A magnetic resonance image reconstruction method and model training method

Abstract

This invention provides a magnetic resonance image reconstruction method and a model training method. The model training method includes: acquiring a thermal ablation magnetic resonance image dataset; generating training samples for continuous-time magnetic resonance images at each fracture site through time downsampling, and adding them to a training sample set; and training a deep learning model using the training sample set to obtain a magnetic resonance image reconstruction model. This invention acquires continuous-time magnetic resonance images at several fracture sites during the thermal ablation process, and generates training samples for each fracture site's magnetic resonance images through time downsampling to train a deep learning model. The model can complete the downsampled time series, reducing the amount of magnetic resonance data collected during thermal ablation and improving the real-time performance of data acquisition. Furthermore, since the dataset covers the entire process and all locations of thermal ablation, the trained model better meets the needs of the thermal ablation scenario, and the accuracy of the reconstructed data is also higher.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a magnetic resonance image reconstruction method and a model training method. Background Technology

[0002] Thermal ablation is a technique that uses thermal effects to cause coagulation, necrosis, vaporization, or carbonization of diseased tissue, thereby achieving the purpose of ablation and inactivation. It includes laser, high-frequency electrosurgical, plasma coagulation, microwave therapy, and radiofrequency therapy. Magnetic resonance imaging provides information support for the thermal ablation process. Through magnetic resonance imaging, temperature changes in various parts of the tissue can be continuously monitored, which makes it easier for doctors to understand the current ablation status of the target area and provides doctors with more information support.

[0003] The drawback of magnetic resonance imaging is that it takes a long time, while the temperature field changes rapidly during thermal ablation, which requires high real-time temperature imaging. Currently, several improvement measures have been proposed to enhance the real-time performance of temperature imaging: ① Reducing the imaging range and acquiring fewer layers (e.g., cyclically acquiring images from three parallel slices passing through the target area). The drawback of this approach is that it has limited improvement in imaging efficiency, and the lesion itself occupies a certain space, resulting in insufficient coverage of the lesion and creating blind spots in temperature monitoring; ② Designing magnetic resonance sequences with shorter acquisition times, such as EPI sequences. The drawback of this approach is that it significantly reduces imaging quality, and the acquired temperature data has large errors, easily leading to under-ablation / false ablation; ③ Using downsampling to reduce the amount of data acquired, improving acquisition efficiency, and using algorithms to supplement missing data (e.g., interpolation algorithms). The drawback of this approach is poor image reconstruction quality. In addition, there is the practice of using deep learning models to supplement missing spatial data. The reasoning ability of this deep learning model is directly affected by the quality of the training data. Since the state of the tissue changes dynamically during thermal ablation, magnetic resonance imaging cyclically acquires images from various slices, but at any given moment, only one slice image can be acquired, making it difficult to obtain comprehensive and high-quality magnetic resonance imaging data for training, thus resulting in poor reasoning ability of the deep learning model.

[0004] Overall, while existing technologies have improved image acquisition efficiency, they have sacrificed image quality and have significant temperature monitoring errors. To overcome or at least partially overcome these defects, this invention provides a magnetic resonance image reconstruction method and a model training method. Summary of the Invention

[0005] This invention provides a magnetic resonance image reconstruction method and a model training method to solve the defects of poor real-time acquisition and low image quality in existing magnetic resonance images.

[0006] In a first aspect, the present invention provides a training method for a magnetic resonance image reconstruction model, comprising:

[0007] Acquire a thermal ablation magnetic resonance image dataset; wherein the thermal ablation magnetic resonance image dataset includes continuous-time magnetic resonance images at at least one fracture site during the thermal ablation process;

[0008] For each fault, continuous-time magnetic resonance images are generated by time downsampling to create training samples, which are then added to the training sample set.

[0009] The deep learning model is trained using the training sample set to obtain the magnetic resonance image reconstruction model.

[0010] Optionally, the thermal ablation magnetic resonance image dataset includes continuous-time magnetic resonance images of at least one fracture site obtained through thermal ablation experiments.

[0011] Optionally, the thermal ablation magnetic resonance image dataset includes continuous-time magnetic resonance images of at least one tomographic region obtained through thermal ablation experiments, and magnetic resonance images of at least one tomographic region from historical clinical thermal ablation data. The step of training a deep learning model using the training sample set to obtain a magnetic resonance image reconstruction model includes:

[0012] The initial deep learning model is pre-trained using training samples corresponding to the thermal ablation experiment, and then a second training is performed using training samples corresponding to the historical clinical thermal ablation data to obtain the deep learning model.

[0013] Optionally, the magnetic resonance image is a magnetic resonance K-space image, amplitude map, or phase map, or a magnetic resonance temperature map, temperature difference map, or ablation map generated based on the phase map.

[0014] Optionally, for each fault location, the continuous-time magnetic resonance image is downsampled to generate training samples, which are then added to the training sample set as follows:

[0015] Time downsampling is performed on continuous-time magnetic resonance images acquired at a fault location;

[0016] Several adjacent frames of magnetic resonance images at the fault location after downsampling are used as samples, and the missing magnetic resonance images are used as corresponding labeled data to form training samples, which are then added to the training sample set.

[0017] Optionally, the training samples may also include tissue characteristic parameters and / or thermal ablation power data corresponding to the magnetic resonance image acquisition time. During model training, the tissue characteristic parameters and / or thermal ablation power data corresponding to the magnetic resonance image acquisition time are used as additional data of the magnetic resonance image to train the deep learning model.

[0018] Further, training samples are generated as follows:

[0019] The magnetic resonance image and ablation power acquired at a certain moment at a fault location are used as samples, and the missing temperature maps at adjacent moments are used as labeled data to form training samples, which are then added to the training sample set.

[0020] Alternatively, training samples can be generated as follows:

[0021] Time downsampling is performed on continuous-time magnetic resonance images acquired at a fault location;

[0022] All magnetic resonance images prior to a certain moment after downsampling at the fault location are used as samples, and the thermal ablation image at that moment is used as labeled data to form training samples, which are then added to the training sample set.

