Model training device, MR scan planning device, and MR scan planning method
Through the multi-task detection model combined with the prediction results of multiple tasks, the problem of target direction definition deviation in MRI scanning planning is solved, achieving higher accuracy and stability.
Patent Information
- Application Number
- CN202311230532.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-21
- Publication Date
- 2025-03-21
AI Technical Summary
In the prior art, when formulating MR scanning plans performed by MRI devices, it is difficult to effectively solve the problem of target direction definition deviation caused by individual differences in human structure, especially when the target structure is small or the lesions exist.
Using a multi-task detection model, the generalization and stability of the planning are improved by combining the prediction results of multiple tasks (such as methods based on feature point detection, image segmentation and linear detection) by including images of the region of interest as input, and combining the prediction results of multiple tasks (such as methods based on feature point detection, image segmentation and linear detection).
Through the use of multi-task detection model, the limitations of single-task detection can be overcome, the accuracy and stability of MR scanning planning can be improved, and the conditions of different individuals and lesions can be adapted to different individuals and lesions.
Smart Images
Figure CN119673413A_ABST
Abstract
Description
Technical Field
[0001] The implementation method relates to a model training device, an MR scan planning device and an MR scan planning method. Background Art
[0002] When a medical imaging device such as an MRI (Magnetic Resonance Imaging) device is used to scan an object, there are differences between different objects even for the same scanning area, so it is necessary to formulate a personalized scanning plan for a specific object. For example, characteristic parameters related to the object are obtained to predetermine a scanning protocol suitable for the object for scanning examination, so that the scanning is performed according to the scanning protocol to obtain medical images that are more conducive to reading.
[0003] For example, when planning an MR (Magnetic Resonance) scan performed by an MRI device, it is necessary to define parameters such as the scan center and the scan plane direction. Taking the scan plane direction as an example, the scan plane direction is usually along the direction of the scan object in the human body. For example, in the case of an orbital or optic nerve MR examination, it is necessary to scan along the optic nerve direction, and in the case of a spinal disc examination, it is necessary to scan along the disc plane direction. In order to determine the direction of the target structure, existing methods include feature point detection, image segmentation, shape detection (such as straight line detection), etc. However, due to individual differences in human body structure, different methods will have different difficulties in practical applications.
[0004] For example, the method based on feature point detection is to determine a direction based on two points. For the detection target, the first and last endpoints of the structure can be detected to define its direction, or multiple feature points can be detected along the main axis of the target to fit the direction of the target. However, when the target structure is small, the defined feature points are close, which makes the detected target direction very sensitive to the position of the feature points, and a small displacement of the feature points will cause a significant deviation in the obtained direction angle. For example, when the direction of the optic nerve is defined by the front and back endpoints of the optic nerve based on feature point detection, a small displacement of the endpoints will cause a significant deviation in the obtained direction angle. Compared with relying on only two points, if the optic nerve is detected as a whole, such as segmenting the optic nerve or detecting the straight line where the optic nerve is located, the detection result will be more stable.
[0005] In addition, the image segmentation-based method is to extract the mask of the target area through image segmentation technology, and the direction of the target structure can be determined by performing shape analysis on the mask. However, there may be under-segmentation and over-segmentation errors in the segmentation results, which directly affect the shape of the extracted mask and indirectly lead to errors in the definition of the target direction. For example, when segmenting the optic nerve through an image segmentation-based method, other tissues such as fat may be over-segmented, resulting in the extracted optic nerve mask being an irregular shape, and the actual direction of the optic nerve cannot be analyzed. On the other hand, if the positions of the front and rear endpoints of the optic nerve can be referenced, the direction of the detection can be corrected.
[0006] In addition, the method based on line detection is that the optic nerve and intervertebral disc appear to be close to straight lines in certain viewing angles, so the straight line structure in the image is directly detected through techniques such as Hough transform to determine the direction of the target. However, the morphology of the target may change due to lesions, which sometimes leads to failure of straight line detection. For example, when the scanning plane direction of the spinal intervertebral disc is detected by the method based on line detection, sometimes the morphology of the spinal vertebrae and intervertebral disc changes significantly due to lesions, the shape of the intervertebral disc becomes irregular, and the straight line structure cannot be detected. However, the direction of the intervertebral disc can be obtained by direct image segmentation or detecting the front and rear endpoints of the intervertebral disc on the plane. Summary of the invention
[0007] The present invention is completed in view of the above problems, and its purpose is to provide a model training device, an MR scanning planning device and an MR scanning planning method that can make MR scanning planning have better generalization and stability.
[0008] A technical solution of the present invention is an MR scan planning device, which comprises: an acquisition unit, which acquires an MR positioning image; a positioning unit, which determines a region of interest including a target structure for MR scanning in the MR positioning image; a multi-task detection unit, which calculates the parameters required for MR scan planning using a multi-task detection model that uses an image including the region of interest as input and outputs the parameters required for MR scan planning, wherein the multi-task detection model uses multiple prediction results of multiple tasks, and the multiple tasks are different detection methods for obtaining prediction results related to the same parameter; and an MR scan planning unit, which performs MR scan planning using the parameters calculated by the multi-task detection unit.
