Magnetic resonance electrical properties tomography
The physically coupled neural network approach for MREPT optimizes conductivity map reconstruction by generating stabilization coefficients, addressing uniformity assumptions and enhancing noise robustness and generalizability, leading to improved accuracy in MREPT.
Patent Information
- Application Number
- PCT/JP2024/016877
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-08
- Filing Date
- 2024-05-02
- Publication Date
- 2024-11-14
AI Technical Summary
Magnetic resonance electrical property tomography (MREPT) faces inaccuracies and numerical errors due to uniformity assumptions in conductivity reconstruction, leading to boundary artifacts and requiring empirical selection of stabilization factors, which limits generalizability and noise robustness, especially in medical imaging where data collection is challenging.
A physically coupled neural network approach that uses a stabilized convective reaction partial differential equation (cr-EPT) with diffusion and convection coefficients generated by separate neural networks, allowing for local or global coefficient maps selection based on noise conditions, optimizing conductivity map reconstruction.
The method achieves high accuracy and robustness in reconstructing conductivity maps even in noisy environments and deviant cases, improving generalizability and reducing artifacts compared to conventional methods.
Smart Images

Figure JP2024016877_14112024_PF_FP_ABST
Abstract
Description
Magnetic Resonance Electrical Properties Tomography
[0001] The present invention relates to magnetic resonance electrical properties tomography, and more particularly to physically coupled neural network magnetic resonance electrical properties tomography, for image reconstruction using electrical properties as tissue contrast.
[0002] The electrical properties (EPs; conductivity σ, permittivity ε) of human tissue are quantitative biomarkers useful for identifying various pathologies and their onset. Images using EPs as tissue contrast (conductivity maps, permittivity maps) may be applicable to the early diagnosis of cancerous tissue.
[0003] Magnetic resonance electric properties tomography (MREPT) is an imaging diagnostic technique that reconstructs EPs using radio frequency magnetic fields (RF fields) measured by an MRI device. Conventionally, MREPT is known to reconstruct conductivity maps by numerically solving a standard analytical model (std-EPT) based on the Maxwell equations. While such analytical MREPT numerical solutions are simple, they suffer from the drawback of boundary artifacts at tissue boundaries, resulting in low reconstruction accuracy due to inaccuracies and numerical errors in the homogeneity assumption underlying the analytical model, i.e., the assumption that electrical conductivity is constant.
[0004] Non-Patent Document 1 proposes a method for flexibly solving the reconstruction problem by removing the assumption of uniformity from an analytical model based on the Maxwell equations. In detail, the reconstruction problem is formulated as a convection-reaction partial differential equation (cr-EPT) and is solved by discretizing it for the region of interest (ROI). This results in unknowns that can be solved without spatial constraints, making the assumption of uniformity unnecessary. However, this method does not solve the B measured by the MRI device. 1 If the field resolution is low, numerical artifacts may occur.
[0005] The effects of numerical artifacts in the cr-EPT method can be reduced by adding a diffusion term that acts as a viscous regularizer. However, if the diffusion coefficient in the diffusion term is large, the diffusion term dominates the cr-EPT, blurring the distribution of EPs and reducing contrast. Therefore, to achieve a trade-off between suppressing rapid fluctuations in the reconstruction of EPs and reducing contrast, it is necessary to provide an appropriate diffusion coefficient for the sample.
[0006] In Non-Patent Document 2, a method is proposed to add a convection coefficient to adjust the weight of the convection term in addition to the diffusion coefficient to the cr-EPT in order to correct boundary artifacts. In detail, Non-Patent Document 2 proposes a method of adding a single convection coefficient empirically selected for the entire ROI, a so-called global convection coefficient map (d∇ 2 ρ+c∇φ・∇ρ+(∇ 2 φ)ρ=ωμ 0 However, different convection coefficients may be required in different regions (boundary and non-boundary) within the ROI to improve reconstruction accuracy. Therefore, it is necessary to assign an appropriate convection coefficient depending on the sample, and also to consider the coordination with the diffusion coefficient.
[0007] FS Hafalir, OF Oran, N. Gurler, and YZ Ider, “Convection-reaction equation based magnetic resonance electrical properties tomography (crmrept),” IEEE Trans Med Imaging, vol. 33, no. 3, pp. 777-793, Mar. 2014. YZ Ider and MN Akyer, “Properties and implementation issues of phase based cr-MRECT for conductivity imaging,” in Proc. Intl. Soc. Mag. Reson. Med. 28, 2020, p. 3190.S. Mandija, EF Meliad`o, NRF Huttinga, PR Luijten, and CAT van den Berg, “Opening a new window on mr-based electrical properties tomography with deep learning,” Scientific Reports, vol. 9, no. 1, Dec. 2019.
[0008] As mentioned above, artifacts associated with analytical MREPT numerical solutions can be reduced by adding stabilization coefficients to the analytical model, i.e., appropriate diffusion and convection coefficients depending on the sample. However, because these stabilization coefficients have been selected empirically, it has been difficult to improve the generalizability of the analytical model.
[0009] In addition, in Non-Patent Document 3, in order to avoid artifacts associated with the above-mentioned analytical numerical solution of MREPT, B is calculated from training samples without using Maxwell's equations. + 1A data-driven, end-to-end neural network (NN)-based MREPT (NN-EPT) has been proposed, which learns a function that maps fields to EPs. However, the NN-EPT proposed in Non-Patent Document 3 is a physics-uncoupled approach that does not capture the wave physics recognized in analytical MREPT. Therefore, a huge amount of training sample data is required to ensure robustness to new samples, i.e., generalizability. Collecting such a large amount of data is particularly difficult in the medical field. In addition, robustness against various types of noise is required in MRI to improve reconstruction accuracy.
[0010] The present invention has been made in light of the above-mentioned problems, and has an object to provide magnetic resonance electrical characteristics tomography that is excellent in noise robustness and general applicability.
[0011] The inventors have been conducting intensive research into the above-mentioned problems and have discovered that by constructing a physically coupled neural network electrical properties tomography system that includes a first means for reconstructing a conductivity map using a physical model based on wave physics with the diffusion coefficient ρ and the convection coefficient β as coefficients, and a second means for learning the diffusion coefficient ρ and the convection coefficient β using a data-driven machine learning model that utilizes parameters of the physical model, it is possible to achieve magnetic resonance electrical properties tomography that is highly noise robust and generalizable, which led to the completion of the present invention.
[0012] In addition, the physical model is tr ・∇γ)+γ∇ 2 φ tr -ρ∇ 2 γ=ωμ 0 and the parameter is φ tr , ∇φ tr , ∇ 2 φ tr where φ tr is the normalized transmission phase. γ is the inverse of the conductivity σ (γ=1 / σ). ω is the Larmor frequency. μ 0 is the magnetic permeability of a vacuum.
[0013] The machine learning model may also be a neural network.
[0014] Alternatively, the diffusion coefficient ρ and the convection coefficient β may be calculated using separate neural networks.
[0015] The apparatus may also include a third means for determining a difference between the conductivity map reconstructed by the first means and the ground truth conductivity map, and feeding the difference back to the second means.
[0016] Furthermore, the diffusion coefficient ρ and the convection coefficient β may each be a global coefficient map or a local coefficient map.
[0017] Furthermore, the diffusion coefficient ρ and the convection coefficient β may be selectable as either a global coefficient map or a local coefficient map depending on the noise condition of the test sample.
[0018] The magnetic resonance tomography of the present invention provides a magnetic resonance tomography with excellent noise robustness and generalizability, specifically, it can reconstruct a conductivity map with high accuracy even in cases where the test sample deviates from the training sample or in noisy environments.