[0023] Optionally, the deep learning model is selected from any of the following: recurrent neural networks, generative adversarial networks, and long short-term memory networks.

[0024] Secondly, the present invention also provides a magnetic resonance image reconstruction method, comprising:

[0025] Acquire time-downsampled magnetic resonance images of the current patient;

[0026] The time-downsampled magnetic resonance image is input into the magnetic resonance image reconstruction model to obtain the time-reconstructed magnetic resonance image, which provides information support for the thermal ablation process;

[0027] The magnetic resonance image reconstruction model is pre-trained according to the training method of the magnetic resonance image reconstruction model described in any of the preceding claims.

[0028] Optionally, the magnetic resonance image reconstruction method further includes inputting the time-reconstructed magnetic resonance image into a spatial reconstruction model to obtain a further spatially reconstructed magnetic resonance image; wherein the spatial reconstruction model is trained using data collected from reproducible thermal ablation experiments.

[0029] Optionally, the magnetic resonance image reconstruction method also includes generating temperature maps and / or ablation maps based on the magnetic resonance images acquired at each time point.

[0030] Thirdly, the present invention also provides a magnetic resonance ablation monitoring method, comprising:

[0031] Acquire time-downsampled magnetic resonance images of the current patient;

[0032] The time-downsampled magnetic resonance image is input into the magnetic resonance image reconstruction model to obtain the reconstructed ablation map, which provides information support for the thermal ablation process;

[0033] The magnetic resonance image reconstruction model is pre-trained according to the training method of the magnetic resonance image reconstruction model described in any of the preceding claims.

[0034] Fourthly, the present invention also provides a magnetic resonance-guided thermal ablation monitoring system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the training method for the magnetic resonance image reconstruction model as described in any of the preceding claims, or implements the steps of the magnetic resonance image reconstruction method as described in any of the preceding claims, or implements the steps of the magnetic resonance ablation monitoring method as described in the preceding claims.

[0035] The magnetic resonance image reconstruction method and model training method provided by this invention have at least the following beneficial effects:

[0036] 1. By acquiring continuous-time magnetic resonance images at at least one fracture point during the thermal ablation process, training samples are generated for each fracture point's magnetic resonance image through time downsampling to train a deep learning model. The model can fill in the missing time series, resulting in better temporal continuity of the output image. This allows for a lower rate of reduction in the amount of magnetic resonance image data acquired during the thermal ablation process, thereby improving the real-time performance of data acquisition.

[0037] 2. By acquiring magnetic resonance image data through reproducible thermal ablation simulation experiments, the magnetic resonance equipment has sufficient time to acquire magnetic resonance images of the "same thermal ablation process" in multiple experiments, which effectively makes up for the problem of insufficient and difficult acquisition of real thermal ablation data, especially solving the problem of limited and difficult acquisition of continuous time magnetic resonance image data at the same fault.

[0038] 3. The model was pre-trained using thermal ablation simulation test data, which gave the model the foundation for transfer learning. Secondary training using historical clinical thermal ablation data enabled the model to better supplement the missing data in thermal ablation scenarios, thereby improving the model's performance.

[0039] 4. Use thermal ablation power data and / or target tissue characteristic parameters as supplementary data for training samples to improve the accuracy of model predictions.

[0040] 5. Some implementations support the reconstruction of magnetic resonance K-space images, some support the reconstruction of magnetic resonance amplitude maps, some support the reconstruction of magnetic resonance phase maps, and some support the reconstruction of phase difference maps, temperature difference maps, temperature maps, or ablation maps generated based on phase maps, thus better meeting the monitoring needs of the thermal ablation process. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0042] Figure 1 This is a flowchart illustrating the training method of the magnetic resonance image reconstruction model provided by the present invention;

[0043] Figure 2 This is an example of a magnetic resonance imaging acquisition method;

[0044] Figure 3 This is a schematic flowchart of the magnetic resonance image reconstruction method provided by the present invention;

[0045] Figure 4 This is a flowchart illustrating the magnetic resonance ablation monitoring method provided by the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0047] The following is combined with Figures 1-3 This invention describes a magnetic resonance image reconstruction method and a model training method. Figure 1 This is a flowchart illustrating a training method for a magnetic resonance image reconstruction model provided by the present invention, as shown below. Figure 1 As shown, the method includes:

[0048] S11. Obtain the thermal ablation magnetic resonance image dataset; wherein, the thermal ablation magnetic resonance image dataset includes continuous-time magnetic resonance images at at least one fracture site during the thermal ablation process;

[0049] Specifically, this thermal ablation magnetic resonance imaging (MRI) dataset comprises MRI images acquired throughout the entire thermal ablation process. The acquired images are continuous-time MRI images at several tomographic locations, such as MRI images at one, two, three, etc., passing through the target region. The dataset can include data from different patients and different tomographic locations, enriching the dataset and improving the model's image reconstruction performance. The continuous-time MRI images in the dataset can be K-space MRI images, amplitude maps, phase maps, or phase difference maps, temperature difference maps, temperature maps, or ablation maps generated based on the phase map—images related to MRI during the thermal ablation process. Furthermore, the dataset may also include energy power and cooling rates at various time points.

[0050] It should be noted that, in this invention, continuous time refers to the acquisition time interval at the same fracture site not exceeding 1 second. Furthermore, in the field of magnetic resonance imaging (MRI) acquisition, for static tissue structures, continuous-time MRI images of the same fracture site are typically not acquired. For thermal ablation processes, MRI images are usually acquired alternately at several fracture sites to maximize the monitoring range, rather than acquiring continuous-time MRI images from a single fracture site. This invention trains a deep learning model by acquiring continuous-time images at fracture sites to reconstruct time series, thereby improving the temporal continuity of the monitoring data stream.