[0009] In addition, in the above-mentioned MR scan planning device, the multi-task detection model may also include: a multi-task detection module, which takes the image containing the region of interest as input and outputs multiple prediction results of multiple tasks; and a parameter calculation unit, which determines the corresponding parameters based on the prediction results.
[0010] Furthermore, in the above-mentioned MR scan planning device, the parameter calculation unit may determine the corresponding parameter by performing weighted averaging on the plurality of prediction results according to respective confidence levels of the plurality of prediction results.
[0011] Another technical solution of the present invention is an MR scan planning method, which includes: an acquisition step of acquiring an MR positioning image; a positioning step of determining a region of interest including a target structure for MR scanning in the MR positioning image; a multi-task detection step of calculating the parameters required for MR scanning planning using a multi-task detection model that uses an image including the region of interest as input and outputs the parameters required for MR scanning planning, wherein the multi-task detection model uses multiple prediction results of multiple tasks, and the multiple tasks are different detection methods for obtaining prediction results related to the same parameter; and an MR scan planning step of performing MR scanning planning using the parameters calculated in the multi-task detection step.
[0012] In addition, another technical solution of the present invention is a model training device, which is used to train a multi-task detection model used in MR scan planning, which uses an image containing the region of interest as input and outputs parameters required in MR scan planning, wherein the multi-task detection model uses multiple prediction results of multiple tasks, and the multiple tasks are different detection methods for obtaining prediction results related to the same parameter. The model training device comprises: an acquisition unit, which acquires training data including MR positioning images; a loss function setting unit, which sets a comprehensive loss function for the multiple tasks as the loss function used when training the multi-task detection model; and a learning unit, which uses the comprehensive loss function set by the loss function setting unit to train the multi-task detection model, thereby outputting the trained multi-task detection model.
[0013] According to the above technical solution of the present invention, by using a multi-task detection model with multiple prediction results of multiple tasks for MR scan planning, the multi-task detection results can be combined to make the MR scan planning have better generalization and stability. In addition, by analyzing the confidence of the multi-task results and assigning reasonable weights, the limitations of single-task detection can be overcome, reliable results can be identified and abnormal results can be excluded, thereby improving the accuracy of scan planning.
[0014] In addition, in terms of model training, learning multiple different but related tasks together can alleviate the overfitting of the model and share the underlying network. In addition, using the result consistency loss function can enable different tasks to achieve complementary effects in model parameter training. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1This is a block diagram showing an example of the functional configuration of an MR scan planning system including a model training device and an MR scan planning device according to the first embodiment.
[0016] Figure 2 (a) is an example diagram of image data in the training data used in the learning of the multi-task detection module; Figure 2 (b) is an example diagram of the form of the optic nerve direction detected by the method based on feature point detection; Figure 2 (c) is an example diagram of the form of the optic nerve direction detected by the image segmentation-based method; Figure 2 (d) is an example diagram of the form of the optic nerve direction detected using a method based on straight line detection.
[0017] Figure 3 It is a schematic diagram of the deep neural network structure of the multi-task detection module in the model training device involved in the first embodiment.
[0018] Figure 4 Schematic diagram of a preset task weight table according to the first embodiment.
[0019] Figure 5 It is a flowchart of the process of creating a multi-task detection model in the model training device involved in the first embodiment.
[0020] Figure 6 Schematic diagram showing the determination of the region of interest according to the first embodiment.
[0021] Figure 7 It is a schematic diagram of the use of a multi-task detection module in the MR scan planning device according to the first embodiment.
[0022] Figure 8 (a) shows an example graph of the output of a classifier, Figure 8 (b) is a schematic diagram showing an example of determining actual weights of multiple tasks using a classifier and a preset task weight table according to the first embodiment.
[0023] Fig. 9 (a) and (b) are schematic diagrams showing a method of eliminating abnormal values according to the first embodiment.
[0024] Fig.10 : is a flowchart showing the MR scan planning process in the MR scan planning apparatus according to the first embodiment.
[0025] Fig.11 This is a block diagram showing an example of the functional configuration of an MR scan planning system including a model training device and an MR scan planning device according to the second embodiment.
[0026] Fig.12 It is a schematic diagram of the deep neural network structure of the multi-task detection module in the model training device involved in the second embodiment.
[0027] Fig.13 It is a flowchart of the process of creating a multi-task detection model in the model training device involved in the second embodiment.
[0028] Fig.14 FIG. 4 is a schematic diagram showing the use of a multi-task detection module in the MR scan planning device according to the second embodiment.
[0029] Fig.15 : is a flowchart showing the MR scan planning process in the MR scan planning apparatus according to the second embodiment.
