Learning device, learning method, and learning program

Abstract

To reduce the training time of the model. [Solution] The learning device according to this embodiment includes an acquisition unit and a training unit. The acquisition unit has a first resolution and acquires first training data based on target data. The training unit trains a first machine learning model using the first training data and generates a first trained model. The training unit has a second resolution with a higher resolution than the first resolution and trains a second machine learning model using second training data based on the target data and the first trained model, and generates a second trained model.

Description

Technical Field

[0001] The embodiments disclosed in this specification and the drawings relate to a learning device, a learning method, and a learning program.

Background Art

[0002] With the development of machine learning, the application of machine learning models has also advanced in the medical field. As methods for training machine learning models, generally, supervised learning and unsupervised learning (including self-supervised learning) can be mentioned. In supervised learning, if there is a large amount of training data, it can be trained with high performance, and it can be said that inference is also fast. However, there are cases where sufficient training data cannot be collected. In addition, so-called hallucination, in which incorrect information is output as if it were correct information depending on the situation, is likely to occur. Therefore, in image reconstruction processing, it is realistic to train a machine learning model by unsupervised learning such as self-supervised learning. However, since training is performed in real time when using the machine learning model, there is a problem that the reconstruction process takes time.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Non-Patent Documents

[0004]

Non-Patent Document 1

Non-Patent Document 2

[0005] One of the problems that the embodiments disclosed herein and in the drawings aim to solve is to reduce the training time of a model. However, the problems that the embodiments disclosed herein and in the drawings aim to solve are not limited to the above problem. Problems corresponding to the effects of each configuration shown in the embodiments described later can also be positioned as other problems. [Means for solving the problem]

[0006] The learning device according to this embodiment includes an acquisition unit and a training unit. The acquisition unit has a first resolution and acquires a first training dataset based on target data. The training unit trains a first machine learning model using the first training dataset and generates a first trained model. The training unit has a second resolution with a higher resolution than the first resolution and trains a second machine learning model using a second training dataset based on the target data and the first trained model, generating a second trained model. [Brief explanation of the drawing]

[0007] [Figure 1] Figure 1 is a block diagram showing a learning device according to this embodiment. [Figure 2] Figure 2 is a conceptual diagram illustrating the differences in resolution within the training dataset. [Figure 3]Figure 3 is a flowchart showing an example of the operation of the learning device according to this embodiment. [Figure 4] Figure 4 is a conceptual diagram illustrating a first example of a machine learning model configuration according to this embodiment. [Figure 5] Figure 5 is a conceptual diagram illustrating a first example of a machine learning model configuration according to this embodiment. [Figure 6] Figure 6 is a conceptual diagram illustrating a first example of a machine learning model configuration according to this embodiment. [Figure 7] Figure 7 is a conceptual diagram illustrating a first example of a machine learning model configuration according to this embodiment. [Figure 8] Figure 8 is a conceptual diagram illustrating a second example of the machine learning model configuration according to this embodiment. [Figure 9] Figure 9 is a conceptual diagram illustrating a second example of the machine learning model configuration according to this embodiment. [Figure 10] Figure 10 shows an example of using a trained model for iterative image reconstruction. [Figure 11] Figure 11 is a list of examples of tasks to which the trained model can be applied. [Modes for carrying out the invention]

[0008] The learning device, learning method, and learning program according to this embodiment will be described below with reference to the drawings. In the following embodiments, parts with the same reference numerals perform similar operations, and redundant explanations will be omitted as appropriate. One embodiment will be described below with reference to the drawings.

[0009] The learning device according to this embodiment will be described with reference to the block diagram in Figure 1. The learning device 10 according to this embodiment includes a processing circuit 11, a memory 12, an input interface 13, and a communication interface 14.

[0010] Note that the learning device 10 according to this embodiment may be included in a console, a workstation, etc., or may be included in a medical image diagnostic device such as an MRI (Magnetic Resonance Imaging) device or an X-ray CT (Computed Tomography) device. Alternatively, it may be stored in a server instead of a local terminal.