[0051] S12. For the continuous-time magnetic resonance images at each fault, generate training samples by time downsampling and add them to the training sample set;

[0052] Specifically, for each continuous-time magnetic resonance imaging (MRI) image at a specific topograph, one or more training samples can be generated through temporal downsampling. For example, if MRI images A1, A2, A3, A4, A5, A6, A7, and A8 are continuously acquired at a topograph, a training sample can be generated using A1 and A3 as samples and A2 as the corresponding annotation, and so on, using A2 and A4 as samples and A3 as the corresponding annotation. Similarly, if MRI images B1, B2, B3, B4, B5, B6, B7, and B8 are continuously acquired at a topograph, a training sample can be generated using B1 and B3 as samples and B4 as the corresponding annotation, and so on, using B2 and B4 as samples and B5 as the corresponding annotation. By generating training samples from different patients and at different topographs, the diversity of training samples is increased. Consequently, the generalization ability of the subsequently trained model can be improved, enabling it to better reconstruct the temporally downsampled MRI images during actual thermal ablation and fill in the missing time series.

[0053] S13. Train the deep learning model using the training sample set to obtain the magnetic resonance image reconstruction model.

[0054] Specifically, training a deep learning model using a training sample set can be divided into a training set, a validation set, and a test set. The deep learning model learns patterns and rules from the training samples in the training set, enabling it to make predictions on unseen data. Since the training samples cover the entire thermal ablation process, the trained model can reconstruct downsampled data from any stage of real thermal ablation. The validation set is used to adjust the hyperparameters of the deep learning model, improving its performance and generalization ability. The test set is used to evaluate the model and select the best-performing model, which can then be used for magnetic resonance image reconstruction during subsequent thermal ablation processes.

[0055] In this embodiment, continuous-time magnetic resonance images are acquired at at least one fracture site during the thermal ablation process. Training samples are generated from the magnetic resonance images at each fracture site through temporal downsampling. These samples are then used to train a deep learning model. The model can complete the downsampled time series, resulting in better temporal continuity of the output images. This reduces the amount of data collected during the ablation process and improves the real-time performance of data acquisition. Furthermore, since the training samples cover the entire thermal ablation process and involve multiple fracture locations, the trained model can accurately reconstruct the time series of magnetic resonance images at all locations throughout the entire thermal ablation process.

[0056] Reference Figure 2 To understand the effects of this invention, existing technology involves cyclically acquiring magnetic resonance images at slices 1, 3, and 5: that is, slice 1 is acquired in the 1st second, slice 3 in the 2nd second, and slice 5 in the 3rd second; slice 1 is acquired in the 4th second, slice 3 in the 5th second, slice 5 in the 6th second, and so on. Figure 2 As shown, the magnetic resonance imaging (MRI) reconstruction model of this invention can acquire odd-numbered layers 1, 3, and 5 in the first cycle, even-numbered layers 2, 4, and 6 in the second cycle, odd-numbered layers 1, 3, and 5 in the third cycle, and even-numbered layers 2, 4, and 6 in the fourth cycle, and so on, cyclically acquiring data. Each acquired data set can be used to supplement missing time-series data using the MRI reconstruction model of this invention. For example, inputting the data of layer 1 acquired at 1s and 7s into the MRI reconstruction model yields the data of layer 1 at 4s; inputting the data of layer 3 acquired at 2s and 8s yields the data of layer 3 at 5s; inputting the data of layer 5 acquired at 3s and 9s yields the data of layer 5 at 6s, thus including MRI image data with continuous tomography in the second cycle, and so on, to supplement data for each cycle.

[0057] It is evident that the method of this invention can also achieve the effect of "expanding the collection range without increasing the amount of data collected".

[0058] Based on the previous embodiment, in one embodiment, the thermal ablation magnetic resonance image dataset includes continuous-time magnetic resonance images of at least one fracture site obtained through thermal ablation experiments.

[0059] Specifically, thermal ablation experiments include ablation tests using animals or ablation models. For example, animal ablation experiments involve performing thermal ablation on live animals or excised animal tissues, acquiring continuous-time magnetic resonance imaging data at certain tomographic locations during the ablation process. Preferably, thermal ablation experiments also include reproducible ablation simulation experiments. These simulations mimic the external environment, biological tissue, and ablation procedures during the ablation process. Reproducible thermal ablation simulation experiments refer to reproducing the external environment, biological tissue model, and ablation procedures during the same ablation process. Traditional animal ablation experiments only perform a single ablation to verify the ablation effect and do not reproduce the ablation process. Moreover, traditional animal ablation experiments, using live animals or excised tissues as test subjects, have poor reproducibility due to potential individual differences among animals. Even by selecting animal breeds and ages, individual differences can only be controlled as much as possible. This method specifically designs a thermal ablation simulation experiment with "a specific external environment, a biological tissue model to be ablated, and an ablation operation." This thermal ablation simulation experiment has good reproducibility, and can reproduce the same thermal ablation process with the same external environment, biological tissue model to be ablated, and ablation operation. By reproducing the thermal ablation simulation experiment multiple times, we have enough time to acquire magnetic resonance images. Furthermore, by selecting the image acquisition location and time point, the images acquired in multiple experiments have a correspondence, which can be used to generate training samples. By changing the external environment, the tissue model to be ablated, and the ablation operation, multiple sets of thermal ablation simulation experiments can be formed, and more training samples can be collected, which makes up for the difficulty in obtaining clinical data and generating sufficient training samples.

[0060] The aforementioned external environment includes, for example, ambient temperature and humidity. For instance, referencing the standard temperature of an MRI room (18°C–24°C) and standard humidity of 40%–60%, the same ambient temperature (22°C) and humidity (50%) are used in the thermal ablation simulation experiment. The biological tissue model is a structural model of the object to be ablated. For example, in laser ablation scenarios, agar blocks are used to simulate brain tissue. The refractive index of the agar is adjusted by adding auxiliary materials such as sugar and acid. Laser light is guided into the agar block using an optical fiber, and the thermal effect of the laser ablates specific areas of the agar block. Another example is using gelatin and agar as the main materials, with the addition of materials such as sodium chloride to adjust conductivity to simulate brain tissue. Silicone tubes simulate blood vessels, and methacrylamide gelatin simulates tumors. Radiofrequency current is applied using a radiofrequency ablation needle to heat specific areas of this "brain model." Yet another example is using agar blocks to simulate brain tissue in ultrasound ablation scenarios. The absorption and reflection rates of ultrasound are adjusted by changing the porosity of the agar blocks. The same thermal ablation operation is applied as described above. For example, in laser ablation scenarios, the same type of optical fiber is used and implanted in the same location in the "simulated thermal ablation environment" in multiple experiments, outputting laser with the same time-power curve. Similarly, in ultrasonic ablation scenarios, ultrasonic transducer arrays are arranged in the same location, and ultrasonic energy is output with the same time-power curve at each time period.