[0030] Description of Reference Numerals
[0031] 100, 100a model training device; 10 training data acquisition unit; 20, 20a loss function setting unit; 30, 30a learning unit; 31, 31a multi-task detection module training unit; 32 parameter calculation unit; 33 image classifier training unit; 200, 200a MR scan planning device; 40 acquisition unit; 50 positioning unit; 60, 60a multi-task detection unit; 61, 61a multi-task detection module; 62 parameter calculation unit; 63 outlier elimination unit; 70 MR scan planning unit. DETAILED DESCRIPTION
[0032] Hereinafter, with reference to the accompanying drawings, preferred embodiments of the model training device, the MR scan planning device, and the MR scan planning method of the present invention will be described in detail.
[0033] For the convenience of explanation, the present invention integrates a model training device that outputs a model and an MR scan planning device that uses the model to form an MR scan planning system for explanation. However, the model training device and the MR scan planning device may also be independent devices, and the model generated by the model training device is incorporated into the MR scan planning device as a product for use in MR scan planning (e.g., generating a scan protocol).
[0034] In addition, the model training device and MR scan planning device of the present invention are respectively composed of multiple functional modules, and can be installed as software in an independent computer or other device having a CPU (central process unit) and a memory, or installed in multiple devices in a dispersed manner, and implemented by a processor executing each functional module of the model training device or MR scan planning device stored in the memory. It can also be implemented in hardware as a circuit capable of executing each function of the model training device or MR scan planning device. The circuit implementing the model training device or MR scan planning device can send and receive data and collect data via a network such as the Internet. In addition, the model training device or MR scan planning device of the present invention can also be directly installed in a medical image acquisition device such as a magnetic resonance imaging device as a part of a medical image acquisition device that requires MR scan planning. In addition, the model training device and the MR scan planning device can be installed in the same device, or in different devices in different forms.
[0035] (First Embodiment)
[0036] First, refer to Figures 1 to 10 The first embodiment is described. In this embodiment, the scan planning in the optic nerve MR examination is described as an example. The optic nerve scanning direction is taken as an example of a parameter used in the scan planning, and an MR scan planning system that can obtain this parameter more accurately is described in detail.
[0037] Figure 1 1 is a block diagram showing an example of the functional configuration of an MR scan planning system including a model training device and an MR scan planning device according to the first embodiment. Figure 1 As shown, the MR scan planning system includes a model training device 100 for training a multi-task detection model and an MR scan planning device 200 using the multi-task detection model.
[0038] Among them, the model training device 100 comprises: a training data acquisition unit 10, which acquires training data including MR positioning images; a loss function setting unit 20, which sets a comprehensive loss function for multiple tasks as a loss function used when training a multi-task detection model; and a learning unit 30, which uses the comprehensive loss function set by the loss function setting unit 20 to train the multi-task detection model, thereby outputting a trained multi-task detection model.
[0039] Specifically, the training data acquisition unit 10 acquires medical image data and label data as training data from a database of a medical facility such as a hospital or an external device such as an image acquisition device.
[0040] Specifically, image data refers to medical image data including the target structure of the subject, such as an MR scan localization map. Depending on the needs of actual detection, the image data can be an original complete localization map or a pre-processed localization map, for example, a local area of the localization map centered on the target structure. Figure 2 (a) illustrates two-dimensional image data including the optic nerve region of the subject obtained by MR localization scanning, and this image data can also be regarded as image data of the region of interest.
[0041] Label data refers to the true value of the detection result corresponding to the image data in the training data. Label data includes, for example, the true position of the target feature point, the true mask of the target structure, or the true direction of the target structure. In this embodiment, the extension direction of the optic nerve is used as the optic nerve scanning direction, so the true extension direction of the optic nerve as the target structure is used as the label of the final prediction result. Label data can be obtained by verification methods such as physician annotation of image data.
[0042] The learning unit 30 uses the training data obtained by the training data acquisition unit 10 to learn the multi-task detection model, and obtains a multi-task detection model that uses image data as input and outputs parameters required in MR scan planning. In this embodiment, the multi-task detection model includes a multi-task detection module that implements multi-task detection in the framework of a neural network and a parameter calculation unit that determines the parameters to be output based on the detection results of the multi-task detection module. Therefore, the learning unit 30 has a multi-task detection module training unit 31 that trains the multi-task detection module, a parameter calculation unit 32, and an image classifier training unit 33 that trains an image classifier for calculating the weights of each of the multiple tasks.
[0043] The multi-task detection module training unit 31 constructs a multi-task detection module by constructing a deep neural network for multi-task learning to associate image data with multiple prediction results of the parameter obtained by multiple detection methods for obtaining the prediction results of a certain parameter in the image data.
[0044] Figure 3 Schematic diagram of the deep neural network structure of the multi-task detection module in the model training device involved in the first embodiment. Figure 3 As shown in the figure, in this neural network, the image data is formed into a feature graph through a shared layer (bottom layer network), and then connected to the respective neuron hierarchical structures (sub-networks) of multiple tasks 1 to 3 for processing, and the prediction results of each task are output. The neuron hierarchical structure of each task can adopt any existing network structure in deep learning, so the detailed description is omitted. Figure 3In the network structure shown, since all tasks are used to obtain prediction results related to the same parameter and are highly related, multiple tasks can share the underlying network, which extracts common image features to obtain a series of feature maps. Then, a sub-network is constructed for each task, and the common features extracted from the underlying network are input to each sub-network, thereby outputting the prediction results of different tasks.
