Training method of magnetic resonance image reconstruction model and magnetic resonance image reconstruction method

Abstract

The invention provides a magnetic resonance image reconstruction method and a model training method. The training method comprises the following steps: acquiring a thermal ablation magnetic resonance image data set; wherein the thermal ablation magnetic resonance image data set comprises a magnetic resonance image in a thermal ablation process; generating a training sample set through downsampling according to magnetic resonance images in the thermal ablation magnetic resonance image data set; and training a deep learning model according to the training sample set to obtain a magnetic resonance image reconstruction model. According to the method, the magnetic resonance image data set in the thermal ablation process is obtained, the training sample set is generated through downsampling to train the deep learning model, and the thermal ablation magnetic resonance image data set contains data of the whole thermal ablation process and all positions; the model obtained by corresponding training has better reconstruction capability for downsampling magnetic resonance data in each stage in the thermal ablation process, so that the magnetic resonance data acquisition efficiency can be improved while the imaging quality is guaranteed, and a larger acquisition range is covered.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a training method for a magnetic resonance image reconstruction model and a magnetic resonance image reconstruction method. Background Art

[0002] Thermal ablation is a technology that uses thermal effects to cause coagulation, necrosis, vaporization, and carbonization of diseased tissue, thereby achieving the purpose of ablation and inactivation treatment. It includes laser, high-frequency electric knife, plasma coagulation, microwave therapy, and radiofrequency therapy. Magnetic resonance imaging provides information support for the thermal ablation process. Magnetic resonance imaging can monitor the temperature state of the tissue, allowing doctors to understand the current ablation status of the target area, adjust the ablation position more accurately, and control the ablation power, thereby achieving the goal of "ablating lesions as much as possible and protecting normal tissue."

[0003] The drawback of magnetic resonance imaging is that it is time-consuming, and the temperature field changes rapidly during thermal ablation, which places high demands on the real-time performance of temperature imaging. Currently, several measures have been proposed to improve the real-time performance of temperature imaging: ① Narrowing the imaging range and acquiring fewer layers (e.g., three layers). This approach has the following drawbacks: 1) Reducing the imaging range and acquiring fewer layers (e.g., three layers) has limited improvements in imaging efficiency, and because the lesion itself occupies a certain space, too few layers may not cover the lesion, resulting in a temperature monitoring blind spot; 2) Designing magnetic resonance sequences with shorter acquisition times, such as EPI sequences, has the following drawbacks: Significantly reduced imaging quality and large errors in the acquired temperature data; 3) Using downsampling to reduce the amount of data acquired, improve acquisition efficiency, and using algorithms to supplement missing data (e.g., interpolation algorithms). This approach has the drawback of poor image reconstruction quality.

[0004] In order to overcome or at least partially overcome the above-mentioned drawbacks, the present invention provides a magnetic resonance image reconstruction method and a model training method. Summary of the Invention

[0005] The present invention provides a magnetic resonance image reconstruction method and a model training method to address the shortcomings of the prior art magnetic resonance imaging, such as poor real-time performance or low imaging quality.

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

[0007] Acquiring a thermal ablation magnetic resonance image dataset; wherein the thermal ablation magnetic resonance image dataset includes magnetic resonance images during the thermal ablation process;

[0008] generating a training sample set by downsampling the magnetic resonance images in the thermal ablation magnetic resonance image dataset;

[0009] A deep learning model is trained according to the training sample set to obtain a magnetic resonance image reconstruction model.

[0010] Optionally, the thermal ablation magnetic resonance image data set includes several groups of magnetic resonance images at consecutive slices during the thermal ablation process.

[0011] Furthermore, generating a training sample set by downsampling the magnetic resonance images in the thermal ablation magnetic resonance image dataset includes:

[0012] Inter-layer downsampling is performed on a set of magnetic resonance images of consecutive slices, and the magnetic resonance images of several adjacent slices after downsampling are used as samples. The missing slices near the several adjacent slices after downsampling are used as corresponding annotation data to form training samples, which are added to the training sample set.

[0013] Optionally, the thermal ablation magnetic resonance image dataset includes magnetic resonance images at several slices during a thermal ablation process, and generating a training sample set by downsampling the magnetic resonance images in the thermal ablation magnetic resonance image dataset includes:

[0014] The magnetic resonance images at each slice are subjected to intra-slice downsampling, and the magnetic resonance images after intra-slice downsampling are used as samples. The data eliminated during the intra-slice downsampling process are used as corresponding labeled data to form training samples, which are added to the training sample set.

[0015] Furthermore, the intra-layer downsampling is intra-layer interval downsampling, or intra-layer random downsampling, or Gaussian distribution sampling.

[0016] Optionally, the magnetic resonance image in the thermal ablation magnetic resonance image dataset is an amplitude image, a phase image, or a phase difference image, a temperature difference image, a temperature image or an ablation image generated based on the phase image.

[0017] Optionally, the thermal ablation magnetic resonance image dataset includes magnetic resonance images acquired through a reproducible thermal ablation simulation test.