[0061] Because the aforementioned thermal ablation simulation experiment is reproducible, the reproducibility allows the MRI equipment sufficient time to acquire continuous-time MRI images of several slices during the same thermal ablation process in multiple experiments. For example, in the first experiment, the MRI equipment acquires an MRI image of the first slice during the entire ablation process; in the second experiment, it acquires an MRI image of the second slice during the entire ablation process, and so on. Through 10 experiments, 10 slices of MRI images are obtained. The data acquired from multiple experiments are combined to form continuous-time MRI images of multiple slices, overcoming the defect of poor real-time data acquisition of MRI equipment and avoiding the defect of poor image quality caused by trying to improve real-time data acquisition, thus obtaining high-quality continuous-time MRI images.

[0062] The above example describes the process of acquiring continuous-time magnetic resonance images (MRMI) of multiple fractures during a thermal ablation procedure. Similarly, more sets of MRMI images can be acquired at other ablation time periods and at other fracture locations to generate training samples. Furthermore, a different "thermal ablation procedure" with different physical conditions and thermal ablation operations can be reproduced to acquire more sets of "MRMI images" and generate even more training samples, thus enriching the training data and improving the generalization ability of the trained MRMI image reconstruction model.

[0063] This embodiment acquires magnetic resonance imaging data through thermal ablation experiments, avoiding the difficulty in obtaining clinical thermal ablation data. In particular, the reproducible thermal ablation experiments allow the magnetic resonance equipment sufficient time to acquire magnetic resonance images of the "same thermal ablation process" in multiple experiments. The magnetic resonance images acquired in multiple experiments are "stitched" into a continuous-time magnetic resonance image, overcoming the defect of poor real-time data acquisition of magnetic resonance equipment. It also avoids the defect of poor image quality caused by trying to improve the real-time data acquisition, and obtains high-quality continuous-time magnetic resonance images.

[0064] Based on any embodiment, in one embodiment, the thermal ablation magnetic resonance image dataset includes continuous-time magnetic resonance images at at least one tomographic location obtained during a thermal ablation experiment, and magnetic resonance images at at least one tomographic location from historical clinical thermal ablation data, S13 including:

[0065] The initial deep learning model was pre-trained using training samples from thermal ablation simulation experiments, and then a second training was performed using training samples from historical clinical thermal ablation data to obtain the deep learning model.

[0066] Specifically, compared to the previous embodiment, the thermal ablation magnetic resonance image dataset in this embodiment includes not only magnetic resonance images acquired in thermal ablation simulation experiments, but also magnetic resonance images collected and saved during past clinical thermal ablation processes. In the thermal ablation simulation experiments, continuous-time magnetic resonance images at several tomographic locations are acquired to monitor the thermal ablation process. Each tomographic continuous-time magnetic resonance image can be used to generate pre-training samples for pre-training the initial deep learning model through time downsampling. Since the thermal simulation experiments are reproducible, more tomographic and shorter time interval magnetic resonance image data of the "same thermal ablation process" can be collected through repeated experiments. This data can be easily and abundantly acquired for initial training of the initial deep learning model, preliminary optimization and adjustment of the model's parameters, giving the model a foundation for transfer learning. Subsequently, training samples generated from magnetic resonance images acquired at several tomographic locations during clinical thermal ablation processes can be used for secondary training of the model to further improve its performance. To obtain high-quality clinical thermal ablation data, data with fewer monitored tomographic fragments during the ablation process can be selected from historical clinical thermal ablation data. This improves the real-time performance and quality of data acquisition at these few fragments, acknowledging the scarcity of such data. Considering the difficulty in acquiring continuous-time MRI data clinically, the aforementioned historical clinical thermal ablation data need not be continuous tomographic MRI images. The model output can be used as input to supplement adjacent time frames until real data corresponds to the supplemented adjacent time frames. Comparing these adjacent time frames with the real data guides the adjustment of the deep learning model's parameters. For example, in clinical practice, magnetic resonance imaging (MRI) images of several tomographic locations are acquired cyclically with a 3-second cycle. MRI image data for the 1st, 4th, 7th, and subsequent seconds are acquired at a specific tomographic location. The data from the 1st second is input into a deep learning model to predict the simulated output for the 2nd second. The simulated output for the 2nd second is then input into the deep learning model to predict the simulated output for the 3rd second. The simulated output for the 3rd second is then input into the deep learning model to predict the simulated output for the 4th second. The simulated output for the 4th second is then compared with the actual acquired data for the 4th second, and the loss function is calculated to guide the adjustment of the deep learning model's parameters.

[0067] This embodiment acquires magnetic resonance imaging data through thermal ablation experiments, effectively making up for the lack and difficulty in obtaining clinical thermal ablation data. In particular, it solves the problem of the scarcity and difficulty in obtaining continuous time magnetic resonance imaging data at the same slice. Through pre-training based on thermal ablation simulation test data and secondary training based on historical clinical thermal ablation data, the performance of the model is improved.

[0068] Based on any embodiment, in one embodiment, the magnetic resonance image is a magnetic resonance K-space image, amplitude map, or phase map, or a magnetic resonance temperature map, temperature difference map, or ablation map generated based on the magnetic resonance time-domain image.