[0011] The processing circuit 11 includes an acquisition function 111, a conversion function 112, a model design function 113, a training function 114, and an output function 115. The processing circuit 11 has a processor not shown as a hardware resource.

[0012] The acquisition function 111 acquires a plurality of target data. The plurality of target data are, for example, a plurality of medical data. Examples of the type of medical data include k-space data collected by an MRI device, MR images, projection data collected by an X-ray CT device, sinogram data, or CT images. Also, the acquisition function 111 acquires a training data set converted from the plurality of target data by the conversion function 112. Note that the type of medical data is not limited to data collected by an MRI device or an X-ray CT device, and may be data collected by other image diagnostic devices such as an ultrasonic diagnostic device.

[0013] The conversion function 112 converts each of the plurality of target data into training data sets with different resolutions. When one training data set is k-space data, it includes a plurality of partial data corresponding to a plurality of resolutions. That is, when it is k-space data, a plurality of training data sets with different resolutions can be easily generated.

[0014] The model design function 113 designs the structure of a machine learning model to be trained using the training data set for each resolution.

[0015] The training function 114 trains the machine learning model designed by the model design function 113 using the training dataset. The machine learning model is a neural network composed of, for example, a deep neural network such as a deep convolutional neural network. The machine learning model may be a U-net, which is an example of an encoder-decoder model, a ResNet including residual blocks, or a DenseNet including Dense blocks. Furthermore, it may be formed by a neural network of the MoDL type or unroll type disclosed in Non-Patent Document 1, or may be formed by a recurrent neural network such as LSTM (Long Short Term Memory) or GRU (Gated Recurrent Unit). When the training of the machine learning model is completed, a learned model is generated. The training function 114 trains a plurality of machine learning models having corresponding structures for training datasets having a plurality of different resolutions, and generates a learned model. The output function 115 outputs information regarding the training status of the machine learning model, the learned model, etc. to the outside.

[0016] In addition, various functions in the processing circuit 11 may be stored in the memory 12 in the form of a program executable by a computer. In this case, it can also be said that the processing circuit 11 is a processor that realizes the functions corresponding to each program by reading and executing the programs corresponding to these various functions from the memory 12. In other words, the processing circuit 11 in the state of having read each program will have a plurality of functions etc. shown in the processing circuit 11 of FIG. 1.

[0017] In Figure 1, these various functions are explained as being realized by a single processing circuit 11, but it is also possible to configure the processing circuit 11 by combining multiple independent processors, and each processor realizes the functions by executing a program. In other words, each of the above-mentioned functions may be configured as a program, and one processing circuit may execute each program, or a specific function may be implemented in a dedicated, independent program execution circuit.

[0018] Memory 12 stores various types of data, such as target data, training datasets for different resolutions, and trained models. Memory 12 can be a semiconductor memory element such as RAM (Random Access Memory) or flash memory, a hard disk drive (HDD), a solid state drive (SSD), or an optical disc. Memory 12 may also be a drive device that reads and writes various information to and from portable storage media such as a CD-ROM drive, DVD drive, or flash memory.

[0019] The input interface 13 has circuits that receive various instructions and information inputs from the user. The input interface 13 has circuits related to pointing devices such as a mouse, or input devices such as a keyboard. However, the circuits of the input interface 13 are not limited to circuits related to physical operating components such as a mouse or keyboard. For example, the input interface 13 may have electrical signal processing circuits that receive electrical signals corresponding to input operations from an external input device provided separately from the learning device 10, and output the received electrical signals to various circuits within the learning device 10.

[0020] The communication interface 14 exchanges data with external devices via wired or wireless connection. The communication method and interface structure are standard communication methods and therefore will not be explained here.