[0018] Optionally, the thermal ablation magnetic resonance image dataset includes magnetic resonance images acquired through reproducible thermal ablation simulation experiments and magnetic resonance images acquired from historical clinical thermal ablation data, and training a deep learning model based on the training sample set to obtain a magnetic resonance image reconstruction model includes:

[0019] The initial deep learning model is pre-trained using the training samples corresponding to the thermal ablation simulation test, and secondary training is performed using the training samples corresponding to the historical clinical thermal ablation data to obtain the deep learning model.

[0020] Optionally, the deep learning model is selected from any one of the following: convolutional neural network, recurrent neural network, and generative adversarial network.

[0021] In a second aspect, the present invention further provides a magnetic resonance image reconstruction method, comprising:

[0022] Obtaining a downsampled magnetic resonance image of the current patient;

[0023] Inputting the downsampled magnetic resonance image into a magnetic resonance image reconstruction model to obtain a reconstructed magnetic resonance image to provide information support for the thermal ablation process;

[0024] The magnetic resonance image reconstruction model is pre-trained according to any of the aforementioned training methods for magnetic resonance image reconstruction models.

[0025] In a third aspect, the present invention also provides a magnetic resonance-guided thermal ablation monitoring system, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps of the training method of the magnetic resonance image reconstruction model as described in any of the foregoing items, or implements the steps of the magnetic resonance image reconstruction method as described in any of the foregoing items.

[0026] The magnetic resonance image reconstruction method and model training method provided by the present invention have at least the following beneficial effects:

[0027] 1. By acquiring magnetic resonance images during thermal ablation, training samples are generated through inter-layer downsampling for training deep learning models. The model can supplement the missing magnetic resonance data, thereby reducing the amount of data collected during thermal ablation and improving the real-time performance of data collection.

[0028] 2. In some implementations, the thermal ablation magnetic resonance image dataset includes several groups of magnetic resonance images at consecutive faults during the thermal ablation process. The downsampling process is inter-layer downsampling. The trained image reconstruction model can reconstruct the inter-layer downsampled magnetic resonance image data during the thermal ablation process, supplement the missing inter-layer data, and obtain a continuous three-dimensional image.

[0029] 3. In some implementations, the thermal ablation magnetic resonance image dataset includes magnetic resonance images of several slices during the thermal ablation process. The downsampling process is intra-layer downsampling. The trained image reconstruction model can reconstruct the intra-layer downsampled magnetic resonance image data during the thermal ablation process, supplement the missing intra-layer data, and improve image acquisition efficiency.

[0030] 4. Magnetic resonance image data were collected through reproducible thermal ablation simulation experiments, which effectively made up for the problem of insufficient and difficult to obtain real thermal ablation data, especially solving the problem of insufficient and difficult to obtain magnetic resonance image data at continuous faults.

[0031] 5. The model was pre-trained using thermal ablation simulation test data, providing the model with a 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 model performance.

[0032] 6. Some implementations support reconstruction of magnetic resonance amplitude images, some implementations support reconstruction of magnetic resonance phase images, and some implementations support reconstruction of phase difference images, temperature difference images, temperature images or ablation images generated based on magnetic resonance phase images, which better meet the monitoring needs of the thermal ablation process. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0034] Figure 1 1 is a flow chart of a method for training a magnetic resonance image reconstruction model provided by the present invention;

[0035] Figure 2 1 is a flow chart of the magnetic resonance image reconstruction method provided by the present invention;

[0036] Figure 3 It is a flow chart of the magnetic resonance temperature imaging method provided by the present invention. DETAILED DESCRIPTION

[0037] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0038] The following combination Figure 1-Figure 3 Describe a magnetic resonance image reconstruction method and model training method of the present invention, Figure 1 FIG. 1 is a flow chart of a training method for a magnetic resonance image reconstruction model provided by the present invention, such as Figure 1 As shown, the method includes:

[0039] S11. Acquire a thermal ablation magnetic resonance image dataset; wherein the thermal ablation magnetic resonance image dataset includes magnetic resonance images during the thermal ablation process;

[0040] Specifically, the thermal ablation magnetic resonance image dataset is magnetic resonance image data collected during the entire thermal ablation process, which includes several tomographic images covering the patient's target area. For example, in the thermal ablation history data of a patient, the size of a target area is 3 cm. 3 During the entire thermal ablation process, the full thermal ablation process data collected from the patient at 10 consecutive faults covering the target area are collected and saved. For another example, full sampling images of the patient at 4 interval faults during the thermal ablation process are collected. The magnetic resonance image dataset may include the original magnetic resonance image, and may also include temperature maps, ablation maps and other data generated based on the original magnetic resonance images during the thermal ablation process. It is understandable that the data in the magnetic resonance image dataset can come from different patients and be collected at different stages of thermal ablation to improve the richness of the data and the generalization ability of the model. In addition, the magnetic resonance images in the above-mentioned thermal ablation magnetic resonance image dataset may be magnetic resonance amplitude images, magnetic resonance phase maps, or phase difference maps, temperature difference maps, temperature maps, ablation maps and other images related to magnetic resonance during the thermal ablation process generated based on the magnetic resonance phase maps.