[0069] Specifically, the magnetic resonance images in the thermal ablation magnetic resonance image dataset can be K-space images. K-space is the space that stores the original magnetic resonance signal. The K-space image can be transformed to the image space through inverse Fourier transform to obtain amplitude and phase maps. Since the phase change is linearly related to the temperature change, a temperature difference map can be obtained by transforming the phase difference map. The temperature difference map combined with the baseline temperature can generate a temperature map. The temperature map can indicate the tissue temperature state at a specific fracture site at a certain moment. Furthermore, based on the cumulative effect of temperature over time, the ablation state of the tissue can be determined, i.e., an ablation map can be generated from the temperature map. In practical applications, downsampled K-space magnetic resonance images can be input into a magnetic resonance image reconstruction model to obtain time-series supplemented K-space magnetic resonance images. Then, based on the supplemented K-space magnetic resonance images, amplitude and phase maps, phase difference maps, temperature difference maps, temperature maps, and ablation maps can be generated. Of course, the magnetic resonance images in the thermal ablation magnetic resonance image dataset can also be amplitude maps and phase maps, phase difference maps, temperature difference maps, temperature maps or ablation maps. Training samples are generated by downsampling to train deep learning models. The trained models can be directly supplemented with time-series magnetic resonance time-domain images, temperature difference maps, temperature maps or ablation maps.

[0070] In some other embodiments, the magnetic resonance image can also be a temperature map. In this case, the training phase of the model uses training samples based on the temperature map, and the object reconstructed in the model application phase is directly the temperature map.

[0071] Based on any embodiment, in one embodiment, S12 generates training samples in the following manner:

[0072] Time downsampling is performed on continuous-time magnetic resonance images acquired at a fault location;

[0073] Several adjacent frames of magnetic resonance images at the fault location were downsampled and used as samples. The missing magnetic resonance images were used as the corresponding labeled data to form training samples, which were then added to the training sample set.

[0074] Specifically, suppose continuous-time magnetic resonance images (MRI) numbered 1 to 100 were acquired at a fault location. These images are then downsampled to remove even-numbered MRI images, leaving only odd-numbered ones. Next-order frames can then be used as samples. For example, frames 3 and 5 can be used as samples, with the missing downsampled frame 6 as the corresponding annotation data, forming a training sample. Similarly, frames 9, 11, and 13 can be used as samples, with the missing downsampled frame 14 as the corresponding annotation data, forming another training sample, and so on, generating a training sample set. Again, for example, if continuous-time MRI images numbered 1 to 100 were acquired at a fault location, frames 1, 4, and 7 can be used as samples, with the missing downsampled frames 8 and 9 as the corresponding annotation data, forming one training sample. Similarly, frames 4, 7, and 11 can be used as samples, with the missing downsampled frames 12 and 13 as the corresponding annotation data, forming another training sample, and so on, generating a training sample set.

[0075] Of course, other methods can also be used for time downsampling, such as removing magnetic resonance images with odd numbers, or removing one frame of magnetic resonance image every two frames.

[0076] The training sample set generated in the above manner is used to train the deep learning model, which enables the model to reconstruct the missing magnetic resonance image based on several frames of images after time downsampling at a certain fault during the thermal ablation process. This makes the magnetic resonance monitoring process have better temporal continuity and also improves the image acquisition efficiency.

[0077] Based on any embodiment, in one embodiment, the training samples further include tissue characteristic parameters and / or thermal ablation power data corresponding to the magnetic resonance image acquisition time. During model training, the tissue characteristic parameters and / or thermal ablation power data corresponding to the magnetic resonance image acquisition time are used as additional data of the magnetic resonance image to train the deep learning model.

[0078] Specifically, in this embodiment, when acquiring magnetic resonance image data, ablation power data and / or target tissue characteristic parameters are also acquired. For laser ablation, the ablation power here is the laser output power; for radio frequency ablation, the ablation power here is the radio frequency power. Target tissue characteristic parameters include thermal conductivity, specific heat capacity, electrical conductivity, and light absorption coefficient. These parameters and / or ablation power affect the temperature change trend and the grayscale value of the magnetic resonance image. Using these parameters and / or ablation power as supplementary data from the magnetic resonance image to train the deep learning model can improve the accuracy of the model's predictions. Specifically, the deep learning model has multiple input channels. The thermal ablation power data and / or target tissue characteristic parameters are input into the deep learning model as parallel data along with the magnetic resonance image data. The deep learning model will then provide a predicted missing magnetic resonance image. By comparing the predicted magnetic resonance image with the standard magnetic resonance image in the training samples, a loss function is calculated. This loss function is used to guide the adjustment of the deep learning model's parameters. Furthermore, a deep learning model can be a single-channel output, that is, predicting the missing data of a single frame, or it can be a multi-channel output, that is, predicting the missing data of multiple frames at once.

[0079] Based on any embodiment, in one embodiment, S12 generates training samples in the following manner:

[0080] The magnetic resonance image and ablation power acquired at a certain moment at a fault location are used as samples, and the missing temperature maps at adjacent moments are used as labeled data to form training samples, which are then added to the training sample set.

[0081] Specifically, a magnetic resonance imaging (MRI) image and ablation power acquired at a certain moment at a fault location are used as joint samples. Missing temperature maps at adjacent moments at the same fault location are used as corresponding labeled data to form training samples. This process is repeated to generate a training sample set for training a deep learning model. The trained model can predict missing temperature maps based on downsampled MRI images and ablation power during thermal ablation, thereby reducing data acquisition volume and improving data acquisition efficiency and the real-time performance of ablation monitoring. For example, an MRI image and ablation power acquired at the 5th second at a fault location are used as joint samples. Missing temperature maps at the 6th and 7th seconds at the same fault location are used as corresponding labeled data to form a training sample set, which is then used to train the deep learning model.

[0082] Based on any embodiment, in one embodiment, S12 generates training samples in the following manner:

[0083] Time downsampling is performed on continuous-time magnetic resonance images acquired at a fault location;

[0084] All magnetic resonance images before a certain moment after downsampling at the fault location are used as samples, and the thermal ablation map at that moment is used as labeled data to form training samples, which are then added to the training sample set.

