resonance region,” Inverse Problems 12(4), pp. 383-393 (1996). |
Cooley, et al., “A portable scanner for magnetic resonance imaging of the brain,” Nature Biomedical Engineering 5, pp. 229-239 (2021). |
Dority & Oldham, “Subarachnoid Hemorrhage: An Update,” Anesthesiology Clinics 34(3), pp. 577-600 (2016). |
Fassbender, et al., “Mobile stroke units for prehospital thrombolysis, triage, and beyond: benefits and challenges,” The Lancet Neurology 16(3), pp. 227-237 (2017). |
Fei, et al., “Ultrahigh Frequency (100?MHz-300?MHz) Ultrasonic Transducers for Optical Resolution Medical Imagining,” Scientific Reports 6:28360, 8 pages (2016). |
Ferns, et al., “De Novo Aneurysm Formation and Growth of Untreated Aneurysms: A 5-Year MRA Follow-Up in a Large Cohort of Patients With Coiled Aneurysms and Review of the Literature,” Stroke 42(2), pp. 313-318 (2011). |
Gabriel, et al., “The dielectric properties of biological tissues: I. Literature survey,” Physics in Medicine & Biology 41(11), pp. 2231-2249 (1996). |
Gabriel, et al., “The dielectric properties of biological tissues: II. Measurements in the frequency range 10 Hz to 20 GHz,” Physics in Medicine & Biology 41(11), pp. 2251-2269 (1996). |
Gabriel, et al., “The dielectric properties of biological tissues: III. Parametric models for the dielectric spectrum of tissues,” Physics in Medicine & Biology 41(11), pp. 2271-2293 (1996). |
Geethanath & Vaughan, “Accessible magnetic resonance imaging: A review,” Journal of Magnetic Resonance Imaging 49(7), pp. e65-e77 (2019). |
Giroud, et al., “The worldwide landscape of stroke in the 21st century,” The Lancet 383(9913), pp. 195-197 (2014). |
Gordon, et al., “Anthropometric Survey of U.S. Army Personnel: Summary Statistics, Interim Report for 1988,” Defense Technical Information Center—Technical report NATIC/TR-89/027, retrieved from https://apps.dtic.mil/sti/citations/ADA209600, 335 pages (1988). |
Hemphill, et al., “Guidelines for the Management of Spontaneous Intracerebral Hemorrhage: A Guideline for Healthcare Professionals From the American Heart Association/American Stroke Association,” Stroke 46(7), pp. 2032-2060 (2015). |
Jamlos, et al., “High performance novel UWB array antenna for brain tumor detection via scattering parameters in microwave imaging simulation system,” 9th European Conference on Antennas and Propagation (EuCAP), 5 pages (2015). |
Kingma & Ba, “Adam: A Method for Stochastic Optimization,” 3rd International Conference for Learning Representations, retrieved from https://arxiv.org/abs/1412.6980, 15 pages (2015). |
Kirsch, “Characterization of the shape of a scattering obstacle using the spectral data of the far field operator,” Inverse Problems 14(6), pp. 1489-1512 (1998). |
Lantigua, et al., “Subarachnoid hemorrhage: who dies, and why?,” Critical Care 19:309, 10 pages (2015). |
Lin & Clarke, “Microwave imaging of cerebral edema,” Proceedings of the IEEE 70(5), pp. 523-524 (1982). |
Miglioretti, et al., “The Use of Computed Tomography in Pediatrics and the Associated Radiation Exposure and Estimated Cancer Risk,” JAMA Pediatrics 167(8), pp. 700-707 (2013). |
Mustafa, et al., “Novel Preprocessing Techniques for Accurate Microwave Imaging of Human Brain,” IEEE Antennas and Wireless Propagation Letters 12, pp. 460-463 (2013). |
Powers, et al., “Guidelines for the Early Management of Patients With Acute Ischemic Stroke: 2019 Update to the 2018 Guidelines for the Early Management of Acute Ischemic Stroke: A Guideline for Healthcare Professionals From the American Heart Association/American Stroke Association,” Stroke 50(12), pp. e344-e418 (2019). |
Salzberg, “On Comparing Classifiers: Pitfalls to Avoid and a Recommended Approach,” Data Mining and Knowledge Discovery 1, pp. 317-328 (1997). |
Saver, “Time Is Brain—Quantified,” Stroke 37(1), pp. 263-266 (2006). |
Semenov & Corfield, “New Electromagnetic Methods and Applications of Antennas in Biomedicine,” International Journal of Antennas and Propagation 2008:254830, 8 pages (2008). |
Semenov, “Microwave tomography: review of the progress towards clinical applications,” Philosophical Transactions A 367(1900), pp. 3021-3042 (2009). |
Semenov, et al., “Microwave-tomographic imaging of the high dielectric-contrast objects using different image-reconstruction approaches,” IEEE Transactions on Microwave Theory and Techniques 53(7), pp. 2284-2294 (2005). |