[0019] 1A and 1B are diagrams showing the configuration of magnetic resonance electrical properties tomography (MRI), i.e., physically coupled neural network electrical properties tomography (PCNN-EPT), according to the present invention. (a) shows an artificial irregular shape (AIG) dataset, (b) shows a digital human head (DHH) dataset, and (c) shows a dataset of a test sample obtained by modifying the DHH dataset. (a) shows a phantom consisting of a cylindrical sample with an AIG and a birdcage coil mounted thereon, and (b) shows a phantom consisting of a DHH and a birdcage coil mounted thereon. (b) shows conductivity maps reconstructed in generalization experiments 1 to 3 using non-learning numerical analysis models (std-EPT, cr-EPT), a physically uncoupled direct learning model (NN-EPT), and physically coupled learning models (PCNN-global-EPT, PCNN-local-EPT) according to the present invention. This figure shows the local coefficient maps, normalized local coefficient maps, and line profiles generated by PCNN-local-EPT in the sixth column of Figure 4 for five types of experiments (a) to (e) corresponding to generalization experiments 1 to 3. This figure shows the mean and standard deviation of the SSIM values of the conductivity maps reconstructed by each model for cases (1) to (9) for five types of experiments (a) to (e) corresponding to generalization experiments 1 to 3. This figure shows the mean and standard deviation of the SSIM values of the conductivity maps reconstructed by each model while decreasing the signal-to-noise ratio (SNR) of various samples from 200 to 0 to examine noise robustness in generalization experiments 1 to 3. This figure shows the conductivity maps reconstructed by each model for the conductivity map of training sample #10 of the AIG dataset at four SNR levels (SNR = ∞, 150, 100, 50). The first row shows the conductivity map of training sample #4 of the AIG dataset, divided into boundary and non-boundary parts; the second row shows the local diffusion coefficient map divided and generated by PCNN-local-EPT for the three diagonal circular structures and their boundary and non-boundary parts; and the third row shows the local convection coefficient map divided and generated in the same way.(a) is a graph comparing the distribution of the average diffusion coefficient for conductivity values in the boundary and non-boundary regions of the test sample, and (b) is a graph comparing the distribution of the average convection coefficient. (c) is a graph comparing the aggregated results of diffusion coefficient values for conductivity values in the boundary and non-boundary regions of the test sample, and (d) is a graph comparing the aggregated results of convection coefficient values. The first row shows a comparison of the reconstruction results of phantom scans trained using the DHH dataset, and the second row shows a comparison of the reconstruction results of phantom scans trained using the AIG dataset. (a) to (e) are graphs showing the average and standard deviation of the SSIM values of the reconstructed conductivity maps under various outer boundary setting (OBS) conditions in generalization experiments 1 to 3. (a) to (e) are graphs showing the results of generalization experiments 1 to 3, using either MSE or SSIM as the loss function during training of the black-box NN-EPT, with φ as the input. tr When using ∇ 2 φ tr , ∇φ tr , φ tr 10 is a graph showing the average and standard deviation of the SSIM values of the conductivity map reconstructed when using (a) PCNN-local-EPT, (b) PCNN-global-EPT, (c) NN-EPT, and (d) are graphs showing the loss curves during learning for PCNN-local-EPT. Note that (d) shows the operation of the learning rate reduction mechanism (see solid line) that prevents the loss curve from getting stuck in a local minimum (see dashed line) during learning for PCNN-local-EPT and reduces the loss curve.
[0020] Hereinafter, embodiments of the present invention will be described with reference to the drawings. However, the present invention can be embodied in many different forms and is not limited to the following embodiments and examples.
[0021] As shown in FIG. 1, the magnetic resonance electrical properties tomography in this embodiment is a so-called physically coupled neural network electrical properties tomography (PCNN-EPT) that includes stabilized cr-EPT as a first means for reconstructing a conductivity map using a physical model based on wave physics to which a diffusion coefficient ρ and a convection coefficient β are added as stabilization coefficients, and coefficient generation NNs (neural networks) as a second means for learning the diffusion coefficient ρ and the convection coefficient β using a data-driven machine learning model that utilizes parameters of the physical model, and is therefore capable of providing magnetic resonance electrical properties tomography that is excellent in noise robustness and generalizability.
[0022] Furthermore, the PCCNN-EPT of this embodiment reconstructs a conductivity map using a stabilized cr-EPT to which the diffusion coefficient ρ and convection coefficient β generated by the coefficient generation NNs are added, and by providing feedback to the coefficient generation NNs based on the difference between this reconstructed conductivity map and the ground truth conductivity map, the coefficient generation NNs learn to optimize the diffusion coefficient ρ and the convection coefficient β, thereby reducing artifacts and enabling the PCCNN-EPT to reconstruct a conductivity map with high accuracy even in cases where the test sample deviates from the training sample or in a noisy environment. In other words, even if the PCCNN-EPT does not know the ground truth of the diffusion coefficient ρ and the convection coefficient β for complex-shaped human tissue or an unknown sample, it is possible to generate an optimized diffusion coefficient ρ and a convection coefficient β from the coefficient generation NNs.
[0023] Furthermore, the PCNN-EPT in this embodiment is equipped with a third means for calculating a loss function based on a structural similarity index measure (SSIM) as the difference (reconstruction error) between the conductivity map reconstructed by the stabilized cr-EPT and the ground truth conductivity map, and for performing feedback using an error backpropagation method, and two NNs (U-nets) serving as data-driven machine learning models constituting the coefficient generation NNs are updated by this feedback. Note that the diffusion coefficient ρ and the convection coefficient β are generated separately from the two NNs constituting the coefficient generation NNs.
[0024] Furthermore, for the diffusion coefficient ρ and convection coefficient β added to the stabilized cr-EPT, there is an optimal combination of a global coefficient map and a local coefficient map depending on the noise conditions of the test sample. Therefore, the coefficient generation NNs can select either the global coefficient map or the local coefficient map for the diffusion coefficient ρ and convection coefficient β to be generated depending on the noise conditions of the test sample. Figure 1 shows a case where the diffusion coefficient ρ and convection coefficient β generated by the coefficient generation NNs are each a local coefficient map.
[0025] The "global coefficient map" is a scalar in which the diffusion coefficient ρ and the convection coefficient β are applied globally (to the entire ROI), in other words, a single-value coefficient, while the "local coefficient map" is a matrix in which the diffusion coefficient ρ and the convection coefficient β are applied locally to each pixel in the ROI.
[0026] PCNN-EPT in this embodiment will be described in detail below.
[0027] (Construction of Stabilized Cr-EPT) First, we will explain the construction of stabilized Cr-EPT used in PCNN-EPT. Phase-based magnetic resonance electrical property tomography (MREPT) is a fast model due to its shortened MRI scan time. As shown in the following equation (1), the simplest form corresponds to the standard EPT formulation (std-EPT) and assumes uniform conductivity (∇γ = 0) in the ROI. Furthermore, we assume that the transmit and receive phases are similar and that the transmit phase can be approximated as half of the transmit and receive phase. This assumption is known as the "half-phase assumption," and its impact is not significant at conditions below 3 Tesla.
[0028]
[0029] Here, φ in the above formula (1) tr is the B + 1 is the transmission phase from the field (normalized transmission phase), ω is the Larmor frequency, γ is the inverse of the conductivity σ (γ = 1 / σ), μ 0is the magnetic permeability of a vacuum. Note that due to the assumption of uniform conductivity, boundary artifacts occur in the conductivity map reconstructed by the above equation (1). Such boundary artifacts can be attenuated by solving for the spatial variation of the conductivity map as an unknown.
[0030]
[0031] The formulation of the convection-reaction partial differential equation (cr-EPT) shown in equation (2) above is a method derived from the Maxwell equations. Although cr-EPT removes the assumption of uniform conductivity, numerical artifacts occur in the solution due to numerical instability associated with discretization for the ROI. In cr-EPT, a diffusion term (ρ∇ 2 γ), and we propose to add a convection coefficient β to the convection term in cooperation with the diffusion term to generate a more accurate contrast. This leads to the stabilized cr-EPT shown in Equation (3) below, which is used in PCNN-EPT.
[0032]
[0033] (Flow of PCNN-EPT) Next, the flow of PCNN-EPT will be described. As shown in FIG. 1, the formulation of the physical model in PCNN-EPT is based on the stabilized cr-EPT, which is a phase-based analytical model shown in the above formula (3). The input of PCNN-EPT includes the normalized transmission phase (φ tr ), and its gradient (∇φ tr ), and its Laplacian (∇ 2 φ tr ) PCNN-EPT is mainly composed of two parts: a stabilized cr-EPT and coefficient generation NNs.
[0034] The coefficient generation NNs have a two-stream NN structure, and each NN generates either the diffusion coefficient ρ or the convection coefficient β, which are stabilization coefficients. The generated diffusion coefficient ρ and convection coefficient β are passed to the stabilization cr-EPT, which calculates a conductivity map as output. The reconstructed (predicted) conductivity map is compared with the ground truth conductivity map, a loss function is calculated based on SSIM, and feedback is performed using the backpropagation method (see the dashed arrows in Figure 1), updating the parameters of the two individual NNs.