[0085] Specifically, all magnetic resonance images prior to a certain data gap point after downsampling are used as samples, and the corresponding ablation maps at that point are used as standard data to form training samples. This process is repeated to generate a training sample set for training the deep learning model. The trained model can predict missing ablation maps based on downsampling magnetic resonance images, reducing the computational load of ablation map calculations. It can also fill in missing frames of ablation maps; that is, when an image is acquired in the current frame, the ablation map for that frame can be directly calculated, and missing ablation maps can be filled in using the above method, improving data acquisition efficiency and the real-time performance of ablation monitoring. It should also be noted that the ablation state of tissue is affected by temperature and the duration of temperature fluctuations. The ablation maps in the thermal ablation magnetic resonance image dataset can be calculated using the Arrenius formula or the CEM43 model.

[0086] Based on any embodiment, in one embodiment, the deep learning model is selected from any of the following: recurrent neural network (RNN), generative adversarial network (GAN), long short-term memory network (LSTM), etc.

[0087] Recurrent Neural Networks (RNNs) are a type of recurrent neural network that takes sequential data as input, recursively moves along the sequence, and all nodes are connected in a chain-like manner. Generative Adversarial Networks (GANs) consist of a generative model and a discriminative model. The generative model is responsible for capturing the distribution of sample data, while the discriminative model is typically a binary classifier that distinguishes between real data and generated samples. During training, one model is fixed while the parameters of the other are updated, and this process is repeated iteratively until the generative model can estimate the distribution of the sample data. Long Short-Term Memory (LSTM) networks are a type of recurrent neural network that selectively remembers and forgets information through gating mechanisms, making them particularly suitable for domains that require understanding or predicting time series data.

[0088] A preferred embodiment of the present invention is described below:

[0089] Step 1: Obtain a "continuous-time magnetic resonance image" at at least one fault location by reproducing the thermal ablation simulation experiment and add it to the thermal ablation magnetic resonance image dataset.

[0090] Specifically, reproducible thermal ablation simulation experiments refer to simulating the external environment and biological tissue during the thermal ablation process, and then performing the thermal ablation operation and acquiring magnetic resonance images based on this simulation. Because these thermal ablation simulation experiments are reproducible, the magnetic resonance equipment has sufficient time to acquire magnetic resonance images of several sections representing the "same thermal ablation process" in multiple experiments. These images are then "stitched" together to form a continuous-time magnetic resonance image, overcoming the poor real-time performance of magnetic resonance equipment and avoiding the loss of image quality due to efforts to improve real-time performance. This results in high-quality continuous-time magnetic resonance images. The magnetic resonance images can be magnetic resonance K-space images, amplitude maps, phase maps, or magnetic resonance temperature maps, temperature difference maps, or ablation maps generated based on the magnetic resonance time-domain images.

[0091] Step 2: For each fault location, generate training samples by time downsampling and add them to the training sample set.

[0092] Specifically, for each continuous-time magnetic resonance imaging (MRI) image at a specific topograph, one or more training samples can be generated through temporal downsampling. For example, if MRI images A1, A2, A3, A4, A5, A6, A7, and A8 are continuously acquired at a topograph, a training sample can be generated using A1 and A3 as samples and A2 as the corresponding annotation, and so on, using A2 and A4 as samples and A3 as the corresponding annotation. Similarly, if MRI images B1, B2, B3, B4, B5, B6, B7, and B8 are continuously acquired at a topograph, a training sample can be generated using B1 and B3 as samples and B4 as the corresponding annotation, and so on, using B2 and B4 as samples and B5 as the corresponding annotation. By generating training samples from different patients and at different topographs, the diversity of training samples is increased. Consequently, the generalization ability of the subsequently trained model can be improved, enabling it to better reconstruct the temporally downsampled MRI images during actual thermal ablation and fill in the missing time series.

[0093] Furthermore, target tissue characteristic parameters and / or ablation power can be used together with magnetic resonance images as associated data to generate training samples, so that the trained magnetic resonance image reconstruction model can supplement the missing time series based on time-downsampled images and power, thereby improving the accuracy of model reconstruction.

[0094] Step 3: Train a deep learning model based on the training sample set to obtain a magnetic resonance image reconstruction model.

[0095] Specifically, a deep learning model is trained using training samples, the model parameters are optimized, and the best-performing model is selected. This model can then be used to reconstruct time-downsampled magnetic resonance images during subsequent thermal ablation processes. Because the training samples in the aforementioned training sample set cover the entire thermal ablation process, the trained model can supplement the time-downsampled sequences at any stage of real thermal ablation, exhibiting good generalization ability.

[0096] Furthermore, continuous-time magnetic resonance images at at least one tomographic location from historical clinical thermal ablation data can be used to generate training samples for secondary training of the magnetic resonance image reconstruction model, further improving the model's performance. Overall, the model is trained once using readily available, reproducible simulated ablation experiments, providing a foundation for transfer learning. Secondary training with more realistic and scarce historical clinical thermal ablation data combines the advantages of both types of data, thus enhancing model performance.

[0097] In this embodiment, continuous-time magnetic resonance images are acquired at at least one fracture site during the thermal ablation process. Training samples are generated from the magnetic resonance images at each fracture site through temporal downsampling. These samples are then used to train a deep learning model. The model can complete the downsampled time series, resulting in better temporal continuity of the output images. This reduces the amount of data collected during the ablation process and improves the real-time performance of temperature monitoring. Furthermore, since the training samples cover the entire thermal ablation process and involve multiple fracture locations, the trained model can accurately reconstruct the time series of magnetic resonance images at all locations throughout the entire thermal ablation process, further improving the quality of temperature monitoring.

[0098] The following describes a magnetic resonance image reconstruction method provided by the present invention. The magnetic resonance image reconstruction method described below and the training method of the magnetic resonance image reconstruction model described above can be referred to and correspond to each other.

[0099] Figure 3 This is a flowchart illustrating a magnetic resonance image reconstruction method provided by the present invention, as shown below. Figure 3 As shown, the method includes:

[0100] S21. Acquire the current patient's time-downsampled magnetic resonance image;

[0101] S22. Input the time-downsampled magnetic resonance image into the magnetic resonance image reconstruction model to obtain the time-reconstructed magnetic resonance image, which provides information support for the thermal ablation process.

[0102] The magnetic resonance image reconstruction model is pre-trained according to the training method of any of the aforementioned magnetic resonance image reconstruction models.