[0045] The so-called multiple tasks are different detection methods for obtaining prediction results related to the same parameter. In this embodiment, as a method for obtaining the scanning direction of the optic nerve, the method based on feature point detection is set as Task 1, the method based on image segmentation is set as Task 2, and the method based on straight line detection is set as Task 3.
[0046] The method based on feature point detection is to predict key feature points in the sub-network of a specific task, such as the starting point and end point of the optic nerve, and output Figure 2 (b) shows a heat map of the key feature point positions or a position vector of two key feature points (x1, y1) and (x2, y2) The optic nerve scanning direction is determined based on the straight line where the feature point positions are located.
[0047] The image segmentation-based approach is to predict the segmentation mask in a task-specific subnetwork, which can output Figure 2 The predicted target area mask as shown in (c) is used to determine the optic nerve scanning direction by analyzing the shape of the mask.
[0048] The method based on line detection is to directly predict the direction vector of the target structure in the sub-network of a specific task. This method can be used, for example, to predict the longitudinal direction of the left ventricle and prostate. It is particularly suitable for predicting the direction of structures that are close to a straight line in a two-dimensional plane, such as the optic nerve and intervertebral disc. It can output Figure 2 A voting map in the Hough transform domain as shown in (d) or a direction vector showing the direction shown in the voting map One dimension θ in the voting map represents the angle of the line where the target structure is located in the plane, and the other dimension ρ represents the distance between the line and the center of the image.
[0049] The multiple tasks mentioned above in the implementation method are only examples. More than two tasks may be selected for parameter prediction, or other tasks other than the above three tasks may be selected for parameter prediction. For example, an oriented-object detection method may be selected as one of the tasks.
[0050] The multi-task detection deep neural network constructed in this way can serve as a multi-task detection module and output the prediction results of task 1, task 2, and task 3 respectively when image data is input. The multi-task detection module training unit 31 uses the multiple sets of training data obtained by the training data acquisition unit 10 to learn the constructed multi-task detection module. That is, the multi-task detection module training unit 31 uses the comprehensive loss function set by the loss function setting unit 20 to adjust the deep neural network in such a way that the prediction results of the multiple tasks output by the multi-task detection module are close to the true value of the label data, thereby optimizing the multi-task detection module and generating a trained multi-task detection module. Details of the comprehensive loss function will be described later.
[0051] The loss function setting unit 20 is used to set the loss function used in the deep neural network. The loss function used when learning the multi-task detection module is a comprehensive loss function formed by combining the loss functions of at least two tasks among the multiple tasks involved in the multi-task detection module. total Including accuracy loss function Loss acc The term and consistency loss function Loss consistency Item, precision loss function Loss acc The purpose of setting is to integrate multiple tasks and make each task obtain more accurate detection results through learning. consistency The purpose of setting is to make the detection results obtained by different tasks consistent, that is, the target directions predicted by different tasks are close. acc And consistency loss function Loss consistency The comprehensive loss function Loss total For example, it is shown in the following formula (1).
[0052] Loss total =αLoss acc +βLoss consistency Formula (1)
[0053] Among them, α and β are weights of different types of loss functions, which can be set arbitrarily according to the requirements or tendencies for accuracy and consistency in the detection project.
[0054] Precision loss function Loss accIt is obtained by weighted addition of the loss functions of at least two tasks among the multiple tasks. Specifically, for each sub-network of the task, the loss function between the output of the sub-network and the true value of the corresponding label data is calculated. For example, the loss function loss1 of task 1 can be the mean square error (MSE) calculated based on the heat map or the Euclidean distance calculated based on the position vector. The loss function loss2 of task 2 can be the calculation of binary cross entropy, DICE, etc. The loss function loss3 of task 3 can be the calculation of the cosine value based on the direction vector or the calculation of the MSE based on the heat map. Therefore, the overall accuracy loss function is shown in the following formula (2), for example.
[0055]
[0056] Among them, loss i is the loss function of the ith sub-network, λi is the weight of the loss function of the ith sub-network, and the λ weight can also be defined in the following three ways.
[0057] Method 1: On the training dataset, train an independent prediction model for predicting the target direction for each detection task, evaluate the accuracy of the prediction results, and set a high λ weight for the loss function of tasks with high accuracy.
[0058] Method 2: Traverse the weight distribution schemes of multiple loss functions and select the one with the best network performance. For example, there are three detection tasks, and the λ weight of each task traverses {0.5, 1.0, 1.5, 2.0}, then there are at most 4 3 =64 weight allocation schemes (due to the overlap of equal proportional coefficients such as 0.5:0.5:0.5, the actual number of weight allocation schemes is less than 64), and the loss function weight allocation scheme with the highest detection accuracy is selected from the multiple weight allocation schemes.