[0021] Next, we will explain the differences in resolution in the training dataset by referring to the conceptual diagram in Figure 2. Figure 2 is a conceptual diagram showing the k-space in MR data acquisition. In this embodiment, resolution is assumed to be spatial resolution. Data corresponding to a part of the k-space becomes partial data, and the larger the area of ​​partial data, the higher the resolution of the partial data. The processing circuit 11 generates training datasets of different resolutions by extracting (cutting out) partial data of different resolutions from the target data using the conversion function 112.

[0022] Specifically, in the example shown in Figure 2, if the target data (not shown) is k-space data collected over the entire k-space, then k-space partial data 21 includes the central part of k-space and is the k-space data with the lowest resolution. Partial data 22, which covers a larger area (data size) than partial data 21, includes the central part of k-space and the area outside of partial data 21, and is k-space data with a higher resolution than partial data 21. Partial data 23, which covers a larger area (data size) than partial data 22, includes the central part of k-space and the area outside of partial data 22, and is k-space data with a higher resolution than partial data 21 and partial data 22. The resolution stages are assumed to be set in descending order of resolution, for example, the target data corresponding to the entire k-space, and then 1 / 2, 1 / 4, 1 / 8, and 1 / 16 of that target data, which are powers of 2, but are not limited to this, and any stages with different resolutions are acceptable. Furthermore, when using k-space data as the target data, since image processing is anticipated, it is assumed that partial data including the central part of k-space will be extracted as partial data regardless of resolution, in order to include low-frequency components. Furthermore, when using images such as MR images as target data, multiple training datasets with different resolutions can be generated by changing the resolution of the images through image processing. For image processing to convert the resolution, the processing circuit 11, using the conversion function 112, can generate multiple images with different resolutions by using at least one of the following: for example, a process that degrades the resolution by compressing the original image, or a process that improves the resolution by super-resolution, and these can be used as training datasets.

[0023] Next, an example of the operation of the learning device 10 according to this embodiment will be described with reference to the flowchart in Figure 3.

[0024] In step SA1, the processing circuit 11 acquires multiple target data using the acquisition function 111. In step SA2, the processing circuit 11 uses the conversion function 112 to convert multiple target data into multiple partial data with lower resolution compared to the target data, thereby generating a training dataset. For example, in the first processing step, one-eighth the size of the target data is extracted to generate the training dataset. The number of resolution levels at which the training dataset is generated may be predetermined, or it may be determined according to the type of target data and the task of the machine learning model. In the first conversion process, a portion of the data converted at the lowest resolution among the pre-set resolutions is generated as a training dataset.

[0025] In step SA3, the processing circuit 11 designs a machine learning model using the model design function 113. Specifically, it designs the machine learning model based on information such as what architecture to use and which parts of the overall architecture to train. For subsequent processing iterations, the model design function 113 designs a new machine learning model that includes the machine learning model that has been trained in previous iterations.

[0026] In step SA4, the processing circuit 11 uses the training function 114 to train a machine learning model using the training dataset and generate a trained model. In this embodiment, the training method is assumed to be self-supervised learning in which the input training dataset is images and the model is trained to reproduce the input images, but it may also be unsupervised learning or supervised learning. Whether or not the training of the machine learning model is complete can be determined, for example, when a predetermined number of epochs of training have been completed, or when the loss value related to the loss function is below a threshold. If training is complete, parameter updates are stopped and a trained model is generated. On the other hand, if training is not complete, the same process is repeated until, for example, the above-mentioned determination of training completion is made.

[0027] In step SA5, the processing circuit 11 uses the training function 114 to determine whether the machine learning model has been trained with training data having a predetermined resolution. Here, it is determined whether the training was conducted using the training dataset with the highest resolution among several pre-set resolution levels. For example, if the model was trained using full-size target data, it may be determined that the machine learning model was trained with a training dataset having a predetermined resolution. Alternatively, if, for example, half the size of the target data, i.e., half the resolution, is set as the maximum resolution, then if the model was trained using a training dataset with half the resolution, it may be determined that the machine learning model was trained with a training dataset having a predetermined resolution. If the machine learning model has been trained with training data of the specified resolution, proceed to step SA7; otherwise, proceed to step SA6.