[0041] S12, generating a training sample set by downsampling the magnetic resonance images in the thermal ablation magnetic resonance image dataset;

[0042] Specifically, downsampling is performed on the images in the thermal ablation MRI dataset. Downsampling can be performed between different slice positions, i.e., removing some slice images from several consecutive slice images. Downsampling can also be performed within a slice of a MRI image. Downsampling strategies include random downsampling, interval downsampling, and Gaussian distribution downsampling.

[0043] S13. Train a deep learning model based on the training sample set to obtain a magnetic resonance image reconstruction model.

[0044] Specifically, a training sample set is used to train a deep learning model, and the 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 through the training samples in the training set so that it can make predictions about unseen data. The training samples of the present invention cover the data of the entire thermal ablation process, so that the trained model can restore and reconstruct the downsampled data at any stage of the actual thermal ablation. The validation set is used to adjust the hyperparameters of the deep learning model and improve the performance and generalization ability of the model. The test set is used to evaluate the model and screen out the best performing model, which can be used for magnetic resonance image reconstruction in the subsequent thermal ablation process.

[0045] This embodiment collects a magnetic resonance image dataset during the thermal ablation process, and generates a training sample set through downsampling to train a deep learning model. Since the thermal ablation magnetic resonance image dataset contains data on the entire process and all positions of thermal ablation, the corresponding trained model has better reconstruction capabilities for the downsampled magnetic resonance data at each stage of the thermal effect process, thereby improving the efficiency of magnetic resonance data acquisition during the thermal ablation process while ensuring imaging quality, and covering a larger acquisition range.

[0046] Based on any one of the embodiments, in one embodiment, the thermal ablation magnetic resonance image dataset includes several groups of magnetic resonance images at consecutive slices, and the downsampling is inter-slice downsampling of the magnetic resonance images at a group of consecutive slices.

[0047] Specifically, in this embodiment, the thermal ablation magnetic resonance image data set includes magnetic resonance images at continuous faults collected in groups. During the clinical thermal ablation process, multiple groups of continuous fault magnetic resonance images can be collected at different ablation stages. Accordingly, the downsampling process is to perform inter-layer downsampling on each group of magnetic resonance images.

[0048] This embodiment generates training samples for training a deep learning model by performing inter-layer downsampling on magnetic resonance images at continuous slices, so that the model can restore and reconstruct the magnetic resonance images after inter-layer downsampling during the thermal ablation process, which is beneficial to improving the acquisition efficiency of magnetic resonance images during thermal ablation and improving the real-time performance of magnetic resonance monitoring.

[0049] Based on the previous embodiment, in one embodiment, the S12 includes: performing inter-layer downsampling on a group of magnetic resonance images of continuous slices, using the magnetic resonance images of several adjacent slices after downsampling as samples, and using the missing slices near the several adjacent slices after downsampling as corresponding annotation data to form training samples, and adding them to the training sample set.

[0050] Specifically, the thermal ablation magnetic resonance image data set includes multiple groups of magnetic resonance images periodically acquired at consecutive slices as the thermal ablation process proceeds. The downsampling process is to perform inter-layer downsampling on each group of magnetic resonance images. For example, images at some slices in a group of magnetic resonance images are removed, and the remaining images are used as samples. The images of the removed slices are used as labeled data to form a training sample, which is added to the training sample set. For another example, inter-layer downsampling is performed on a portion of a group of continuous magnetic resonance images. The downsampled magnetic resonance images are used as samples, and the magnetic resonance images of the portion before downsampling are used as labeled data to form a training sample, which is added to the training sample set.

[0051] For example, a set of magnetic resonance images at consecutive slices includes A1, A2, A3, A4, A5, A6, A7, A8, and A9. The even-numbered layers A2, A4, A6, and A8 are removed through inter-slice downsampling. Then, the remaining odd-numbered layers A1, A3, A5, A7, and A9 can be used as samples, and the removed even-numbered layers can be used as corresponding labeled data to form a training sample and added to the training sample set. For another example, a set of magnetic resonance images at consecutive slices includes A1, A2, A3, A4, A5, A6, A7, A8, and A9. A3, A6, and A9 are removed through inter-slice downsampling. Then, the remaining A1 and A2 can be used as samples, and A3 can be used as corresponding labeled data to form a training sample; the remaining A4 and A5 can be used as samples, and A6 can be used as corresponding labeled data to form a training sample; the remaining A7 and A8 can be used as samples, and A9 can be used as corresponding labeled data to form a training sample. For another example, a set of magnetic resonance images at consecutive slices include A1, A2, A3, A4, A5, A6, A7, and A8. A2, A4, A6, and A9 are eliminated through inter-layer downsampling. Then, A1 and A3 can be used as samples, and A4 can be used as the corresponding labeled data to form a training sample; A3 and A5 can be used as samples, and A6 can be used as the corresponding labeled data to form a training sample; A5 and A7 can be used as samples, and A8 can be used as the corresponding labeled data to form a training sample.