[0035] In detail, the coefficient generation NNs in this embodiment learns to generate optimized stabilization coefficients (diffusion coefficient ρ and convection coefficient β) from the difference (reconstruction error) between the input and the conductivity map without explicitly specifying the ground truth of the diffusion coefficient ρ and convection coefficient β, which are unknown in the MREPT research area. Note that the U-net that constitutes the coefficient generation NNs is a general model for semantic segmentation, so a detailed description will be omitted.
[0036] The output of the NN may be optimized by various settings, for example, bounded by a sigmoid activation function at the output so that the stabilization coefficient does not overwhelm the stabilized cr-EPT in equation (3) above. Specifically, the diffusion coefficient ρ is set to 1×10 because the diffusion term (viscosity term) may cause the values of neighboring pixels to be too strongly coupled, thereby attenuating all spatial variations. -6 The value is limited to 0.1. The maximum value of the diffusion coefficient ρ is selected by grid search to select a global value that does not cause numerical artifacts in the samples, and then the global value is doubled to increase the degrees of freedom of the neural network. The minimum value of the diffusion coefficient ρ is a non-zero value to prevent the diffusion coefficient ρ from being deleted from the following equation (4). On the other hand, the value of the convection coefficient β is limited to 0.01 to 5 to prevent the convection term from overtaking the reconstruction during training and obscuring the formulation. The maximum value of the convection coefficient β is limited to twice the global convection coefficient value to increase the degrees of freedom of the neural network, and the minimum value of the convection coefficient β is selected to prevent the neural network from getting stuck in a local minimum by deleting the convection term from the following equation (4). Note that the following equation (4) shows the optimization formulation for the i-th iteration.
[0037]
[0038] where ρ θi and β θi is the NN parameter θ belonging to the parameter space (θ) i are the diffusion coefficient ρ and the convection coefficient β at the iteration when σ(ρ) is given. L is the loss function defined by (5) using the SSIM of the following equation (6). θi-1 , β θi-1 ) is the conductivity predicted in the previous iteration. σ is the conductivity of the ground truth. In the above equation (4), ˆσ(ρ θi-1 , β θi-1 ) is compared with σ through the loss function L. The NN parameters θ i is optimized in the first learning iteration to minimize the loss function L according to the gradient using the RMSprop optimization algorithm with a learning rate of 0.001 and momentum of 0.9.
[0039]
[0040]
[0041] where σ and σ are the ground truth conductivity and the reconstructed conductivity, respectively. μ is the mean value of a collective patch of pixels. SD is the standard deviation of the same patch. 1 and C 2 is a prefix coefficient. SSIM is chosen from various options for constructing a loss function to overcome highly local artifacts that appear in the conductivity reconstructed by Equation (3) above. This is because the calculation of μ and SD in Equation (6) above is performed over regions, i.e., patches of pixels, thereby mitigating the impact of local artifacts. If a pixel-wise error measure such as mean squared error (MSE) were used, these local artifacts would bias the training, leading to high diffusion values (ρ ≈ 0.1), reducing the overall spatial variance and reconstructing a conductivity map with very low contrast.
[0042] For each model of magnetic resonance electrical properties tomography, three generalization experiments were conducted using two simulated data sets, in which the test samples gradually deviated from the training samples, and a noise robustness experiment.
[0043] In each experiment, we compared the physically coupled learning model PCNN-EPT with PCNN-local-EPT (see Figure 1), in which stabilization coefficients are generated as local coefficient maps, which are matrices applied locally to each pixel in the ROI, and PCNN-global-EPT, in which stabilization coefficients are scalars applied globally to the ROI. In PCNN-global-EPT, two parallel two-layer fully connected architectures are used to generate the global coefficient maps from the coefficient-generating neural networks. Specifically, in the first layer, the input is compressed to the square root of the phase image size, and in the second layer, it is further compressed to a scalar output.
[0044] In addition to the two PCNN-EPTs, we also compared the non-learning numerical analysis models std-EPT and cr-EPT, and the physics-uncoupled direct learning model NN-EPT. The loss function of the NN in NN-EPT is based on the mean squared error (MSE). This is because NN-EPT can benefit more from MSE than SSIM due to the absence of local artifacts in the end-to-end approach (see Figure 13).
[0045] The stabilized cr-EPT used in PCCNN-EPT is based on the formulation of the above equation (3), and receives stabilized coefficients (local coefficient map or global coefficient map) output from the coefficient generation NNs as input, generating a conductivity map as output. In this embodiment, PCCNN-EPT focuses on two-dimensional reconstruction to reduce computational costs, but it can also be extended to three-dimensional reconstruction. The spatial resolution is 2 mm in the x, y, and z directions. A central difference is taken for each point in the ROI, and a mesh is created using the finite difference method to solve the formulation of the above equation (3). More specifically, the formulation of the phase-stabilized cr-EPT after discretization is as shown in the following equation (7). Here, i and j are pixel index numbers in the x and y directions, respectively, and φ tr The derivatives are calculated using a second-order Savistky-Golay filter. Discretization of equation (7) below gives a set of linear equations for the ROI, σ i,j We can form an Ax = b system where x is an unknown sequence.
[0046]
[0047] Here, we describe the simulated dataset used for each experiment. EPT Algorithm / Network φ tr , ∇φ tr , ∇ 2 φ tr The input of B is a simulation of different phantoms. + 1 Simulation B + 1 The datasets were generated using Sim4Life (ZMT AG, Zurich). For the constructed phantom (see Figure 3), the transmit coil was a high-pass 16-run birdcage coil shielded by a 50 cm diameter shield, with an inner diameter of 28 cm for head imaging. The birdcage coil was excited with two ports geometrically separated by 90°, using a 127.78 MHz harmonic excitation, with a 90° phase shift between the ports to create an orthogonal mode. The complex B associated with a specific slice for reconstruction was calculated. + 1 The field is extracted and the transmit / receive phase φ trand the corresponding derivatives are used as inputs for the model. To generate the training samples required for the data-driven model, we mounted different phantoms (see Fig. 3(a) and (b)) on the coil and further performed complex B + 1 The field was chosen to reduce the end-ring effect and suppress the value of δBz, so that δBz≈0 can be assumed.
[0048] In this example, two types of training data sets of numerically simulated phantoms are created: a two-dimensional artificial irregular shape (AIG) data set shown in Figure 2(a) and a digital human head (DHH) data set generated from the virtual population 3.0 shown in Figure 2(b).
[0049] The AIG dataset consists of 30 samples, each containing a different foreground embedded in a cylindrical main structure measuring 16 cm in diameter and 24 cm in length. The length of the cylinder is chosen to minimize edge effects. The structures vary in size and conductivity (0.1-2.5 S / m). This AIG dataset was created to provide diverse information for the reconstruction model, presenting samples with different conductivity values as well as different main structures.
[0050] The DHH dataset consists of 30 samples, consisting of 15 samples with ±40% conductivity variation. The 15 samples correspond to eight male digital models ("Duke") and seven female models ("Ella"). From each model, two consecutive slices from the ROI were selected to reconstruct a total of 30 samples (#31-46 for "Duke" and #47-60 for "Ella"). Although the male and female models are at different heights, the ROIs are set at a constant height, resulting in different anatomical structures being focused on for the male and female models. This DHH dataset was created to represent samples with anatomical structures and complex boundaries similar to those found in clinical settings.
[0051] (Modified Test Dataset) The test dataset used in this example is a subset of the DHH dataset samples, with variations added, consisting of 16 samples with modifications such as rotation (up to 15° left or right) and the addition of masses simulating pathological tissue. The masses simulating pathological tissue were created by placing spheres with diameters of 0.8 to 2.4 cm at random locations within the ROI in the brain, increasing the white matter conductivity by 60 to 120% to mimic brain tumors. These modified samples were created to represent variations that may appear in clinical practice.
[0052] (Generalization Experiment) To investigate the generalizability of PCNN-EPT, we conducted Generalization Experiments 1 to 3. The test samples were designed with three levels of deviation from the training samples, with Generalization Experiment 1 showing the smallest deviation, followed by Generalization Experiment 2 and Generalization Experiment 3.
[0053] In generalization experiment 1, the AIG dataset (Fig. 2(a)) and the DHH dataset (Fig. 2(b)) are partitioned using a five-fold cross-validation method. This five-fold cross-validation employs a "leave-one-group-out" training and testing partition, i.e., one group for testing and four groups for training. Therefore, five training schemes (one for each test group) were performed for each dataset. The average reconstruction accuracy of the five test groups summarizes the model's performance on the dataset.