[0059] Method 3: When the numerical differences in the loss functions of different tasks are large (for example, different orders of magnitude), it is necessary to first adjust the order of magnitude of λ to achieve a balance between different tasks in the overall loss function.
[0060] Consistency loss function Loss consistency Calculate the difference between the results of different tasks and evaluate the consistency of the prediction results between at least two tasks in multiple tasks. For example, for the difference between task 1 and task 2, the intersection between the heat map of the key feature points and the segmentation mask can be calculated, and the negative value of the intersection size can be used as the consistency loss function Loss consistency One of the items.
[0061] In addition, for example, for the difference between Task 1 and Task 3, the angle between the predicted direction output by Task 3 and the straight line where the key feature points output by Task 2 are located can also be calculated as the consistency loss function Loss consistency For example, assuming that the position vector of the key feature point output by task 2 is The direction vector of the predicted direction output by task 3 is In the case of and The cosine of the angle between the two points Use -cosθ or 1-cosθ as the consistency loss function Loss consistency One of the items.
[0062] The loss function setting unit 20 sets the comprehensive loss function of the multiple tasks as described above, so that the learning unit 30 can train the multi-task detection module by linking the multiple tasks using the comprehensive loss function.
[0063] In addition, the loss function setting unit 20 may use only the accuracy loss function as the comprehensive loss function, or only the consistency loss function as the comprehensive loss function, and may also implement the training of multiple tasks linked by the learning unit 30 .
[0064] The parameter calculation unit 32 is used to combine the detection results of each task to determine the parameters ultimately used for MR scanning planning (in this embodiment, the ultimate optic nerve scanning direction).
[0065] For example, in this embodiment, the parameter calculation unit 32 can perform a weighted average on the prediction results of task 1, the prediction results of task 2, and the prediction results of task 3, and use the result of the weighted average as the final optic nerve scanning direction.
[0066] Since the actual conditions of the image may be diverse and complex (such as lesions and artifacts), the performance of each detection task under different image states may also be different. Therefore, different weights can be set for the task for each image state. The parameter calculation unit 32 pre-saves, for example, a preset task weight table that is preset according to the performance of different tasks under different image states, and has an image classifier that identifies and scores the image state, so that for the image that needs to be detected, it can calculate the weights of multiple tasks based on the identified image state and the preset weights in the preset task weight table to perform weighted averaging on the multiple tasks. Details of the action of the parameter calculation unit 32 in actual detection will be described in detail later.
[0067] Figure 4 Schematic diagram of the preset task weight table involved in the first embodiment. Figure 4As shown, preset weights are set in advance for each of the method based on feature point detection, the method based on image segmentation, and the method based on straight line detection in terms of the three states of "normal", "blurred boundary", and "lesion".
[0068] In addition, the image classifier used therein can also be directly obtained from an existing image classifier. However, in this embodiment, the learning unit 30 also has an image classifier training unit 33, which can train an image classifier that scores the performance of the image in the three states of "normal", "blurred boundary" and "lesion", and the trained image classifier is incorporated into the parameter calculation unit 32 for use.
[0069] The learning unit 30 outputs such a parameter calculation unit that can combine the prediction results of multiple tasks together with the multi-task detection module trained by the multi-task detection module training unit 31 as a multi-task detection model that takes an image containing a region of interest as input and outputs the parameters required in MR scan planning.
[0070] Thus, the model training device 100 generates a multi-task detection model and outputs it as a product. The overall processing of the model training device 100 according to the first embodiment will be described below. Figure 5 It is a flowchart of the process of creating a multi-task detection model in the model training device involved in the first embodiment.
[0071] First, after starting to create a multi-task detection model, the training data acquisition unit 10 acquires training data including medical image data and label data from the outside (step S501). Then, in step S502, the loss function setting unit 20 sets the comprehensive loss function Loss used in the deep learning of the multi-task detection module based on the neural network. total , for example, the comprehensive loss function Loss total Set to include the accuracy loss function Loss acc The term and consistency loss function Loss consistency Item.
[0072] Next, the process proceeds to step S503 , where the multi-task detection module training unit 31 uses a comprehensive loss function to train the multi-task detection module in the deep neural network structure of the multi-task detection module to obtain a trained multi-task detection module.
[0073] Next, entering step S504, the image classifier training unit 33 obtains training data from the outside for training an image classifier that recognizes (scores) multiple states of an image, and uses the training data to train an image classifier that inputs image data and outputs scores of multiple states of the image data, and then incorporates the trained image classifier into the parameter calculation unit 32 (step S505).
[0074] Next, the parameter calculation unit that can use the trained image classifier and the preset task weights to assign weights to the multiple detection results output by the trained multi-task detection module and output them together with the multi-task detection module as a multi-task detection model by weighted averaging the calculated parameters (step S506).
[0075] exist Figure 5 In the process shown, the execution order of each step is not limited to this. Steps S504 and S505 may be executed first, and then steps S501 to S503. The execution order may be changed appropriately.
[0076] The MR scan planning device 200 can obtain the multi-task detection model by purchasing products or other means, and use the multi-task detection model to perform MR scan planning.