[0052] This embodiment generates training samples for training a deep learning model by performing inter-layer downsampling on magnetic resonance images at continuous slices, so that the model can restore and reconstruct the magnetic resonance images after inter-layer downsampling during the thermal ablation process, which is beneficial to improving the acquisition efficiency of magnetic resonance images during thermal ablation and improving the real-time performance of magnetic resonance monitoring.

[0053] Based on any embodiment, in one embodiment, the thermal ablation magnetic resonance image data set includes magnetic resonance images at several slices during the thermal ablation process, and the S12 includes: performing intra-layer downsampling on the magnetic resonance images at each slice, using the magnetic resonance images after intra-layer downsampling as samples, and using the data eliminated during the intra-layer downsampling process as corresponding labeled data to form training samples, which are added to the training sample set.

[0054] Specifically, unlike the above embodiment, the positions of the slices in the "magnetic resonance images at several slices" obtained in this embodiment do not have to be continuous. In addition, this embodiment performs downsampling within a single slice image, and generates training samples through inter-layer downsampling to train the magnetic resonance image reconstruction model, so that the model has the ability to recover the missing data of the downsampling.

[0055] More specifically, in one embodiment, the S12 includes:

[0056] S121. Acquire a first magnetic resonance image from the thermal ablation magnetic resonance image dataset;

[0057] S122. Perform intra-slice downsampling on the first magnetic resonance image to obtain a second magnetic resonance image;

[0058] S123: Using the second magnetic resonance image as a sample and the first magnetic resonance image as a corresponding annotation, form a training sample, and add the sample to a training sample set;

[0059] The first magnetic resonance image is any image in the thermal ablation magnetic resonance image dataset, and the above steps S121 to S123 are repeatedly performed until a final training sample set is obtained.

[0060] In this embodiment, the magnetic resonance images in the thermal ablation magnetic resonance image dataset are downsampled to form training samples, which can reduce the data acquisition amount of one frame of magnetic resonance image during the thermal ablation process. Then, the deep learning model trained in this embodiment is used to restore the downsampled magnetic resonance images during the thermal ablation process, thereby improving the acquisition efficiency of magnetic resonance images and the real-time performance of temperature monitoring.

[0061] Based on the previous embodiment, in one embodiment, the downsampling is intra-layer interval downsampling, or intra-layer random downsampling, or Gaussian distribution downsampling.

[0062] Specifically, this embodiment is to downsample a single magnetic resonance image. The downsampling process can be interval sampling, for example, row interval sampling, column interval sampling, adjacent interval sampling (that is, the four pixels adjacent to a sampling pixel in the orthogonal direction are no longer sampled); the downsampling process can also be random downsampling, with a sampling ratio of, for example, 50%, 40%, 30%, 20%, etc.; the downsampling can also be Gaussian distribution downsampling, for example, the closer to the center of the image, the higher the sampling ratio, and the closer to the edge of the image, the lower the sampling ratio, and the sampling ratio is Gaussian distributed as a whole.

[0063] Based on the previous embodiment, in one embodiment, the magnetic resonance image in the thermal ablation magnetic resonance image dataset is an amplitude image, a phase image, or a phase difference image, a temperature difference image, a temperature image, or an ablation image generated based on the phase image.

[0064] That is, the method provided in this embodiment can use the magnetic resonance images in the thermal ablation magnetic resonance image dataset to train a magnetic resonance image reconstruction model. This model can supplement the reconstruction of the downsampled magnetic resonance images during the subsequent thermal ablation process, thereby improving data acquisition efficiency while also improving data quality. Conventional magnetic resonance images include amplitude maps and phase maps. The change in phase is linearly related to the change in temperature. Therefore, a temperature difference map can be obtained by transforming the phase difference map. The temperature difference map can be combined with the base temperature to generate a temperature map. The temperature map can indicate the temperature state of the tissue at a specific section at a certain moment. Furthermore, based on the cumulative effect of temperature over time, the ablation state of the tissue can be determined, that is, an ablation map can be generated based on the temperature map. In actual application, the downsampled magnetic resonance amplitude image can be input into the magnetic resonance image reconstruction model to obtain a supplemented magnetic resonance amplitude map, or the downsampled magnetic resonance phase map can be input into the magnetic resonance image reconstruction model to obtain a supplemented magnetic resonance phase map. Then, based on the supplemented phase map, a phase difference map, a temperature difference map, a temperature map, an ablation map, etc. can be generated. Of course, the magnetic resonance images in the thermal ablation magnetic resonance image dataset can also be directly phase images, phase difference images, temperature difference images, temperature images or ablation images. Training samples are generated by downsampling for training deep learning models. The trained model can directly supplement the missing data in the "downsampled phase image, phase difference image, temperature difference image, temperature image or ablation image" to reconstruct the complete image.