[0054] In generalization experiment 2, the model using the training group of the DHH dataset in Figure 2(b) from generalization experiment 1 was tested using 16 test samples obtained by modifying the DHH dataset, such as by rotating it or adding pathological tissue, as shown in Figure 2(c).
[0055] In Generalization Experiment 3, we investigate dataset generalization, i.e., responses in two cases that are challenging for data-driven PCNN-EPT and NN-EPT: when the presented test samples are completely different from the samples used during training. Specifically, in Generalization Experiment 3(A), we tested the training group from Generalization Experiment 1 and models trained on the DHH dataset (Figure 2(b)) with the corresponding test group from Generalization Experiment 1. In Generalization Experiment 3(B), we tested the models trained using the training group from the AIG dataset (Figure 2(b)) with modified test samples from the DHH dataset (Figure 2(c)).
[0056] (Noise Robustness Experiment) To investigate the noise robustness of PCNN-EPT, a noise robustness experiment was conducted. Using the model trained in the generalization experiment described above, noise was added to the corresponding test samples from each experiment to evaluate the noise robustness without further training. In detail, the transmit / receive phase (φ tr ) is contaminated with additive zero-mean Gaussian noise, and the added noise is + 1 B generated by the search method + 1 It is calculated by simulating the noise in the magnitude image. Therefore, the standard deviation (SD) of the noise, SDφ, is tr The relationship between the signal-to-noise ratio (SNR) of the grade image and the grade image is expressed as the following equation (8).
[0057]
[0058] The noise level is set within the range of a clinical MRI device, and the noisy transmission and reception phase (φ tr ) and its derivatives are used as the model input. Also, to avoid noise explosion, a low-pass Gaussian filter with a standard deviation of 1 is applied as preprocessing before numerical differentiation. This process is applied equally to all models being compared when noisy samples are tested.
[0059] (Consideration of local coefficient maps) Both the diffusion coefficient map and the convection coefficient map, which are stabilized coefficients generated in PCCNN-EPT, are very important, especially near tissue boundary regions. To analyze the variation of the coefficient maps generated in PCCNN-EPT, the coefficient maps are classified into boundary and non-boundary parts of the conductivity map. The conductivity boundary part is determined by calculating the sum of the gradient values of the conductivity in the x and y directions and adjusting the gradient value to the conductivity difference of the boundary of each test sample, i.e., the absolute value of the difference between the two conductivities that make up the boundary part.
[0060] The gradient values near the boundary are modeled as a Gaussian distribution, and the boundary region is defined as the region within the range of the mean value ±1 standard deviation to capture the fluctuations before and after the boundary of the coefficient map. The non-boundary conductivity region is determined by separating each test sample according to its conductivity value and deleting the previously defined boundary conductivity region. Using the boundary and non-boundary regions of the conductivity map, each coefficient map is divided into two: a boundary coefficient map and a non-boundary coefficient map.
[0061] (Phantom Scan) A Magphan® SMR-170 MRI phantom was used for the phantom scan. The background was filled with saline (2.5 g / L NaCl, estimated conductivity 0.5 S / m), and the foreground was filled with saline (10 g / L NaCl, estimated conductivity 2.0 S / m). Measurements were performed using a 3T "Skyra" MR scanner (Siemens, Erlangen, Germany) installed at Hubei Provincial Cancer Hospital, using a standard quadrature body coil. A multiple gradient echo (m-GRE) sequence was used to measure the transmit / receive phase. The imaging parameters were TR = 84 ms, first echo time (TE) = 1.91 ms, ΔTE = 2.67 ms, flip angle = 15°, and number of echoes = 6. The scan resolution was 1.06 × 1.06 × 2 mm. The unwrapped phase map at echo time is then fitted to the transmit and receive phase using temporal fitting to obtain the phase measurement at TE=0 as shown in equation (9) below, which is used to reconstruct the conductivity: where r is the phase, TE is the echo time, and B 0 is a constant uniform magnetic field, B 1 is B 0is the magnetic field that crosses the
[0062]
[0063] The phase map was also filtered for conductivity reconstruction as described in the noise robustness experiment above. Specifically, a 3 mm thick wall between the background and foreground nulled the signals of the pixels between the background and foreground, which caused artifacts in the reconstruction. Therefore, the filtering process smoothed out these nulled signals.
[0064] Fig. 4 shows the conductivity maps reconstructed by the non-learning numerical analysis models (std-EPT, cr-EPT), the physically uncoupled direct learning model (NN-EPT), and the physically coupled learning models (PCNN-global-EPT, PCNN-local-EPT) in generalization experiments 1 to 3, along with the SSIM values of the reconstructed conductivity maps. Fig. 5 also shows the local coefficient map generated by PCNN-local-EPT, as well as the normalized local coefficient map and line profile.
[0065] The first column in Figure 4 is the ground truth conductivity map. Columns 2 to 6 in Figure 4 are the conductivity maps reconstructed by std-EPT, cr-EPT, NN-EPT, PCNN-global-EPT, and PCNN-local-EPT, respectively, in generalization experiments 1 to 3. Columns 7 and 8 in Figure 5 (see the numbers in the figures) are the local diffusion coefficient map and local convection coefficient map used in PCNN-local-EPT in column 6 in Figure 4, where each pixel has a learned diffusion coefficient ρ or convection coefficient β. Columns 9 to 12 in Figure 5 are the normalized local diffusion coefficient map and local convection coefficient map and their respective line profiles.
[0066] Specifically, the local diffusion coefficient map and the local convection coefficient map shown in columns 7 and 8 of Fig. 5 are normalized to maximum values of 0.1 and 1, respectively, for better visibility, and are shown in columns 9 and 11 of Fig. 5. The line profiles of the local diffusion coefficient map and the local convection coefficient map shown in columns 10 and 12 of Fig. 5 are accompanied by line profiles of the normalized ground truth conductivity map (see dashed lines). The line profile of the local diffusion coefficient map is magnified by five times to visualize its spatial shift.
[0067] The first and second rows of Figures 4 and 5 show the conductivity maps reconstructed in Generalization Experiment 1 using the 5-fold cross-validation method. The first row shows the results when sample #26 from the AIG dataset shown in Figure 2(a) was used as the test sample, and the second row shows the results when sample #33 from the DHH dataset shown in Figure 2(b) was used as the test sample. The third row shows the conductivity map reconstructed in Generalization Experiment 2, showing the results when sample #75, in which the DHH dataset was modified and shown in Figure 2(c), was used as the test sample. The fourth and fifth rows show the conductivity maps reconstructed in Generalization Experiment 3, the most difficult experiment in which the test data and training data are completely different. The fourth row for Generalization Experiment 3(A) shows the results when sample #26 from the AIG dataset shown in Figure 2(a) was used as the test sample, and the fifth row for Generalization Experiment 3(B) shows the results when sample #75, in which the DHH dataset was modified and shown in Figure 2(c), was used as the test sample.
[0068] The sixth line in FIG. 4 shows the correlation coefficient (CC) between the line profile at the center of mass of the largest additional tissue added in the third line and the corresponding ground truth.
[0069] Figures 6(a) to (e) show the mean and standard deviation of the reconstruction accuracy for the generalization experiments shown in rows 1 to 5 of Figure 4, using different cases: case (1) std-EPT, case (2) cr-EPT, case (3) NN-EPT, case (4) PCNN-global-EPT, and case (5) PCNN-local-EPT. Furthermore, to examine the roles of the global and local coefficient maps, we also performed generalization experiments for cases (6) to (9). Cases (6) and (7) show the average and standard deviation of the reconstruction accuracy for PCNN-global-EPT and PCNN-local-EPT, respectively, when only the diffusion coefficient ρ is learned and the convection coefficient β is set to 1 (fixed global coefficient map). Case (8) shows the average and standard deviation of the reconstruction accuracy for PCNN-global-EPT and PCNN-local-EPT, respectively, when the global diffusion coefficient map and the local convection coefficient map are learned in combination. Case (9) shows the average and standard deviation of the reconstruction accuracy for PCNN-global-EPT and PCNN-local-EPT, respectively, when the local diffusion coefficient map and the global convection coefficient map are learned in combination.