[0028] In step SA6, the processing circuit 11 uses the conversion function 112 to generate a training dataset with a higher resolution than the training dataset from the previous training session. For example, if the first processing generates a training dataset with one-eighth the resolution of the target data, the second processing should generate a training dataset with one-quarter the resolution of the target data.

[0029] After that, the process returns to step SA3 and the same process is repeated. In the second and subsequent steps of step SA3, the model design function 113 causes the processing circuit 11 to design the configuration of a new machine learning model so that it includes at least a part of the configuration of the trained model. For example, if a task-specific MLP (Multi-Layer Perceptron) head was connected as the output layer of the machine learning model, the configuration of the trained model other than the MLP head may be included in the new machine learning model. In the second and subsequent steps SA4, the processing circuit 11, using the training function 114, performs transfer learning on a new machine learning model based on the parameters of the trained model, using a training dataset that is based on the same target data but with a higher resolution than the previous processing. Specifically, for example, in the case of a hierarchical multiscale model like U-net, the new machine learning model should be trained with a block containing fixed parameters of the trained model included in the new machine learning model.

[0030] In step SA7, the processing circuit 11 outputs a trained model that has been trained to a predetermined resolution via the output function 115.

[0031] Next, a first example of a machine learning model configuration to be trained with the learning device 10 according to this embodiment will be described with reference to Figures 4 to 7. In the first example, the training method assumes a U-net. In a U-net, downsampling is performed in the encoder unit and upsampling is performed in the decoder unit. Furthermore, each layer has a skip connection that links the feature map of the encoder unit with the feature map of the upsampling unit. Here, we assume that the target data will be prepared as training datasets at four different resolutions: 1 / 8, 1 / 4, 1 / 2, and full size.

[0032] First, regarding the machine learning model shown in Figure 4, the machine learning model 41 is trained using a training dataset with the lowest resolution set for the target data, which is one-eighth the resolution of the target data, and the trained model 42 is obtained. Here, as an example, the machine learning model 41 is set to have three layers of convolutional blocks connected to perform convolutional processing, but it is not limited to convolutional blocks; any model architecture is acceptable as long as it is possible to stack or extend the network.

[0033] Next, as shown in Figure 5, the processing circuit 11 uses the model design function 113 to design the trained model 42, which has been trained, as the lowest-level feature extraction block in the U-net. It then designs a two-stage U-net as a new machine learning model 53, with an encoder unit 51 and a decoder unit 52, each having skip connections, as feature extraction blocks in the layers above the lowest-level feature extraction block. In other words, the trained model is transferred to the new machine learning model through transfer learning. Here, the two-stage U-net is trained using a training dataset with the second lowest resolution, which has one-quarter the resolution of the target data, and a trained model 54 is obtained.

[0034] Next, as shown in Figure 6, the processing circuit 11 uses the model design function 113 to design a new machine learning model 63, which is a three-stage U-net with feature extraction blocks for the encoder unit 61 and the decoder unit 62 set on top of the trained model 54, which is a two-stage U-net that has already been trained. Here, the three-stage U-net is trained using a training dataset with half the resolution of the target data, which is the third lowest resolution, and a trained model 64 is obtained.

[0035] Finally, as shown in Figure 7, the processing circuit 11 uses the model design function 113 to design a new machine learning model 73 by setting the encoder unit 71 and the decoder unit 72 feature extraction blocks, respectively, on the upper stage of the three-stage U-net that has been trained. Here, the four-stage U-net is trained using full-size target data with the highest resolution as the training dataset, and the final trained model 74 is obtained.

[0036] The parameters (weights, biases, etc.) of the lower-level feature extraction blocks that have already been trained may be fixed, or they may be used as initial values ​​for those feature extraction blocks when training a new machine learning model. Furthermore, in this case, we used 1 / 8th size data as the lowest resolution data for the target data, resulting in a 4-layer U-net. However, the model should be designed and trained so that the number of U-net layers increases according to the resolution level set for the target data.