[0065] Based on the previous embodiment, in one embodiment, the thermal ablation magnetic resonance image dataset includes magnetic resonance images acquired through a reproducible thermal ablation simulation test.

[0066] Specifically, thermal ablation simulation tests simulate the external environment, biological tissue, and thermal ablation procedures during the thermal ablation process. Reproducible thermal ablation simulation tests replicate the external environment, biological tissue model, and thermal ablation procedures during the same thermal ablation process. Traditional thermal ablation tests only perform a single thermal ablation to verify the ablation effect and do not replicate the thermal ablation process. Furthermore, traditional thermal ablation tests use live animals or ex vivo tissue as test subjects. Due to potential differences between individual animals, their reproducibility is poor. Screening animals by species, age, and other factors can only minimize individual differences. This method is specifically designed for thermal ablation operations, a thermal ablation simulation test with "specific external environment, biological tissue model to be ablated, and ablation operation". The thermal ablation simulation test 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 repeating the thermal ablation simulation test multiple times, we have enough time to acquire magnetic resonance images, and by selecting the image acquisition position and the image acquisition time point, the images of multiple tests have a corresponding relationship and can be used to generate training samples. By changing the external environment, the biological tissue model to be ablated, and the ablation operation, multiple groups of thermal ablation simulation tests are formed, and more training samples are collected, which makes up for the problem that clinical data is difficult to obtain and difficult to generate sufficient training samples.

[0067] The aforementioned external environment, for example, includes ambient temperature and humidity. For example, referring to the standard temperature of 18° to 24° and the standard humidity of 40% to 60% within the MRI room, the same ambient temperature of 22° and humidity of 50% were used in the replicated thermal ablation simulation test. A biological tissue model is a model of the tissue structure of the target for thermal ablation. For example, in laser ablation scenarios, an agar block is 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 then introduced into the agar block using optical fibers, and the thermal effect of the laser is used to ablate specific areas of the agar block. Another example is using gelatin and agar as the primary materials, with sodium chloride and other materials added to adjust the conductivity, to simulate brain tissue. Silicone tubes are used to simulate blood vessels, and methacrylated gelatin is used to simulate tumors. A radiofrequency ablation needle is used to apply radiofrequency current to heat specific areas of the "brain model." Another example is using an agar block to simulate brain tissue for ultrasound ablation scenarios. The agar block's porosity is adjusted to adjust the absorption and reflectivity of ultrasound. For example, in the laser ablation scenario, the same thermal ablation operation is applied. In multiple experiments, optical fibers of the same specifications are implanted in the same location in the "simulated thermal ablation environment," and laser energy is output with the same time-power curve. Another example is ultrasound ablation, where an ultrasound transducer array is placed in the same location and ultrasound energy is output with the same time-power curve at each time period.

[0068] For example, in the process of acquiring magnetic resonance images through the above-mentioned replication test, in the first test, the magnetic resonance device acquired magnetic resonance images of layers 1, 4, and 7 in the time period of 10 to 13 seconds. In the second test, the magnetic resonance device acquired magnetic resonance images of layers 2, 5, and 8 in the time period of 10 to 13 seconds. In the third test, the magnetic resonance device acquired magnetic resonance images of layers 3, 6, and 9 in the time period of 10 to 13 seconds. The magnetic resonance images acquired in the above three tests constitute the magnetic resonance images of consecutive slices 1 to 9. The above example describes the process of acquiring a set of magnetic resonance images of consecutive slices in a certain time period. Similarly, more sets of magnetic resonance images of consecutive slices can be acquired in other ablation time periods.

[0069] Another example of the above-mentioned process of obtaining magnetic resonance images through a replication test is that in a first test, images of the 1st, 3rd, 5th, and 7th columns within a certain slice are acquired with high quality, and in a second test, images of the 2nd, 3rd, 6th, and 8th columns within a certain slice are acquired with high quality, and the image data of the two replication tests are combined to obtain high-quality complete images of columns 1 to 8 within the slice.

[0070] Similarly, more magnetic resonance images can be collected during other ablation time periods and at other slice positions to generate training samples. Furthermore, a reproduction experiment can be conducted on another "thermal ablation process" with different external environments, biological tissue models, and thermal ablation operations to collect more magnetic resonance images to generate more training samples. This enriches the training data and improves the generalization ability of the trained magnetic resonance image reconstruction model.

[0071] This embodiment collects magnetic resonance image data through a reproducible thermal ablation simulation test, effectively making up for the problem of insufficient and difficult to obtain real thermal ablation data, and especially solves the problem of difficulty in collecting large-scale and high-quality magnetic resonance image data within a limited time.