[0070] The reconstruction results in the generalization experiment will be explained in more detail. In the following explanation, figure X(i, j) means to refer to row i, column j in figure X.
[0071] (Generalization Experiment 1) Figure 4 (1, 1-6) shows the results of reconstructing test sample #26 from the AIG dataset, and Figure 4 (2, 1-6) shows the results of reconstructing test sample #33 from the DHH dataset. It can be seen that the conductivity maps in columns 4-6 reconstructed using the learning method (learning model) have significantly reduced artifacts compared to the conductivity maps in columns 2-3 reconstructed using the non-learning method (analysis model). Furthermore, comparing columns 5 and 6, it can be seen that the PCNN-global-EPT (Figure 4 (1-2, 5)) has lower reconstruction accuracy than the PCNN-local-EPT (Figure 4 (1-2, 6)). Furthermore, comparing columns 4 and 6, for test sample #26 from the AIG dataset shown in the first row, PCNN-local-EPT (Fig. 4(1,6)) shows higher reconstruction accuracy than NN-EPT (Fig. 4(1,4)). On the other hand, for test sample #33 from the DHH dataset, which has less geometric variation, PCNN-local-EPT (Fig. 4(2,6)) shows lower reconstruction accuracy than NN-EPT (Fig. 4(2,4)). Note that looking at the line profiles of the local diffusive flow coefficient map and the local convective flow coefficient map in columns 10 and 12, respectively, it can be seen that there is variation corresponding to the boundaries.
[0072] Figures 6(a) and (b) show the SSIM values, i.e., the mean and standard deviation of the reconstruction accuracy, of generalization experiment 1 using the 5-fold cross-validation method for the AIG dataset and the DHH dataset, respectively. Comparing cases (1) to (5) in Figures 6(a) and (b), it can be seen that the learning method (cases (3) to (5)) has higher reconstruction accuracy than the non-learning method (cases (1) to (2)). Furthermore, as shown in Figures 6(a) and (b), when the sample is changed from the highly diverse AIG dataset in Figure 2(a) to the less diverse DHH dataset in Figure 2(b), it can be seen that the reconstruction accuracy of both PCNN-global-EPT (case (4)) and PCNN-local-EPT (case (5)) decreases.
[0073] Furthermore, in both Figures 6(a) and (b), it can be seen that in generalization experiment 1, the test samples do not deviate from the training samples, and therefore NN-EPT (case (3)) exhibits the highest reconstruction accuracy among the learning methods.
[0074] 6(a) and 6(b), comparing the results of cases (6) and (7) with cases (4) and (5) confirms that there is almost no degradation in reconstruction accuracy when only the diffusion coefficient ρ is learned as a global diffusion coefficient map or a local diffusion coefficient map and the convection coefficient β is 1 (fixed global coefficient map). Furthermore, comparing the results of case (4) in which both the diffusion coefficient ρ and the convection coefficient β are learned as global coefficient maps, cases (8) and (9) in which one of the coefficients is learned as a global coefficient map, and case (5) in which both coefficients are learned as local coefficient maps, it confirms that learning at least one of the diffusion coefficient ρ and the convection coefficient β as a local coefficient map leads to high reconstruction accuracy, and among these, case (5) in which both coefficients are learned as local coefficient maps has the highest reconstruction accuracy.
[0075] (Generalization Experiment 2) Test sample #75 used in generalization experiment 2 is a modified sample from the DHH dataset that adds rotation and spherical tissue simulating a brain tumor (boxed area in Figure 4 (3,1)). This sample is more difficult to reconstruct than generalization experiment 1. As shown in the line profiles in Figure 4 (6,2-6), all of the non-learning methods (Figure 4 (6,2-3)) exhibit wavy artifacts, while all of the learning methods (Figure 4 (6,4-6)) suppress the occurrence of wavy artifacts. Among the learning methods, the line profile of PCNN-global-EPT in Figure 4 (6,5) has a smaller waveform than the line profile of the ground truth conductivity map (see dashed line). As shown in Figure 4 (3,5), the reconstruction fails to adequately capture the contrast of cerebrospinal fluid (CSF) tissue with a conductivity of 2.14 S / m.
[0076] Furthermore, the NN-EPT line profile in Figure 4(6,4) shows an overestimation of contrast near the beginning of the line profile, and the added spherical tissue simulating the brain tumor exhibits an unstable line profile relative to the added mass, as indicated by the slope of the reconstructed conductivity. On the other hand, the PCNN-local-EPT results in Figures 4(3,6) and 4(6,6) show that the contrast of the CSF tissue is only slightly overestimated, and the contrast of the added spherical tissue simulating the brain tumor is appropriately reconstructed. Furthermore, the arrows in the conductivity map and local diffusion coefficient map profiles in Figures 4(6,6) and 4(3,10) confirm that the local diffusion coefficient map is reconstructed adaptively to the added spherical tissue simulating the brain tumor by increasing its value near the boundary of the added spherical tissue simulating the brain tumor and smoothing the transition to the surrounding white matter tissue.
[0077] As shown in Figure 6(c), cases (1) to (5) show similar trends to those in Figure 6(b). However, a major difference is that the reconstruction accuracy of NN-EPT (case (3)) and PCNN-local-EPT (case (5)) each decreases. The decrease in reconstruction accuracy of NN-EPT is particularly significant, resulting in the reconstruction accuracy of NN-EPT and PCNN-local-EPT being roughly consistent. Furthermore, when using a global diffusion coefficient map with PCNN-global-EPT, the impact of setting the convection coefficient β to 1 is small when comparing cases (4) and (6), and the difference in reconstruction accuracy is negligible. However, when using a local diffusion coefficient map, comparing cases (5) and (7), the reconstruction accuracy of case (7) decreases due to the convection coefficient β being set to 1. This confirms that a fixed value for either the diffusion coefficient ρ or the convection coefficient β is insufficient for the reconstruction of the conductivity map, and that it is important to learn both. Furthermore, comparing cases (8) and (9) with cases (4) and (5), it can be seen that it is particularly important to learn the diffusion coefficient ρ as a local diffusion coefficient map in order to improve the reconstruction accuracy.
[0078] (Generalization Experiment 3) Generalization Experiment 3 is the most difficult of the three generalization experiments because the test sample deviates most from the training sample, requiring higher generalization to reconstruct the conductivity map. Figure 4 (4,1-6), corresponding to Generalization Experiment 3(A), shows the results using the same test sample #26 used in the 5-fold cross-validation experiment shown in Figure 4 (1,1-6). By directly comparing the first and fourth rows of Figure 4, the impact of the similarity between the test sample and the training sample on the reconstruction can be confirmed. It can be seen that PCNN-local-EPT (Figure 4 (4,6)) achieves reconstruction with the same high accuracy as Figure 4 (1,6) in Generalization Experiment 1. On the other hand, it can be seen that the reconstruction using NN-EPT (Figure 4 (4,4)) exhibits significant unexpected artifacts that do not appear in Generalization Experiment 1 (Figure 4 (1,4)).
[0079] Note that comparing Fig. 4(4,3) and Fig. 4(4,5), the reconstruction of PCNN-global-EPT (Fig. 4(4,5)) shows similar artifacts as cr-EPT (Fig. 4(4,3)) due to the low value of the global diffusion coefficient map generated from the coefficient generation NNs (ρG ≈ 0).
[0080] Figure 4 (5,1-6), which corresponds to Generalization Experiment 3 (B), shows the results using the same test sample #75 as in Generalization Experiment 2 shown in Figure 4 (3,1-6). A direct comparison of the third and fifth rows of Figure 4 confirms that the SSIM values are nearly identical regardless of whether the non-learning method or the learning method is used. It can be seen that the reconstruction using PCNN-local-EPT (Figure 4 (5,6)) has a 0.08 reduction in SSIM value compared to Figure 4 (3,6). It can also be seen that the reconstruction using NN-EPT (Figure 4 (5,4)) has a 0.01 reduction in SSIM value compared to Figure 4 (3,4).