[0077] return Figure 1 As described above, the MR scan planning device 200 comprises: an acquisition unit 40 for acquiring an MR positioning image; a positioning unit 50 for determining a region of interest including a target structure for MR scanning in the MR positioning image acquired by the acquisition unit 40; a multi-task detection unit 60 for calculating the parameters required for MR scan planning using a multi-task detection model that takes an image including the region of interest as input and outputs the parameters required for MR scan planning; and an MR scan planning unit 70 for performing MR scan planning using the parameters calculated by the multi-task detection unit 60.
[0078] Specifically, the acquisition unit 40 acquires a positioning image of a subject requiring MR scanning, for example, an image of a sagittal plane of the head through the optic nerve. The positioning image can be obtained by performing a pre-scan on the subject before the actual MR scanning.
[0079] The positioning unit 50 determines the region of interest including the optic nerve from the positioning image obtained by the obtaining unit 40. For example, the positioning unit 50 detects the center of the optic nerve as the target structure in the positioning image, and then expands a preset size range around the detected center as the region of interest. In the case where the area ratio of the region of interest in the positioning image is less than a preset threshold, the positioning unit 50 may also Figure 6As shown, the region of interest or a part of the region including the region of interest is cropped from the positioning image as image data for multi-task detection. The positioning image may also be subjected to other image preprocessing such as format conversion to obtain image data for multi-task detection.
[0080] The multi-task detection unit 60 comprises: a multi-task detection module 61, which takes an image containing a region of interest as input and outputs multiple prediction results of multiple tasks; a parameter calculation unit 62, which determines the corresponding parameters used in the MR scan planning according to the prediction results of the multi-task detection module 61; and an outlier exclusion unit 63, which uses a pre-set outlier judgment condition to judge whether the multiple prediction results of the multi-task detection module 61 are outliers.
[0081] The multi-task detection module 61 and the parameter calculation unit 62 constitute a multi-task detection model, which can be obtained from the model training device 100 and incorporated into the multi-task detection unit 60. The multi-task detection module 61 receives the image data including the region of interest after the region of interest is determined by the positioning unit 50, and outputs the detection results of each of the multiple tasks. The image data including the region of interest can be either the original medical image data or the image data obtained by the positioning unit 50. Figure 6 The cropped image data is shown on the right side of .
[0082] Figure 7 FIG. 1 is a schematic diagram of the use of a multi-task detection module in the MR scan planning device according to the first embodiment. Figure 7 As shown, the cropped image including the optic nerve is input to the multi-task detection module 61, so that the multi-task detection module 61 outputs the position vector of the direction dir1 as the output result of task 1. Or a heat map of the key feature point locations, a target region mask in direction dir2 as the output of task 2, and a direction vector in direction dir3 as the output of task 3 Or voting charts.
[0083] The abnormal value eliminating unit 63 determines whether there are abnormal results in the multiple directions dir1, dir2, and dir3 output by the multi-task detection module 61 according to a preset abnormal value judgment condition. Fig. 9 (a) and (b) are schematic diagrams showing the method of eliminating abnormal values involved in the first embodiment. For example, a maximum angle difference threshold is pre-set. Fig. 9 In the case where the angle difference between a certain direction dir3 and the other two directions dir1 and dir2 shown in (a) is greater than the maximum angle difference threshold, the direction dir3 is judged as an abnormal value. In addition, the abnormal value judgment condition can also be set to Fig. 9When the angle differences between all directions dir1, dir2, and dir3 are large (all larger than the maximum angle difference threshold) as shown in (b), it is determined that there is no abnormal value.
[0084] In addition, the present invention is not limited to this, and the abnormal value eliminating unit 63 may be omitted, and it may be directly regarded that there is no abnormal value.
[0085] In addition, the parameter calculation unit 62 determines the corresponding parameters based on the prediction results of the multi-task detection module 61 after excluding the prediction results determined as abnormal values. Specifically, the parameter calculation unit 62 performs weighted average on the directions remaining after excluding the abnormal values in the multiple directions dir1, dir2, and dir3, and uses the weighted average result as the optic nerve scanning direction Final_dir.
[0086] For example, when the outlier excluding unit 63 is not provided, or when the outlier excluding unit 63 determines that there is no outlier, the optic nerve scanning direction Final_dir is obtained as in the following equation (3).
[0087] Final_dir=w1*Dir1+w2*Dir2+w3*Dir3 Formula (3)
[0088] When the abnormal value eliminating unit 63 determines that the direction dir3 is an abnormal value, the optic nerve scanning direction Final_dir is obtained as shown in the following equation (4).
[0089] Final_dir=w1*Dir1+w2*Dir2 Formula (4)
[0090] Among them, w i is the weight assigned by the parameter calculation unit 62 to the measurement results of each task. The parameter calculation unit 62 has the image classifier and Figure 4 The preset task weight table shown assigns weights to the measurement results of each task by combining the scoring results of the image classifier and the preset task weight table.