[0103] Specifically, acquiring MRI images of the patient via time downsampling during thermal ablation can accelerate acquisition efficiency and improve the real-time performance of MRI monitoring. By inputting time-downsampled images from the same slice of the patient into an MRI image reconstruction model pre-trained using the aforementioned training method, the model can output MRI images with supplemented time-series reconstructions, improving the temporal continuity of the monitoring process and providing information support for the thermal ablation procedure. It is understood that the time downsampling method needs to be consistent with the time downsampling strategy used in the training samples during the model training phase.

[0104] In this embodiment, during the model training phase, training samples are generated using time downsampling based on a thermal ablation magnetic resonance image dataset to train the deep learning model. The dataset covers data from the entire thermal ablation process and all locations, resulting in a deep learning model that more closely resembles real-world thermal ablation scenarios. During the model application phase, the time-downsampled magnetic resonance images are input into the trained deep learning model to obtain time-series supplemented magnetic resonance images. This reduces the amount of magnetic resonance data collected during thermal ablation, improving real-time monitoring while ensuring data quality.

[0105] Furthermore, in some embodiments, the method further includes inputting the time-reconstructed magnetic resonance image into a spatial reconstruction model to obtain a further spatially reconstructed magnetic resonance image; wherein the spatial reconstruction model is trained using data collected from reproducible thermal ablation experiments.

[0106] Specifically, after time-series supplementation of magnetic resonance images, spatial reconstruction models can be used to supplement spatial data of the magnetic resonance images, that is, to supplement data at missing tomographic locations. The aforementioned spatial reconstruction model can be a model trained based on data collected from thermal ablation experiments. For example, the spatial reconstruction model can be trained in the following way:

[0107] Step 1: Obtain the thermal ablation magnetic resonance image dataset; the thermal ablation magnetic resonance image dataset includes magnetic resonance images of continuous tomography during the thermal ablation process;

[0108] Specifically, this thermal ablation magnetic resonance image dataset consists of magnetic resonance image data acquired during the thermal ablation process. It includes multiple sets of continuous tomographic images. For example, in a patient's thermal ablation history data, a target area with a size of 3 cm may be present. 3During the thermal ablation process, magnetic resonance imaging (MRI) images of 10 consecutive tomographic locations covering the target area were acquired and stored from the patient. It is understood that MRI images are cyclically acquired at these consecutive tomographic locations during the thermal ablation process, and a set of MRI images from each cycle can be used to generate training samples. Preferably, the consecutive tomographic MRI images in the thermal ablation MRI image dataset are obtained through thermal ablation experiments, such as animal ablation experiments; more preferably, the thermal ablation experiments include reproducible thermal ablation simulation experiments.

[0109] Step 2: Based on the magnetic resonance images of continuous tomography in the thermal ablation magnetic resonance image dataset, generate a spatial reconstruction training sample set by downsampling;

[0110] Specifically, image downsampling is performed on the thermal ablation dataset. Downsampling can be between different tomographic locations, i.e., removing some tomographic images from several consecutive tomographic images. Alternatively, downsampling can be performed on magnetic resonance amplitude images, magnetic resonance phase maps, or images related to magnetic resonance during the thermal ablation process, such as phase difference maps, temperature difference maps, temperature maps, and ablation maps generated based on the magnetic resonance phase map.

[0111] Step 3: Train the spatial reconstruction model based on the spatial reconstruction training sample set.

[0112] Specifically, the deep learning model is trained using data from the spatial reconstruction training sample set. This training sample set can be split into a training set, a validation set, and a test set. The deep learning model learns patterns and rules from the training samples in the training set, enabling it to make predictions about unseen data. In this embodiment, the spatial reconstruction training samples cover the entire thermal ablation process, thus allowing the trained model to reconstruct downsampled data from any stage of real thermal ablation. The validation set is used to adjust the hyperparameters of the deep learning model, improving its performance and generalization ability. The test set is used to evaluate the model and select the best-performing model. This model can then be used for spatial reconstruction of magnetic resonance images during subsequent thermal ablation processes, supplementing missing tomographic data and providing more information support for the thermal ablation process.

[0113] Furthermore, in the above embodiments, the reconstructed data can be a magnetic resonance K-space image, an amplitude map, or a phase map. Since phase changes are linearly related to temperature changes, a temperature difference map can be obtained by transforming the phase difference map. Combining the temperature difference map with the baseline temperature generates a temperature map. The temperature map indicates the tissue temperature state at a specific fracture site at a given time. Furthermore, based on the cumulative effect of temperature over time, the ablation state of the tissue can be determined, i.e., an ablation map can be generated from the temperature map. In other embodiments, the reconstructed data can also be a temperature difference map, a temperature map, or an ablation map.

[0114] The following describes a magnetic resonance thermal ablation monitoring method provided by the present invention. The magnetic resonance thermal ablation monitoring method described below can be referred to in correspondence with the training method of the magnetic resonance image reconstruction model described above.

[0115] Figure 4 This is a flowchart illustrating a magnetic resonance thermal ablation monitoring method provided by the present invention, as shown below. Figure 4 As shown, the method includes:

[0116] S31. Acquire the current patient's time-downsampled magnetic resonance image;

[0117] S32. Input the time-downsampled magnetic resonance image into the magnetic resonance image reconstruction model to obtain the reconstructed ablation map, which provides information support for the thermal ablation process;

[0118] The magnetic resonance image reconstruction model is pre-trained according to the training method of any of the aforementioned magnetic resonance image reconstruction models.

[0119] Specifically, acquiring the patient's MRI images during thermal ablation using time downsampling can accelerate acquisition efficiency and improve the real-time performance of MRI monitoring. By inputting the time-downsampled images of the same slice of the patient into an MRI image reconstruction model pre-trained using the aforementioned training method, the model can output ablation maps, providing information support for the thermal ablation process. It is understood that the time downsampling method needs to be consistent with the time downsampling strategy used during model training.