[0081] As shown in FIG. 6(d), in generalization test 3(A), among the learning methods (cases (3) to (5), the reconstruction accuracy by NN-EPT (case (3)) further decreased, becoming similar to that of PCNN-global-EPT (case (4)), and lower than that of PCNN-local-EPT (case (5)). Furthermore, when a global diffusion coefficient map is used in PCNN-global-EPT, the impact of setting the convection coefficient β to 1 is small even when comparing cases (4) and (6), and the reconstruction accuracy is similar. Furthermore, when a local diffusion coefficient map is used in PCNN-local-EPT, the impact of setting the convection coefficient β to 1 is small even when comparing cases (5) and (7), and the reconstruction accuracy is similar. Furthermore, when comparing cases (8) and (9) with cases (4) and (5), case (5), in which both the diffusion coefficient ρ and the convection coefficient β were learned as local coefficient maps, showed the highest SSIM value, while case (4), in which both were learned as global coefficient maps, showed the lowest SSIM value. This confirms the importance of coordinating both the diffusion coefficient ρ and the convection coefficient β as local coefficient maps in order to improve reconstruction accuracy.
[0082] Furthermore, as shown in Figure 6(e), in generalization experiment 3(B), among the learning methods (cases (3) to (5), the reconstruction accuracy of NN-EPT (case (3)) was reduced by 0.08 in SSIM value compared to the results in Figure 6(c). The reconstruction accuracy of PCNN-global-EPT and PCNN-local-EPT significantly decreased from cases (4) to (9), lagging behind that of NN-EPT (case (3)). However, it was confirmed that their reconstruction accuracy was higher than that of the non-learning methods (cases (1) and (2)). Furthermore, when comparing cases (6) and (7) with cases (4) and (5), the reconstruction accuracy was similar, but it was confirmed that case (7), in which the convection coefficient was set to 1 and the diffusion coefficient ρ was learned as a local diffusion coefficient map, achieved the highest reconstruction accuracy. Furthermore, comparing cases (4) and (5) with cases (8) and (9), it can be confirmed that when either the diffusion coefficient ρ or the convection coefficient β is trained as a global coefficient map, the reconstruction accuracy is comparable to that when a local diffusion coefficient map is trained, and the relative reconstruction accuracy is improved. In other words, PCNN-local-EPT is the only model that can maintain the contrast of the reconstructed conductivity map across generalization experiments 1 to 3.
[0083] (Noise Robustness Experiment) FIG. 8 shows the conductivity maps of test sample #10 reconstructed at four SNR levels (SNR = ∞, 150, 100, 50) from the AIG dataset of the above-mentioned generalization experiment 1 in the noise robustness experiment. The conductivity maps reconstructed by the non-learning numerical analysis models (std-EPT, cr-EPT), the physically uncoupled direct learning model (NN-EPT), and the physically coupled learning models (PCNN-global-EPT, PCNN-local-EPT) are shown together with the SSIM values of the reconstructed conductivity maps.
[0084] In addition, in the reconstruction by std-EPT in the second column, it can be seen that boundary artifacts are characteristic even without noise (SNR = ∞). In the reconstruction by cr-EPT in the third column, it can be seen that numerical artifacts are observed even without noise, and as noise increases, the reconstruction becomes filled with artifacts. In the reconstruction by NN-EPT in the fourth column, it can be seen that NN-EPT has high reconstruction accuracy when there is no noise, but as noise increases, the reconstruction shows random artifacts. In the reconstruction by PCNN-global-EPT in the fifth column, the reconstruction contrast is low, confirming the advantage of excellent noise robustness regardless of the SNR level. The reconstruction by PCNN-local-EPT in the sixth column shows high contrast without noise, and even when the SNR is increased to the standard noise level of an MRI scanner (SNR = 100), the SSIM value of the reconstructed conductivity map remains at 0.39, demonstrating the highest reconstruction accuracy among all models. It can be seen that the reconstruction by PCNN-global-EPT shows destructive artifacts when the noise level is increased to SNR = 50.
[0085] Figure 7 summarizes the mean and standard deviation of the SSIM conductivity maps for generalization experiments 1 to 3, ranging from SNR = 200 to 0. Figures 7(a) and (b) show the results of generalization experiment 1, Figure 7(c) shows the results of generalization experiment 2, and Figures 7(d) and (e) show the results of generalization experiments 3(A) and 3(B). As shown in Figure 7(a), when using an AIG dataset with diverse information, it can be confirmed that the training methods (NN-EPT, PCNN-global-EPT, and PCNN-local-EPT) are more noise-robust than the non-training methods (std-EPT and cr-EPT). As shown in Figure 7(b), when using a DHH dataset with less diversity, the training methods maintain their superior noise robustness compared to the non-training methods. However, when the SNR is higher than 70, the accuracy of PCNN-global-EPT is lower than the other two learning methods (NN-EPT, PCNN-local-EPT) in terms of reconstruction accuracy (SSIM value). However, as the SNR decreases, the decrease in reconstruction accuracy becomes very small, and therefore, when the SNR is 70 or lower, the reconstruction accuracy becomes higher than the other two learning methods.
[0086] Furthermore, comparing the results of Fig. 7(b) with Fig. 7(c) for generalization experiment 2, it can be seen that the reconstruction accuracy of both NN-EPT and PCNN-local-EPT decreases, but PCNN-local-EPT exhibits higher reconstruction accuracy than NN-EPT up to SNR = 50. Furthermore, as shown in Fig. 7(d) for generalization experiment 3(A), when the test sample deviates significantly from the training sample, PCNN-local-EPT maintains the same tendency as in generalization experiment 2, confirming its excellent noise robustness and generalizability to the diversity of test samples and training samples.
[0087] Furthermore, comparing Figure 7(d) for generalization experiment 3(A) with Figure 7(e) for generalization experiment 3(B), we find that the reconstruction accuracy of PCNN-local-EPT decreases by approximately 0.1 across the entire noise spectrum, while that of NN-EPT and PCNN-global-EPT increases by approximately 0.1, resulting in PCNN-local-EPT having lower reconstruction accuracy than NN-EPT and PCNN-global-EPT. This is primarily due to erroneous coefficients generated by PCNN-local-EPT, which are affected by noise, causing artifacts and reducing reconstruction accuracy. The artifacts caused by the erroneous coefficients can be confirmed by comparing the conductivity map reconstructed by PCNN-local-EPT (shown in the sixth column of Figure 8) with the conductivity maps reconstructed by NN-EPT and PCNN-global-EPT (shown in the fourth and fifth columns).
[0088] (Study of Diffusion Coefficient and Convection Coefficient) Figure 9 shows the conductivity map of sample #4 of the AIG dataset divided into boundary and non-boundary parts (first row), and the diffusion coefficient ρ (second row) and convection coefficient β (third row) generated by PCNN-local-EPT for three circular structures on the diagonal and their boundary and non-boundary parts. As shown in Figure 9, even for structures with the same conductivity, it can be seen that the generated coefficients are different depending on the surroundings, because different boundary effects must be treated.
[0089] Figure 10 compares the coefficients in the boundary and non-boundary regions of the test sample, with Figure 10(a) showing the diffusion coefficient and Figure 10(b) showing the distribution of the average coefficients for the conductivity values in the boundary and non-boundary regions for the convection coefficient. To compare the coefficient values in the boundary and non-boundary regions, two straight lines were fitted to the average diffusion coefficients for the conductivity values in the boundary and non-boundary regions using linear approximation.
[0090] In Fig. 10(a), both fitted lines show a decrease in the mean diffusivity with increasing conductivity, but the fitted line for the non-boundary region (dashed line) has a steeper slope. Also, in Fig. 10(b), the fitted line for the boundary region (solid line) is almost flat, but the fitted line for the non-boundary region (dashed line) shows a slight increase in the mean convection coefficient with increasing conductivity.
[0091] Figure 10(c) shows the aggregated values of the diffusion coefficients for the boundary and non-boundary areas, and it can be seen that there is a large difference between the coefficients for the boundary and non-boundary areas. Furthermore, as shown in Figure 10(d), the average diffusion coefficient value for the boundary area is lower than the average diffusion coefficient value for the non-boundary area, and a similar tendency can be seen for the convection coefficient.
[0092] (Phantom Scan Reconstruction) Figure 11 shows the results of phantom scan reconstruction. The first row shows the learning results for the DHH dataset, and the second row shows the learning results for the AIG dataset. In detail, the first column shows the reconstruction results using Ground Truth, the second column shows the reconstruction results using std-EPT with a global diffusion coefficient ρ = 0 and a global convection coefficient β = 0, the third column shows the reconstruction results using cr-EPT with a global diffusion coefficient ρ = 0 and a global convection coefficient β = 1, and the fourth column shows the reconstruction results using stabilized cr-EPT with a global diffusion coefficient ρ = 1 and a global convection coefficient β = 1. The fifth column shows the reconstruction results using NN-EPT, the sixth column shows the reconstruction results using PCNN-global-EPT, and the seventh column shows the reconstruction results using PCNN-local-EPT. The eighth and ninth columns show the local diffusion coefficient map and the local convection coefficient map, respectively, generated by PCNN-local-EPT in the seventh column.