[0072] Furthermore, in the case of supplementing missing inter-layer data, considering that it takes a certain amount of time to acquire magnetic resonance images during the actual thermal ablation process, there is a certain time difference between the acquisition times of each layer image within a sampling cycle. In other words, the acquisition time points of the images at each fault layer acquired within a cycle are not completely consistent, and the temperature field may have slightly changed during this time difference. Therefore, preferably, when acquiring the thermal ablation magnetic resonance image dataset through the thermal ablation simulation test, the image acquisition time at each fault layer in the thermal ablation simulation test is designed with reference to the time difference in magnetic resonance image acquisition during the actual thermal ablation process.

[0073] For example, in the actual thermal ablation process, the image at the 2nd layer is collected at 1s, the image at the 4th layer is collected at 2s, and the image at the 6th layer is collected at 3s. Referring to the time difference of this process, by reproducing the thermal ablation simulation test, the image at the 1st layer is collected at 0.5s, the image at the 2nd layer is collected at 1s, the image at the 3rd layer is collected at 1.5s, the image at the 4th layer is collected at 2s, the image at the 5th layer is collected at 2.5s, the image at the 6th layer is collected at 3s, and the image at the 7th layer is collected at 3.5s.

[0074] Based on any of the embodiments, in one embodiment, the thermal ablation magnetic resonance image dataset includes magnetic resonance images acquired through a reproducible thermal ablation simulation test and magnetic resonance images acquired from historical clinical thermal ablation data, and S13 includes:

[0075] The initial deep learning model is pre-trained using the training samples corresponding to the thermal ablation simulation test, and the training samples corresponding to the historical clinical thermal ablation data are used for secondary training.

[0076] Specifically, compared with the previous embodiment, the thermal ablation magnetic resonance image data set of this embodiment includes not only magnetic resonance images obtained in the thermal ablation simulation test, but also magnetic resonance images collected and saved in the past clinical thermal ablation process. Magnetic resonance images are obtained through thermal ablation simulation tests and downsampled to generate pre-training samples for pre-training the initial deep learning model. It is understandable that since the thermal simulation test is reproducible, it is possible to collect magnetic resonance image data of more sections and shorter time intervals of the "same thermal ablation process" through repeated tests. These data can be easily obtained and used for the initial training of the initial deep learning model, and the preliminary optimization and adjustment of the model parameters so that the model has the basis for transfer learning. After that, training samples can be generated based on the magnetic resonance images in the historical clinical thermal ablation data, and the model can be trained for the second time to further improve the performance of the model. It is understandable that historical clinical thermal ablation data is scarce.

[0077] This embodiment collects magnetic resonance image data through thermal ablation simulation tests, effectively making up for the problem of insufficient and difficult to obtain clinical thermal ablation data, especially solving the problem of difficulty in collecting large-scale and high-quality magnetic resonance image data within a limited time. The performance of the model is improved through pre-training based on thermal ablation simulation test data and secondary training based on clinical thermal ablation data.

[0078] Based on any one of the embodiments, in one embodiment, the deep learning model is selected from any one of the following: convolutional neural network (CNN), recurrent neural network (RNN), and generative adversarial network (GAN).

[0079] A convolutional neural network (CNN) is a type of feedforward neural network with a deep structure that incorporates convolutional computations. It possesses representational learning capabilities and can process input information according to its hierarchical structure. A recurrent neural network (RNN) is a type of recursive neural network that takes sequential data as input, performs recursive operations in the direction of the sequence's evolution, and has all nodes connected in a chain-like fashion. A generative adversarial network (GAN) consists 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 determines whether the input is real data or a generated sample. During training, one of the two models is fixed while the parameters of the other are updated, alternating iterations. Ultimately, the generative model can estimate the distribution of the sample data.

[0080] A preferred embodiment of the present invention is described below:

[0081] Step 1: Acquire high-quality magnetic resonance images by reproducing thermal ablation simulation experiments and add them to the thermal ablation magnetic resonance image dataset.

[0082] Specifically, for the same thermal ablation process, by reproducing the external environment, biological tissue model and thermal ablation operation, the reproduced thermal ablation process allows the magnetic resonance equipment to have sufficient time to acquire magnetic resonance images. The continuous tomographic magnetic resonance images acquired within the same time period of the same thermal ablation process can constitute a set of data for generating training samples.

[0083] In the first test, high-quality images of the 1st, 4th, and 7th columns within a certain fault were collected. In the second test, high-quality images of the 2nd, 5th, and 8th columns within a certain fault were collected. In the third test, high-quality images of the 3rd, 6th, and 9th columns within a certain fault were collected. The image data of the three repeated tests were combined to obtain high-quality complete images of columns 1 to 9 within the layer.

[0084] Step 2: For the magnetic resonance images in the thermal ablation magnetic resonance image dataset, generate training samples by intra-layer downsampling and add them to the training sample set.