[0091] Figure 8 (a) shows an example graph of the output of a classifier, Figure 8 (b) shows a schematic diagram of an example of determining the actual weights of multiple tasks using a classifier and a preset task weight table according to the first embodiment. Figure 7After the MR positioning image input into the multi-task detection module 61 is also input into the classifier, the classifier outputs the scoring result of the image status: normal 0.1, boundary blur 0.85, lesion 0.05. Then the parameter calculation unit 62 substitutes the scoring result into the preset task weight table, and adds the weighted weights of the corresponding states of multiple tasks. For example, for the task of feature point detection, normal 0.1 is multiplied by the preset weight 0.7, boundary blur 0.85 is multiplied by the preset weight 0.9, and lesion 0.05 is multiplied by the preset weight 0.7, and then each item is added to obtain the following: Figure 8 The parameter calculation unit 62 normalizes the actual weights of the tasks and uses them as weights assigned to the measurement results of the tasks, and calculates the optic nerve scanning direction Final_dir by weighted averaging the measurement results of the tasks.
[0092] The MR scan planning unit 70 performs MR scan planning using the optic nerve scan direction Final_dir. In addition to using the detection parameters detected by the multi-task detection unit 60 in MR scan planning, the MR scan planning unit 70 can use various existing MR scan planning methods, so detailed descriptions are omitted.
[0093] Fig.10 FIG. 4 is a flowchart showing the MR scan planning process in the MR scan planning device according to the first embodiment. Fig.10 As shown, the acquisition unit 40 acquires an MR positioning image (step S1001), then the positioning unit 50 determines the region of interest including the target structure of the MR scan in the MR positioning image acquired by the acquisition unit 40 (step S1002), inputs the image including the region of interest into the multi-task detection module in the multi-task detection unit 60, and acquires multiple prediction results of multiple tasks (step S1003). In step S1004, the outlier exclusion unit 63 excludes outliers from the multiple prediction results, so that the parameter calculation unit 62 determines the corresponding parameters according to the prediction results after excluding the outliers (step S1005), and the MR scan planning unit 70 performs MR scan planning using the parameters output by the parameter calculation unit 62 (step S1006).
[0094] In addition, in the first embodiment, the training data acquisition unit 10 corresponds to the "training data acquisition unit", the loss function setting unit 20 corresponds to the "loss function setting unit", the learning unit 30 corresponds to the "learning unit", the acquisition unit 40 corresponds to the "acquisition unit", the positioning unit 50 corresponds to the "positioning unit", the multi-task detection unit 60 corresponds to the "multi-task detection unit", and the MR scan planning unit 70 corresponds to the "MR scan planning unit".
[0095] According to the first embodiment, by using a multi-task detection model that uses multiple prediction results of multiple tasks to perform MR scan planning, the multi-task detection results can be combined to make the MR scan planning have better generalization and stability. In addition, by assigning reasonable weights to the detection results of multiple tasks when generating parameters, the limitations of single-task detection can be overcome, reliable results can be identified and abnormal results can be excluded, thereby improving the accuracy of scan planning.
[0096] In addition, in terms of model training, learning multiple different but related tasks together can alleviate the overfitting of the model and share the underlying network. In addition, using the result consistency loss function can enable different tasks to achieve complementary effects in model parameter training.
[0097] (Modification of the first embodiment)
[0098] In the first embodiment, when the parameter calculation unit 62 assigns weights to multiple tasks and performs weighted averaging to calculate the corresponding parameters, the weight of each task is calculated by an image classifier that is trained to score the image state. However, the present invention is not limited to this, and any method can be used to assign weights to multiple tasks. For example, an image classifier that classifies whether the performance of the task in detecting parameters on the image is reliable can also be trained for each task, and the weights of each task are adjusted according to the classification results of multiple image classifiers. For example, when the image is classified as a normal image and the result confidence of task 1 when the image is a normal image is high, the weight of task 1 is set higher. For example, when the image is classified as a difficult image and the result confidence of task 3 when the image is a difficult image is low, the weight of task 3 is set lower. Thus, the multiple prediction results of each task are weighted averaged with the adjusted weights to determine the corresponding parameters.
[0099] In addition, the parameter calculation unit 62 can of course directly set weights for each task. For example, it can also set the weights of each task based on the respective confidence levels of multiple prediction results, such as based on the intensity of the heat map, and perform weighted averaging on multiple prediction results to determine the corresponding parameters.
[0100] In addition, the parameter calculation unit 62 can also set the weight of each task according to the accuracy performance of multiple tasks in the verification set during the training phase, and perform weighted averaging on multiple prediction results of each task to determine the corresponding parameters. For example, for optic nerve detection, since the distance between key feature points is close, the direction calculated based on key feature points is easily affected and the accuracy performance is low, so the weight is set lower than other tasks. In addition, for the detection of 3D target direction, such as the detection of the direction of the left ventricle, direct direction prediction (based on straight line detection) is more difficult than the method based on feature point detection and the method based on image segmentation, so the weight of the method based on straight line detection is set lower than other tasks.