[0093] In Figure 11, the std-EPT and cr-EPT reconstructions (columns 2 and 3) show typical artifacts in both rows 1 and 2, respectively. Even the stabilized cr-EPT (column 4) exhibits artifacts associated with conductivity transitions at the foreground-background partition at very high diffusion coefficient values (ρ = 1). This is presumably due to the adverse effects of the global diffusion coefficient. The NN-EPT reconstruction (column 5) exhibits poor reconstruction accuracy in both rows 1 and 2. The PCCNN-global-EPT (column 6) exhibits a central artifact in both rows 1 and 2, similar to the stabilized cr-EPT (column 4). However, the area of the artifact is reduced in the second row, where the training samples were generated using the AIG dataset. The PCCNN-local-EPT reconstruction (column 7) shows the best reconstruction results, but overestimates the foreground conductivity, indicating that this is the only way to accurately characterize the ground truth structure.
[0094] As shown in the results of the conductivity map reconstruction in columns 2 and 3 of Figure 4, PCCNN-EPT (PCCNN-global-EPT, PCCNN-local-EPT) includes formulations derived from std-EPT, which suffers from severe boundary artifacts due to its derivation assumptions, and cr-EPT, which lacks flexibility in its representation, such as fixed coefficients. However, considerable flexibility can be achieved by introducing two multidimensional variable coefficients, namely, the stabilization coefficients (diffusion coefficient ρ and convection coefficient β), which require data optimization. This is evident by comparing Figure 4 (5, 6) with Figure 4 (2, 3).
[0095] Furthermore, by combining the optimization of the two stabilization coefficients using coefficient-generating NNs with the analytical model (stabilized cr-EPT), we achieve the generalization of PCNN-EPT, which is difficult to achieve through end-to-end learning. This is an important point that is essential for future clinical applications of this method. As shown in the results of generalization experiment 2 (Fig. 4(3,1-6) and Fig. 4(5,1-6)), for an additional spherical brain tumor-like tissue not found in the training sample (boxed area in Fig. 4(3,1)), the analytical models (std-EPT, cr-EPT) were able to reconstruct the additional tissue approximately. However, significant errors were observed around the tissue, as indicated by the low correlation coefficient (Fig. 4(6,2-3)) of the line profile across one of the additional tissues (see Fig. 4(3,1-6)). It can be seen that these errors were effectively dealt with with the help of the local diffusion coefficient map and the local convection coefficient map optimized by the coefficient generation NNs of PCNN-local-EPT, as shown in the line profile (see Fig. 4(5,6)). Therefore, the combination of the formulation based on a physical model with wave physics and the learning of stabilization coefficients in PCNN-EPT plays the most important role for high generalization in the reconstruction of conductivity maps.
[0096] This is also evidenced by the results of generalization experiment 3(A), where the PCNN-local-EPT reconstruction showed the highest SSIM value (Fig. 4(4,1-6)), and generalization experiment 3(B), where the PCNN-local-EPT reconstruction showed the second highest SSIM value after the NN-EPT reconstruction (Fig. 4(5,1-6)). Furthermore, the reconstruction results of the phantom scan in Fig. 11 support this assertion. The reconstructions by std-EPT, cr-EPT, and stabilized cr-EPT in columns 2-4 of Fig. 11 show harmful artifacts. The end-to-end NN-EPT reconstruction in column 5 also shows very noticeable artifacts. Meanwhile, the PCNN-global-EPT reconstruction in column 6 shows significantly fewer artifacts than the stabilized cr-EPT reconstruction in column 4 by balancing the global diffusion coefficient map and the global convection coefficient map. Furthermore, it can be seen that the reconstruction by PCNN-local-EPT in the seventh column is able to remove the artifacts that first appeared in the reconstruction by stabilized cr-EPT in the fourth column by combining the local diffusion coefficient map and the local convection coefficient map, thereby showing improved reconstruction accuracy.
[0097] The line profile of the local diffusion coefficient map in Figure 4 (4,10) shows smaller values near the boundary compared to Figure 4 (1,10). However, when compared to Figure 4 (4,2-5), it can be confirmed that the line profile plays a role in reducing artifacts, as shown in Figure 4 (4,6). Furthermore, the line profile of the local diffusion coefficient map in Figure 4 (5,10) shows smaller values near the boundary compared to Figure 4 (3,10). This can be confirmed by the tendency for the peaks of the local diffusion coefficient map trained with the AIG dataset to be smaller, as shown in Figure 4 (1,10). Furthermore, in the function mapping of the physically uncoupled NN-EPT, reconstruction is determined solely by the training sample. Therefore, as shown in Case (3) of Figures 6 (a)-(e), reconstruction accuracy decreases as the test sample deviates from the training sample. PCNN-local-EPT also uses coefficient generation NNs, so similar effects appear. However, due to the formulation of the physical model, as shown in case (5) of Figures 6(a) to (d), as long as the coefficients do not produce extreme artifacts that cause the accuracy degradation shown in Figure 6(e), it is robust to changes in the test samples used, i.e., new samples.
[0098] The noise robustness of PCNN-local-EPT is comparable to that of NN-EPT, as shown in Fig. 7(a), (b) and Fig. 8(3, 6) for generalization experiment 1. Moreover, the noise robustness of PCNN-local-EPT is higher than that of NN-EPT until the noise level increases to SNR = 50 in the test samples, as shown in Fig. 7(c) for generalization experiment 2 and Fig. 7(d) for generalization experiment 3(A).
[0099] Here, as shown in the fourth column of Figure 8, NN-EPT generates more spurious noise-related artifacts as the noise level increases. This is presumably because NN-EPT maps the input to the reconstructed image based on the pattern matching rules acquired by the NN, and the noise is directly transferred to the reconstructed image in an interpolated manner, further reducing generalizability. This problem is due to the B + 1The process of obtaining the signal can be exacerbated by several noise sources.
[0100] On the other hand, in PCNN-local-EPT, even if the test sample contains noise, the noise affects both the stabilization coefficient and the analytical model (stabilized cr-EPT) terms, but can ultimately be suppressed by the stabilization coefficient. Therefore, in generalization experiments 2 and 3(A), even if the test sample contained a structure that was partially or completely different from the training sample, and thus the local diffusion coefficient map and local convection coefficient map could not be accurately generated, PCNN-local-EPT was presumably able to perform relatively accurate reconstruction. However, the inaccurate local diffusion coefficient map and local convection coefficient map generated by PCNN-local-EPT not only generate destructive artifacts when the noise level increases to SNR = 50, as shown in the fourth column of Figure 8, but also, as shown in Figure 7(e) compared to Figure 7(d), the reconstruction accuracy of PCNN-local-EPT is generally lower than that of PCNN-global-EPT, which may adversely affect the reconstruction accuracy of some samples.
[0101] In this way, PCCNN-EPT not only inherits the noise robustness of the formulated stabilized cr-EPT realized by the global coefficient map, but also significantly improves the reconstruction accuracy by optimizing the local coefficient map. Furthermore, by allowing the diffusion coefficient ρ and convection coefficient β generated by the coefficient generation NNs in PCCNN-EPT to be selected between the global coefficient map and the local coefficient map depending on the noise conditions of the test sample, noise robustness and generalization can be improved.
[0102] The role of the diffusion coefficient ρ and the convection coefficient β in PCNN-EPT will be explained in detail. In PCNN-EPT, the diffusion coefficient ρ and the convection coefficient β are respectively the diffusion term (∇ 2 γ) and part of the convection term (∇φ tr , ∇γ).
[0103] The diffusion coefficient ρ determines how a pixel influences its neighbors. Applying a pixel-wise diffusion coefficient ρ, i.e., a local diffusion coefficient map, allows for the viscous effect to be tailored to the circumstances of each region for better reconstruction of the conductivity map. Specifically, at the boundary, ρ is set to a low value to generate accurate conductivity transitions, followed by a high value to avoid boundary artifacts. Meanwhile, at the non-boundary, high values are set to mitigate small numerical artifacts. This is well reflected in the aggregated diffusion coefficient values for the boundary and non-boundary regions in Figure 10(c) and the line profiles of the local diffusion coefficient maps learned from the two training samples (AIG dataset and DHH dataset) in Generalization Experiment 1 (see Figures 5(1,10) and 5(2,10)).