[0037] In this way, a model is trained on a small scale (architecture), and the training results from that scale are transferred to the model of the next scale for training. In other words, the training results of the model are transferred from a small scale to a large scale, and the model is trained stepwise for the desired model size. This can speed up model training.

[0038] Next, a second example of a training method for the learning device 10 according to this embodiment will be described with reference to Figures 8 and 9.

[0039] Figure 8 shows a ResNet (Residual Network), which assumes a configuration in which residual blocks are connected in series in multiple layers. In Figure 8, a model with one residual block 81 is trained using the training dataset with the lowest resolution set for the target data. Here, the model with residual block 81 is trained using a training dataset with one-eighth the resolution of the target data, and a trained model 82 is generated.

[0040] In Figure 9, a machine learning model 92 is designed by connecting a new residual block 91 to the trained model 82, and the machine learning model 92 is trained using a training dataset with the second lowest resolution. Here, the machine learning model 92 is trained using a training dataset with the second lowest resolution of one-quarter. After that, depending on the resolution level, residual blocks are connected in series to the subsequent stages of the trained model after training is complete to design the configuration of a new machine learning model, and then training of this new machine learning model can be carried out. The handling of parameters of the lower-level feature extraction blocks that have already been trained is the same as in the first example of the training method.

[0041] In the examples shown in Figures 8 and 9, residual blocks are connected in series to the later stage, but it is also possible to connect them in series one after another to the earlier stage of residual blocks that have finished training. That is, for example, residual blocks trained using a training dataset with one-eighth the resolution of the target data may be connected in series to the earlier stage of residual blocks trained using a training dataset with one-quarter the resolution of the target data.

[0042] The trained model obtained using the learning device according to this embodiment can be used, for example, in iterative image reconstruction. An example of using a pre-trained model for iterative image reconstruction is explained with reference to Figure 10.

[0043] Figure 10 is an example of a conceptual diagram showing iterative image reconstruction using processing blocks. Iterative image reconstruction is a method that repeatedly performs image reconstruction processing while maintaining data consistency between the target data and the output from the trained model, within an image reconstruction process that performs noise reduction using a neural network. In this embodiment, the target data is collected data, and when the target is an MRI device, it may be the MR signal itself collected by the MRI device (e.g., raw data), k-space data in which the collected MR signals are placed in k-space, data after signal processing with filters, etc., or MR images obtained by transforming k-space data.

[0044] Iterative image reconstruction can often be formulated as an image optimization problem. If the collected data is y, then, for example, the search process should be performed to find an image x that minimizes equation (1) below.

[0045]

number

[0046] A represents the transformation from image to k-space, and "Ax" transforms image x into k-space data. For A, for example, the Discrete Fourier Transform (DFT), Non-Uniform Discrete Fourier Transform (NUDFT), and each of these transformations can be multiplied from the right by a transformation corresponding to the sensitivity of each receiving coil in multicoil acquisition. By reducing the squared error between "Ax" and the acquired data y, if the acquired data y is k-space data, the error of the k-space data Ax on the k-space data y is reduced. λR(x) represents the error term on the output value of the noise reduction process. λ is a parameter related to the error term and is a value learned during the training of the machine learning model. For optimization methods of equation (1), for example, methods such as VSQP (Variable-splitting with the quadratic penalty) and ADMM (alternate directions method of multipliers) can be used, and a detailed explanation is omitted. In iterative image reconstruction, noise reduction is not essential. A trained model may be generated by performing an optimization process that increases the degree of agreement between the output of the machine learning model and the target data in the training dataset.

[0047] Here, we will describe the case where the training method for a machine learning model using the learning device 10 according to this embodiment is applied to iterative image reconstruction. In the example in Figure 10, partial data at a scale of 1 / 8 of the collected data, that is, partial data having 1 / 8 the resolution compared to the full-size collected data, is used as the first training data. The processing circuit 11, using the training function 114, trains model 102, which is a machine learning model as shown in Figure 4, within the framework of iterative image reconstruction, using a first training dataset, which is a collection of first training data generated from multiple collected data.