[0085] Specifically, the downsampling process can be interval sampling, for example, row interval sampling, column interval sampling, adjacent interval sampling (that is, the four pixels adjacent to a sampled pixel in the orthogonal direction are no longer sampled); the downsampling process can also be random downsampling, with a sampling ratio of 50%, 40%, 30%, 20%, etc.; the downsampling can also be Gaussian distribution downsampling, for example, the closer to the center of the image, the higher the sampling ratio, the closer to the edge of the image, the lower the sampling ratio, and the sampling ratio is Gaussian distributed as a whole. The downsampled image is used as the sample, and the original image or the data eliminated by downsampling is used as the corresponding labeled data to form a joint training sample, which is added to the training sample set.

[0086] Step 3: Train the deep learning model based on the training sample set to obtain a magnetic resonance image reconstruction model.

[0087] Specifically, the deep learning model was trained using the training samples, and the model parameters were optimized to identify the best-performing model. This model was then used to reconstruct magnetic resonance images during subsequent thermal ablation procedures. Because the training samples in this set cover the entire thermal ablation process, the trained model is capable of recovering and reconstructing downsampled data from any stage of a real thermal ablation procedure, demonstrating strong generalization capabilities.

[0088] This embodiment acquires magnetic resonance image data through reproducible thermal ablation simulations, allowing the MRI device ample time to capture images of the same thermal ablation process across multiple trials. This effectively addresses the inadequacy and difficulty in acquiring real thermal ablation data, particularly the difficulty in acquiring high-quality magnetic resonance images within a limited timeframe and the scarcity of training data. The method of this embodiment can train a deep learning model to perform high-precision reconstruction of inter-slice downsampled magnetic resonance images, improving the real-time performance of temperature monitoring through inter-slice downsampling without significantly degrading image quality.

[0089] A magnetic resonance image reconstruction method provided by the present invention is described below. The magnetic resonance image reconstruction method described below and the training method of the magnetic resonance image reconstruction model described above can refer to each other.

[0090] Figure 2 FIG. 1 is a flow chart of a magnetic resonance image reconstruction method provided by the present invention, such as Figure 2 As shown, the method includes:

[0091] S21, obtaining a downsampled magnetic resonance image of the current patient;

[0092] S22, inputting the downsampled magnetic resonance image into a magnetic resonance image reconstruction model to obtain a reconstructed magnetic resonance image to provide information support for the thermal ablation process;

[0093] The magnetic resonance image reconstruction model is pre-trained according to any of the aforementioned training methods for magnetic resonance image reconstruction models.

[0094] Specifically, acquiring the patient's magnetic resonance images through downsampling during the thermal ablation process can speed up acquisition efficiency and improve the real-time performance of magnetic resonance monitoring. The downsampled images acquired from the patient are input into a magnetic resonance image reconstruction model pre-trained using the aforementioned magnetic resonance image reconstruction model training method. The model can output reconstructed and supplemented magnetic resonance images to provide information support for the thermal ablation process. It is understood that the downsampling method needs to be consistent with the downsampling strategy of the training samples used during the model training phase.

[0095] In this embodiment, during the model training stage, training samples are generated by downsampling based on the thermal ablation magnetic resonance image dataset for training the deep learning model. The dataset covers the entire process and all positions of thermal ablation. The deep learning model obtained through training is more in line with the thermal ablation scenario. During the model application stage, the downsampled magnetic resonance images are input into the trained deep learning model to obtain reconstructed and supplemented magnetic resonance images, which reduces the amount of data collected by the magnetic resonance equipment during the thermal ablation process and improves the real-time monitoring while ensuring data quality.

[0096] Furthermore, in the above embodiments, what is reconstructed can be a magnetic resonance amplitude map or a magnetic resonance phase map, and the change in phase is linearly related to the change in temperature. Therefore, a temperature difference map can be obtained by transformation based on the phase difference map, and the temperature difference map can be combined with the base temperature to generate a temperature map. The temperature map can indicate the temperature state of the tissue at a specific fault at a certain moment. Furthermore, based on the cumulative effect of temperature over time, the ablation state of the tissue can be judged, that is, an ablation map can be generated based on the temperature map. In other embodiments, what is reconstructed can also be a phase map, a phase difference map, a temperature difference map, a temperature map, or an ablation map.

[0097] A magnetic resonance temperature imaging method provided by the present invention is described below. The magnetic resonance temperature imaging method described below and the training method of the magnetic resonance image reconstruction model described above can be referenced to each other.

[0098] Figure 3 FIG. 1 is a flow chart of a magnetic resonance temperature imaging method provided by the present invention, such as Figure 3 As shown, the method includes:

[0099] S31, obtaining a downsampled magnetic resonance temperature map of the current patient;

[0100] S32, inputting the downsampled magnetic resonance temperature map into a magnetic resonance image reconstruction model to obtain a reconstructed magnetic resonance temperature map to provide information support for the thermal ablation process;

[0101] The magnetic resonance image reconstruction model is pre-trained according to any of the aforementioned training methods for magnetic resonance image reconstruction models.