[0104] By simultaneously training the diffusion coefficient ρ and the convection coefficient β as local coefficient maps, an effective mutual interaction between the diffusion coefficient ρ and the convection coefficient β can be achieved. This can be confirmed by comparing the trained local diffusion coefficient maps (see (1,10) and (2,10) in Figure 5 ) with the line profiles of the trained local convection coefficient maps (see (1,12) and (2,12) in Figure 5 ). This can be seen by comparing the different cases in Figures 6(a) and (b) of generalization experiment 1. Case (5), in which both coefficients were trained as local coefficient maps, exhibits higher reconstruction accuracy than cases (4), (7), and (9), in which only one of the coefficients was set or trained as a global coefficient map. This is also true in generalization experiment 2 (see Figure 6(c)), in which pathological tissue not seen in the training samples was added to the test sample. However, as shown in Figures 6(d) and 6(e), in generalization experiments 3(A) and (B) where the test samples deviate significantly from the training samples, this advantage does not exist, and the accuracy difference between cases (5), (7), and (9) is smaller. This is because the convection coefficient β is fixed to a global coefficient map (β = 1), which reduces the possibility of generating erroneous coefficients that interfere with reconstruction, especially for test samples that are significantly different from the training samples.
[0105] As described above, these embodiments and examples provide a PCNN-EPT (physically coupled neural network electrical properties tomography) that effectively combines a stabilization cr-EPT and a coefficient generation NNs (U-net) and flexibly optimizes two stabilization coefficients (diffusion coefficient ρ and convection coefficient β) to generate accurate reconstructions. This enables the provision of a magnetic resonance electrical properties tomography that can achieve high reconstruction accuracy for two simulation datasets (AIG dataset and DHH dataset) even in cases where an end-to-end approach is difficult, such as when the test sample is most different from the training sample or in a noisy environment, i.e., has excellent noise robustness and generalizability.
[0106] In addition, PCNN-local-EPT, which takes into account different needs for tissue boundaries and optimizes two stabilization coefficients together, particularly as a local coefficient map, can provide magnetic resonance electrical property tomography with excellent noise robustness and generalizability.
[0107] In the above-described embodiments and examples, the manner in which a conductivity map is reconstructed by magnetic resonance electrical properties tomography has been mainly described, but the image reconstructed by magnetic resonance electrical properties tomography may be an image that uses EPs as tissue contrast, and may be, for example, a map based on a dielectric constant map.
[0108] Furthermore, although a neural network is used as a data-driven machine learning model in the above description, the present invention is not limited to this, and other techniques such as ensemble learning may be used as the machine learning model.
[0109] Furthermore, as shown in Figure 12, the results of the average and standard deviation of the SSIM values of the conductivity maps reconstructed under various outer boundary setting (OBS) conditions in generalization experiments 1 to 3 show that the reconstruction accuracy of each model is slightly improved if the OBS is accurate, but the reconstruction is accurate even without the OBS.
[0110] We also used the machine learning library PyTorch® to generate the backpropagation algorithm. This NN framework creates a computational graph that organizes operations into a hierarchical chart, enabling automatic differentiation. The NN is trained for up to 1,000 epochs, and the model with the best accuracy for the training sample is saved for testing. Training uses an annealing method, and if the loss does not decrease after 50 epochs, the training speed is halved. Furthermore, to accelerate training, PCNN-EPT also halves the learning rate if artifacts (spikes in the loss function) accumulate during reconstruction and cause a sudden increase in loss. This further prevents the model from getting stuck in a local optimum. Because the number of training samples is small, training is performed without creating a validation set for each training group. While PCNN-local-EPT converges, artifacts appear in the reconstruction when the local coefficient map is underestimated or overestimated, resulting in spikes in the loss function, as shown in Figure 14(a). These are usually easily corrected, but the loss continues to decrease. PCNN-global-EPT training decreases slightly because the reconstruction is affected only by global parameters, as shown in Figure 14(b). That is, the NN coefficient generation is a highly constrained modification to the reconstruction. NN-EPT training is very smooth, with the loss function decreasing at every epoch, as shown in Figure 14(c). For PCNN-EPT, there is a point during training where the learning stops decreasing, a spike appears, and the training stops at a local minimum (see the dashed line in Figure 14(d)). Subsequently, lowering the learning rate further drives the NN to converge, eventually requiring more than the limit of 1000 epochs to converge, as shown by the solid line in Figure 14(d). In such cases, the model with the lowest loss is retained for testing, provided it is within the limit of 1000 epochs. The training time ranged from 30 minutes for NN-EPT to 1 hour for PCNN-EPT, in either global or local format, and all training was performed on an Nvidia GeForce RTX 2070 GPU.
[0111] To create an independent comparison point for the end-to-end NN-EPT, we also compared two different inputs, one in which the NN directly outputs the final conductivity map and does not perform any intermediate steps. One model uses the transmit / receive phase φ as input to avoid the noise explosion associated with the numerical differentiation calculation. tr received (NN-EPTφ tr ). Another model is the transmission and reception phase φ tr In addition to the transmit and receive phase differential ∇ 2 φ tr , ∇φ tr These features provide relevant information based on many analytical methods, so they are concatenated into a multi-channel input (NN-EPT∇ 2 φ tr , ∇φ tr , φ tr ). Furthermore, since NN-EPT can be trained with different loss functions, we compared training with the normal loss function MSE of the end-to-end method and SSIM used in the embodiment and examples. As shown in Figure 13, there were slight differences in NN-EPT in each experiment, but the NN-EPT∇ trained with MSE in case (3) was 2 φ tr , ∇φ tr , φ tr It can be confirmed that the reconstruction accuracy improves across generalization experiments 1 to 3. [Industrial Applicability]
[0112] The present invention provides magnetic resonance electrical properties tomography with excellent noise robustness and generalizability by using physically coupled neural network electrical properties tomography (PCNN-EPT), which effectively combines stabilized cr-EPT and coefficient generation NNs, flexibly optimizing two stabilization coefficients (diffusion coefficient ρ and convection coefficient β) to generate accurate reconstructions.Since this invention can achieve higher reconstruction accuracy than conventional end-to-end NN-EPT, it has industrial applicability as an important step toward the practical application of magnetic resonance electrical properties tomography in medical diagnosis.
Claims
1. A magnetic resonance electrical property tomography comprising: a first means for reconstructing a conductivity map using a physical model based on wave physics with a diffusion coefficient ρ and a convection coefficient β as coefficients; and a second means for learning the diffusion coefficient ρ and the convection coefficient β using a data-driven machine learning model that utilizes parameters of the physical model.
2. The physical model is tr ・∇γ)+γ∇ 2 φ tr -ρ∇ 2 γ=ωμ 0 and the parameter is φ tr , ∇φ tr , ∇ 2 φ tr (where φ tr is the normalized transmit phase. γ is the inverse of the conductivity σ (γ=1 / σ). ω is the Larmor frequency. μ 0 The magnetic resonance electrical property tomography method according to claim 1, wherein:
3. The magnetic resonance electrical property tomography apparatus according to claim 1, wherein the machine learning model may be a learning model using a neural network.
4. The magnetic resonance electrical characteristic tomography according to claim 3, wherein the diffusion coefficient ρ and the convection coefficient β may be determined by different neural networks.
5. A magnetic resonance electrical characteristic tomography apparatus as described in any one of claims 1 to 4, which may further comprise a third means for calculating a difference between the conductivity map reconstructed by said first means and a ground truth conductivity map, and feeding the difference back to said second means.
6. The magnetic resonance electrical property tomography method according to claim 1, wherein the diffusion coefficient ρ and the convection coefficient β may each be a global coefficient map.
7. The magnetic resonance electrical property tomography of claim 1, wherein the diffusion coefficient ρ and the convection coefficient β may each be a local coefficient map.
8. The magnetic resonance electrical characteristic tomography of claim 1, wherein the diffusion coefficient ρ and the convection coefficient β can be selected as either a global coefficient map or a local coefficient map according to the noise conditions of a test sample.
Citation Information
Patent Citations
Information processing device and information processing method
JP2022115092A
Device and method for reconstructing low-frequency conductivity images using MRI without current injection
US20190142296A1