[0048] Specifically, in the example shown in Figure 10, an inverse fast Fourier transform 101 (IFFT101) is performed on the collected data acquired by the MRI device, in this case k-space data, to generate an image. The generated image is input to model 102 to generate an image with noise reduction. A fast Fourier transform 103 (FFT103) is performed on the noise-reduced image to generate k-space data. An optimization process (i.e., data consistency processing, DC104 in Figure 10) is performed to improve the degree of agreement between the generated k-space data and the k-space data of the collected data. By performing an inverse fast Fourier transform 105 (IFFT105) on the result of the processing by DC104, a data-consistency processed image is generated. The data-consistency processed image is input to model 102 again, and the same process is repeated so that the image output from IFFT105 is optimized. When training of Model 102 is complete, a portion of the collected data at a 1 / 4 scale, which has a higher resolution than the 1 / 8 scale of the first training dataset, is used as the second training dataset. The training function 114 causes the processing circuit 11 to train Model 106, which includes Model 102 and is a machine learning model as shown in Figure 5, using the second training dataset, which is a collection of second training data generated from multiple collected data, in the same way as with the first training dataset, within the framework of the iterative image reconstruction described above. Subsequently, the training function 114 allows the processing circuit 11 to increase the resolution of the training dataset and enlarge the model configuration until a predetermined resolution is reached, thereby continuing to train the model.

[0049] Here, the noise subject to noise reduction processing includes, but is not particularly limited to, noise caused by static magnetic field inhomogeneity if the input data is MR data, as well as other artifacts such as noise caused by data acquisition, noise caused by signal processing on the MR signal, and noise caused by data processing on k-space data and spectra. Furthermore, if the input data is CT data, it includes, but is not particularly limited to, noise caused by data acquisition such as metal artifacts and low-count artifacts.

[0050] Although the explanation uses an iterative reconstruction process as an example, the processing of a trained model trained with the learning device according to this embodiment is not limited to this. Figure 11 shows a list of examples of tasks to which the trained model according to this embodiment can be applied.

[0051] The list in Figure 11 shows examples of the application of the trained model and its relationship to input and output. For example, it may be used for medical image segmentation. In the case of segmentation, both the input data and the ground truth data for the trained model should be trained as low-resolution (reduced data). Alternatively, it may be used for disease identification tasks. In the case of disease identification tasks, the model should be trained to produce images of increasing resolution, starting from low-resolution images (such as partial images), and then the final output should be produced. In other words, the trained model according to this embodiment can be applied to all kinds of processing that improves the accuracy, quality, or image quality of data.

[0052] Furthermore, beyond the examples shown in Figure 11, it can also be used for grading related to cancer progression stages, clinical decision support (CDS), and outputting surgical procedure recommendations.

[0053] According to the embodiment described above, the learning device trains a machine learning model using a first training dataset having a first resolution and generates a trained model. The learning device uses the same target data as the first training dataset as a reference and uses a second training dataset having a second resolution higher than the first resolution to train a new machine learning model, including the trained model, using so-called transfer learning and generating a new trained model. The learning device generates a trained model for the target data by performing the above-described transfer learning until a predetermined resolution is reached. In other words, by starting training at a small scale and sequentially transferring the training results from the completed scale to a larger scale, the convergence speed of learning can be improved and the training time of the model can be shortened compared to training the model at a large scale from the beginning.

[0054] In addition, each function according to the embodiment can also be realized by installing a program that performs the processing on a computer such as a workstation and loading it into memory. In this case, the program that can cause the computer to execute the method can also be stored and distributed on a storage medium such as a magnetic disk (hard disk, etc.), optical disk (CD-ROM, DVD, etc.), or semiconductor memory.