[0102] Specifically, during the thermal ablation process, collecting the patient's magnetic resonance temperature map by downsampling can speed up the acquisition efficiency and improve the real-time performance of magnetic resonance temperature monitoring. The downsampled magnetic resonance temperature map is input into the magnetic resonance image reconstruction model pre-trained based on the training method of the aforementioned magnetic resonance image reconstruction model. The model can output a temperature map to provide information support for the thermal ablation process. It is understandable that the downsampling method needs to be consistent with the downsampling strategy of the training samples used in the model training phase.

[0103] In the model training stage, this embodiment generates training samples for training the deep learning model by downsampling based on the thermal ablation magnetic resonance image dataset. The dataset covers the entire process and all positions of thermal ablation. The deep learning model obtained through training is more in line with the thermal ablation scenario. In the model application stage, the magnetic resonance temperature map after time downsampling is input into the trained deep learning model to obtain a reconstructed and supplemented temperature map, which reduces the data collection amount of the magnetic resonance equipment during the thermal ablation process, and improves the real-time monitoring while ensuring data quality.

[0104] The present invention also provides a magnetic resonance image reconstruction system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of any one of the aforementioned methods for training a magnetic resonance image reconstruction model are implemented.

[0105] The present invention also provides a magnetic resonance thermal ablation monitoring system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the aforementioned magnetic resonance image reconstruction methods when executing the program.

[0106] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0107] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, or of course, by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or certain parts of the embodiments.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A training method for a magnetic resonance image reconstruction model, characterized in that: include: Acquiring a thermal ablation magnetic resonance image dataset; wherein the thermal ablation magnetic resonance image dataset includes magnetic resonance images during the thermal ablation process; generating a training sample set by downsampling the magnetic resonance images in the thermal ablation magnetic resonance image dataset; A deep learning model is trained according to the training sample set to obtain a magnetic resonance image reconstruction model.

2. The training method for a magnetic resonance image reconstruction model according to claim 1, wherein: The thermal ablation magnetic resonance image dataset includes several sets of magnetic resonance images at consecutive slices during the thermal ablation process.

3. The training method for a magnetic resonance image reconstruction model according to claim 2, wherein: The generating a training sample set by downsampling the magnetic resonance images in the thermal ablation magnetic resonance image dataset comprises: Inter-layer downsampling is performed on a set of magnetic resonance images of consecutive slices, and the magnetic resonance images of several adjacent slices after downsampling are used as samples. The missing slices near the several adjacent slices after downsampling are used as corresponding annotation data to form training samples, which are added to the training sample set.

4. The training method for a magnetic resonance image reconstruction model according to claim 1, wherein: The thermal ablation magnetic resonance image dataset includes magnetic resonance images of several slices during the thermal ablation process, and generating a training sample set by downsampling the magnetic resonance images in the thermal ablation magnetic resonance image dataset includes: The magnetic resonance images at each slice are subjected to intra-slice downsampling, and the magnetic resonance images after intra-slice downsampling are used as samples. The data eliminated during the intra-slice downsampling process are used as corresponding labeled data to form training samples, which are added to the training sample set.

5. The training method for a magnetic resonance image reconstruction model according to claim 4, wherein: The intra-layer downsampling is intra-layer interval downsampling, or intra-layer random downsampling, or Gaussian distribution sampling.

6. The training method for a magnetic resonance image reconstruction model according to claim 1, wherein: The magnetic resonance image in the thermal ablation magnetic resonance image dataset is an amplitude image, a phase image, or a phase difference image, a temperature difference image, a temperature image or an ablation image generated based on the phase image.

7. The training method according to claim 1, characterized in that The thermal ablation magnetic resonance image dataset includes magnetic resonance images acquired through a reproducible thermal ablation simulation test.

8. The method for training a magnetic resonance image reconstruction model according to claim 1, wherein: The thermal ablation magnetic resonance image dataset includes magnetic resonance images obtained through reproducible thermal ablation simulation experiments and magnetic resonance images obtained from historical clinical thermal ablation data. The deep learning model is trained based on the training sample set to obtain a magnetic resonance image reconstruction model, which includes: The initial deep learning model is pre-trained using the training samples corresponding to the thermal ablation simulation test, and secondary training is performed using the training samples corresponding to the historical clinical thermal ablation data to obtain the deep learning model.

9. The method for training a magnetic resonance image reconstruction model according to claim 1, wherein: The deep learning model is selected from any one of the following: convolutional neural network, recurrent neural network, and generative adversarial network.

10. A magnetic resonance image reconstruction method, characterized in that: include: Obtaining a downsampled magnetic resonance image of the current patient; Inputting the downsampled magnetic resonance image into a magnetic resonance image reconstruction model to obtain a reconstructed magnetic resonance image to provide information support for the thermal ablation process; The magnetic resonance image reconstruction model is pre-trained according to the training method for a magnetic resonance image reconstruction model according to any one of claims 1 to 9.

11. A magnetic resonance-guided thermal ablation monitoring system, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the method for training a magnetic resonance image reconstruction model according to any one of claims 1 to 9 are implemented, or the steps of the method for magnetic resonance image reconstruction according to claim 10 are implemented.

Images