[0120] In this embodiment, during the model training phase, training samples are generated using time downsampling based on a thermal ablation magnetic resonance image dataset to train the deep learning model. The dataset covers the entire process and all locations of real ablation, resulting in a deep learning model that is more consistent with the thermal ablation scenario. During the model application phase, the time downsampling magnetic resonance images are input into the trained deep learning model to obtain the ablation map, reducing the amount of magnetic resonance data acquisition during thermal ablation and improving the real-time performance of thermal ablation monitoring while maintaining data continuity.

[0121] With the support of this method, users can cyclically acquire magnetic resonance images at various tomographic locations during clinical thermal ablation (due to the time-consuming image acquisition, there is a time difference in the acquisition of images at each tomographic location). For a certain tomographic location, the acquisition cycle of the magnetic resonance image at that location is the aforementioned cycle cycle. The thermal ablation-related image generation model trained by this method can perform time-series supplementation on the magnetic resonance image at that tomographic location (equivalent to the magnetic resonance image after time downsampling) to obtain a continuous-time magnetic resonance image. By supplementing the magnetic resonance images at each tomographic location, thermal ablation-related images with a higher refresh rate can be obtained, improving monitoring quality and providing more comprehensive information support for the thermal ablation process.

[0122] The present invention also provides a magnetic resonance image reconstruction system (electronic device), including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the training method of any of the aforementioned magnetic resonance image reconstruction models.

[0123] The present invention also provides a magnetic resonance thermal ablation monitoring system (electronic device), including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the aforementioned magnetic resonance image reconstruction methods.

[0124] The present invention also provides a magnetic resonance thermal ablation monitoring system (electronic device), including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the aforementioned magnetic resonance ablation monitoring methods.

[0125] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0126] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A training method for a magnetic resonance image reconstruction model, characterized in that, include: Acquire a thermal ablation magnetic resonance image dataset; wherein the thermal ablation magnetic resonance image dataset includes continuous-time magnetic resonance images at at least one fracture site during the thermal ablation process; For each fault, continuous-time magnetic resonance images are generated by time downsampling to create training samples, which are then added to the training sample set. A deep learning model is trained using the training sample set to obtain a magnetic resonance image reconstruction model. The thermal ablation magnetic resonance image dataset includes continuous-time magnetic resonance images of at least one fracture site obtained through thermal ablation experiments; the thermal ablation experiments include animal ablation experiments and / or thermal ablation simulation experiments.

2. The training method according to claim 1, characterized in that, The thermal ablation magnetic resonance image dataset includes continuous-time magnetic resonance images of at least one tomographic region obtained through thermal ablation experiments, as well as magnetic resonance images of at least one tomographic region from historical clinical thermal ablation data. Training a deep learning model using the training sample set to obtain a magnetic resonance image reconstruction model includes: The initial deep learning model is pre-trained using training samples corresponding to the thermal ablation experiment, and then a second training is performed using training samples corresponding to the historical clinical thermal ablation data to obtain the deep learning model.

3. The training method according to claim 1, characterized in that, The magnetic resonance image is a magnetic resonance K-space image, amplitude map, or phase map, or a magnetic resonance temperature map, temperature difference map, or ablation map generated based on the phase map.

4. The training method according to claim 1, characterized in that, For each fault location, the continuous-time magnetic resonance image is used to generate training samples through time downsampling, and these training samples are added to the training sample set as follows: Time downsampling is performed on continuous-time magnetic resonance images acquired at a fault location; Several adjacent frames of magnetic resonance images at the fault location after downsampling are used as samples, and the missing magnetic resonance images are used as corresponding labeled data to form training samples, which are then added to the training sample set.

5. The training method according to claim 4, characterized in that, The step of generating training samples from continuous-time magnetic resonance images at each fracture site by time downsampling and adding them to the training sample set also includes: using the target area tissue characteristic parameters and / or the thermal ablation power data corresponding to the magnetic resonance image acquisition time as additional data for training the deep learning model.

6. The training method according to claim 5, characterized in that, Training samples are generated as follows: The magnetic resonance image and ablation power acquired at a certain moment at a fault location are used as samples, and the missing temperature maps at adjacent moments are used as labeled data to form training samples, which are then added to the training sample set.

7. The training method according to claim 1, characterized in that, Training samples are generated as follows: Time downsampling is performed on continuous-time magnetic resonance images acquired at a fault location; All magnetic resonance images prior to a certain moment after downsampling at the fault location are used as samples, and the thermal ablation image at that moment is used as labeled data to form training samples, which are then added to the training sample set.

8. The training method according to claim 1, characterized in that, The deep learning model is selected from any of the following: recurrent neural network, generative adversarial network, long short-term memory network.

9. A magnetic resonance image reconstruction method, characterized in that, include: Acquire time-downsampled magnetic resonance images of the current patient; The time-downsampled magnetic resonance image is input into the magnetic resonance image reconstruction model to obtain the time-reconstructed magnetic resonance image, which provides information support for the thermal ablation process; The magnetic resonance image reconstruction model is pre-trained using the training method for the magnetic resonance image reconstruction model according to any one of claims 1-8.

10. A magnetic resonance image reconstruction method according to claim 9, characterized in that, Also includes: The time-reconstructed magnetic resonance image is input into the spatial reconstruction model to obtain a further spatially reconstructed magnetic resonance image; wherein, the spatial reconstruction model is trained using data collected from reproducible thermal ablation experiments.

11. The magnetic resonance image reconstruction method according to any one of claims 9 or 10, characterized in that, It also includes generating temperature maps and / or ablation maps based on magnetic resonance images acquired at various times.

12. A method for monitoring magnetic resonance ablation, characterized in that, include: Acquire time-downsampled magnetic resonance images of the current patient; The time-downsampled magnetic resonance image is input into the magnetic resonance image reconstruction model to obtain the reconstructed ablation map, which provides information support for the thermal ablation process; The magnetic resonance image reconstruction model is pre-trained using the training method for the magnetic resonance image reconstruction model according to any one of claims 1-8.

13. An electronic device, characterized in that, The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the training method for the magnetic resonance image reconstruction model as described in any one of claims 1 to 8, or the steps of the magnetic resonance image reconstruction method as described in any one of claims 9 to 11, or the steps of the magnetic resonance ablation monitoring method as described in claim 12.

Images