Semenov, et al., “Neuroimaging patterns of cerebral hyperperfusion,” Journal of Physics: Conference Series 886:012014, 5 pages (2017). |
Skodvin, et al., “Cerebral Aneurysm Morphology Before and After Rupture: Nationwide Case Series of 29 Aneurysms,” Stroke 48(4), pp. 880-886 (2017). |
Sneth, et al., “Assessment of Brain Injury Using Portable, Low-Field Magnetic Resonance Imaging at the Bedside of Critically Ill Patients,” JAMA Neurology 78(1), pp. E1-E7 (2020). |
Stancombe, et al., “Portable Microwave Head Imaging System Using Software-Defined Radio and Switching Network,” IEEE Journal of Electromagnetics, RF and Microwaves in Medicine and Biology 3(4), pp. 284-291 (2019). |
Trefna & Persson, “Antenna array design for brain monitoring,” 2008 IEEE Antennas and Propagation Society International Symposium, 4 pages (2008). |
Van Den Berg & Kleinman, “A contrast source inversion method,” Inverse Problems 13(6), pp. 1607-1620 (1997). |
Wald, et al., “Low-cost and portable MRI,” Journal of Magnetic Resonance Imaging 52(3), pp. 686-696 (2020). |
World Health Organization, “Global Status Report on Noncommunicable Diseases 2014,” retrieved from https://apps.who.int/iris/bitstream/handle/10665/148114/9789241564854_eng.pdf, 302 pages (2014). |
Ahmed, et al., “Advanced Microwave Imaging,” IEEE Microwave Magazine 13(6), pp. 26-43 (2012). |
Amineh, et al., “Three-Dimensional Near-Field Microwave Holography Using Reflected and Transmitted Signals,” IEEE Transactions on Antennas and Propagation 59(12), pp. 4777-4789 (2011). |
Arunachalam, et al., “A Computational Investigation of Microwave Breast Imaging Using Deformable Reflector,” IEEE Transactions on Biomedical Engineering 55(2), pp. 554-562 (2008). |
Fear, et al., “Enhancing breast tumor detection with near-field imaging,” IEEE Microwave Magazine 3(1), pp. 48-56 (2002). |
Hassanien & Kim, “Breast cancer MRI diagnosis approach using support vector machine and pulse coupled neural networks,” Journal of Applied Logic 10(4), pp. 277-284 (2012). |
International Search Report & Written Opinion for PCT/US2020/013204 dated Apr. 8, 2020, 8 pages. |
Kaatze, “Measuring the dielectric properties of materials. Ninety-year development from low-frequency techniques to broadband spectroscopy and high-frequency imaging,” Measurement Science and Technology 24(1), 31 pages (2012). |
Li, et al., “Deep Learning Based Imaging Data Completion for Improved Brain Disease Diagnosis,” International Conference on Medical Image Computing and Computer-Assisted Intervention 2014, pp. 305-312 (2014). |
Meaney, et al., “A clinical prototype for active microwave imaging of the breast,” IEEE Transactions on Microwave Theory and Techniques 48(11), pp. 1841-1853 (2000). |
Mohammed, et al., “Microwave System for Head Imaging,” IEEE Transactions on Instrumentation and Measurement 63(1), pp. 117-123 (2014). |
Mojabi, et al., “Microwave tomography techniques and algorithms: A review,” 15th International Symposium on Antenna Technology and Applied Electromagnetics, 4 pages (2012). |
Nikolova, “Microwave imaging for breast cancer,” IEEE Microwave Magazine 12(7), pp. 78-94 (2011). |
Semenov, et al., “Microwave tomography: review of the progress towards clinical applications,” Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 367(1900), pp. 3021-3042 (2009). |
Semenov, et al., “Microwave-tomographic imaging of the high dielectric-contrast objects using different image-reconstruction approaches,” EEE Transactions on Microwave Theory and Techniques 53(7), pp. 2284-2294 (2005). |
Wang, et al., “Imaging of 3-D Dielectric Objects Using Far-Field Holographic Microwave Imaging Technique,” Progress In Electromagnetics Research B 61, pp. 135-147 (2014). |
Wang, et al., “Three-Dimensional Far-Field Holographic Microwave Imaging: an Experimental Investigation of Dielectric Object,” Progress In Electromagnetics Research B 61, pp. 169-184 (2014). |
Number | Date | Country | |
---|---|---|---|
20210335493 A1 | Oct 2021 | US |
Number | Date | Country | |
---|---|---|---|
62791669 | Jan 2019 | US |
Number | Date | Country | |
---|---|---|---|
Parent | PCT/US2020/013204 | Jan 2020 | US |
Child | 17370983 | US |