[0055] In the above explanation, the term "processor" refers to circuits such as CPUs (Central Processing Units), GPUs (Graphics Processing Units), Application Specific Integrated Circuits (ASICs), and Programmable Logic Devices (e.g., Simple Programmable Logic Devices (SPLDs), Complex Programmable Logic Devices (CPLDs), and Field Programmable Gate Arrays (FPGAs)).

[0056] While several embodiments have been described, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These embodiments can be implemented in a variety of other forms, and various omissions, substitutions, modifications, and combinations of embodiments can be made without departing from the spirit of the invention. These embodiments and their variations are included in the scope and spirit of the invention, as well as in the claims and their equivalents. [Explanation of Symbols]

[0057] 10 Learning device 11 Processing Circuit 12 memory 13 Input Interfaces 14. Communication Interface 21, 22, 23 Partial data 41, 53, 63, 73, 92 Machine Learning Models 42, 54, 64, 74, 82 Pre-trained models 51, 61, 71 Encoder section 52, 62, 72 Decoder section 81,91 Residual Blocks 101,105 Inverse Fast Fourier Transform Models 102, 106 103 Fast Fourier Transform 111 Acquisition function 112 Conversion function 113 Model Design Functions 114 Training Functions 115 Output function

Claims

1. An acquisition unit having a first resolution and acquiring a first training dataset based on target data, The system comprises a training unit that trains a first machine learning model using the first training dataset and generates a first trained model, The training unit has a second resolution that is higher than the first resolution, and trains a second machine learning model using a second training dataset based on the target data and the first trained model, thereby generating a second trained model. Learning device.

2. The second machine learning model includes at least a portion of the configuration of the first machine learning model, The learning device according to claim 1, wherein the training unit trains the second machine learning model based on the parameters of the first trained model.

3. The first machine learning model is part of U-net, which is an encoder-decoder model. The learning device according to claim 1, wherein the second machine learning model is a U-net including a convolutional layer of an encoder and decoder having skip connections in a higher layer than the first machine learning model.

4. The first machine learning model includes residual blocks, The learning apparatus according to claim 1, wherein the second machine learning model has a structure that includes a new residual block connected in series with the first machine learning model.

5. The learning device according to claim 1, further comprising a conversion unit that converts the target data into a first learning dataset with a first resolution and / or a second learning dataset with a second resolution.

6. The aforementioned target data is k-space data, The first training dataset is data that includes the central portion of the k-space of the target data, The learning device according to claim 1, wherein the second learning dataset is data including the central portion and the area outside the central portion of the k-space of the target data.

7. The aforementioned training unit, An optimization process is performed to increase the degree of agreement between the data based on the output from the first machine learning model when the first training data from the first training dataset is input and the first training data, thereby generating the first trained model. The learning device according to claim 1, which generates the second trained model by performing an optimization process to increase the degree of agreement between the data based on the output from the second machine learning model when the second training data from the second training dataset is input and the second training data.

8. The learning apparatus according to claim 1, wherein the first machine learning model and the second machine learning model are models for reducing noise in the data.

9. The learning apparatus according to claim 1, wherein the first machine learning model and the second machine learning model are trained by unsupervised learning or self-supervised learning.

10. The learning device according to claim 1, wherein the target data is one of MR (Magnetic Resonance) images, k-space data, projection data, sinogram data, or CT (Computed Tomography) images.

11. Having a first resolution, a first training dataset based on the target data is obtained, The first machine learning model is trained using the first training dataset to generate the first trained model. A second machine learning model is trained using a second training dataset based on the target data and the first trained model, which has a second resolution higher than the first resolution, and a second trained model is generated. Learning methods.

12. On the computer, A first resolution is obtained, and an acquisition function is provided to acquire a first training dataset based on the target data. A program for implementing a training function that trains a first machine learning model using the first training dataset and generates a first trained model, The training function has a second resolution that is higher than the first resolution, and trains a second machine learning model using a second training dataset based on the target data and the first trained model, thereby generating a second trained model. Learning program.

Images