[0101] In addition, in the first embodiment, taking the scan planning in the MR examination of the optic nerve as an example, the optic nerve scanning direction is explained as an example of a parameter used in the scan planning. However, the present invention is not limited to this, and can also be applied to the determination of parameters such as the plane direction of the intervertebral disc in the planning of other MR scans such as the intervertebral disc of the spine. In addition to determining the direction of the target, the parameters also include determining the scan center by locating the target position and determining the scan range by detecting the target range. For example, the scanning range of the target in each direction can also be determined by a multi-task detection method. The multi-task detection method in this case includes locating the boundary points of the target anatomical structure using feature point detection, extracting the target area range using image segmentation, and predicting the bounding box of the target structure based on target detection. The parameters required to be obtained can be determined according to different inspection contents, and multiple methods (tasks) used to obtain the parameters can be selected, so that the present invention is applicable.
[0102] In addition, in the first embodiment, an MR scan planning system capable of obtaining an optic nerve scanning direction for MR scan planning is used as an example for explanation. However, the present invention is not limited to obtaining a single parameter, and the model training device may learn respective multi-task detection models for multiple scan parameters such as scan direction and scan center, so that the MR scan planning device uses the learned multiple multi-task detection models to obtain multiple scan parameters for MR scan planning.
[0103] (Second Embodiment)
[0104] Reference Figures 11 to 15The second embodiment is described. The MR scan planning system of the second embodiment is different from the first embodiment in that the model training device 100a has a loss function setting unit 20a instead of the loss function setting unit 20, and has a learning unit 30a instead of the learning unit 30, and the learning unit 30a does not have a parameter calculation unit 32 and an image classifier training unit 33. In addition, the MR scan planning device 200a has a multi-task detection unit 60a instead of the multi-task detection unit 60. The following mainly describes the differences, and appropriately omits repeated descriptions.
[0105] Fig.11 1 is a block diagram showing an example of the functional configuration of an MR scan planning system including a model training device and an MR scan planning device according to the second embodiment. Fig.11 As shown, the scanning protocol recommendation system includes a model training device 100a for training a multi-task detection model and an MR scanning planning device 200a using the multi-task detection model.
[0106] Among them, the model training device 100a comprises: a training data acquisition unit 10, which acquires training data including MR positioning images; a loss function setting unit 20a, which sets a comprehensive loss function for multiple tasks as the loss function used when training a multi-task detection model; and a learning unit 30a, which uses the comprehensive loss function set by the loss function setting unit 20a to train the multi-task detection model, thereby outputting a trained multi-task detection model.
[0107] Specifically, the training data acquisition unit 10 acquires medical image data and label data as training data from a database of a medical facility such as a hospital or an external device such as an image acquisition device.
[0108] The learning unit 30a uses the training data obtained by the training data acquisition unit 10 to learn the multi-task detection model, and obtains a multi-task detection model that takes image data as input and outputs parameters required for MR scan planning. In this embodiment, the multi-task detection model is a multi-task detection module that implements multi-task detection in the framework of a neural network. The learning unit 30a has a multi-task detection module training unit 31a that trains the multi-task detection module.
[0109] The multi-task detection module training unit 31a constructs a multi-task detection module by constructing a deep neural network for multi-task learning to associate image data with multiple prediction results of the parameter obtained by multiple detection methods for obtaining the prediction results of a certain parameter in the image data.
[0110] Fig.12Schematic diagram of the deep neural network structure of the multi-task detection module in the model training device involved in the second embodiment. Fig.12 As shown, in this neural network, the image data is formed into a feature map via a shared layer (bottom layer network) and then connected to a sub-network based on a feature point detection method as Task 1 and a sub-network based on an image segmentation method as Task 2. Furthermore, the output of the sub-network of Task 1, the output of the sub-network of Task 2, and the image data are input into a sub-network based on a line detection method as Task 3, and the sub-network of Task 3 directly predicts the optic nerve scanning direction based on the detection results of Task 1, the detection results of Task 2, and the image data. The optic nerve scanning direction output as the detection result of Task 3 is used as the final parameter.
[0111] The multi-task detection deep neural network constructed in this way can directly output the final detection result as a multi-task detection module when image data is input. The multi-task detection module training unit 31a uses the multiple sets of training data obtained by the training data acquisition unit 10 to learn the constructed multi-task detection module. That is, the multi-task detection module training unit 31a uses the comprehensive loss function set by the loss function setting unit 20a to adjust the deep neural network in a way that the final detection result output by the multi-task detection module is close to the true value of the label data, so as to optimize the multi-task detection module and generate a trained multi-task detection module. Details about the comprehensive loss function are described later.
[0112] The loss function setting unit 20a is used to set the loss function used in the deep neural network. The loss function used when learning the multi-task detection module is a comprehensive loss function that combines the loss functions of at least two tasks among the multiple tasks involved in the multi-task detection module. total It is the weighted sum of the loss function loss(Loss1, Loss2) of the output of the first two detection tasks and the loss function loss(Loss3) of the final output result, as shown in the following formula (5).
[0113]
[0114] Among them, loss i is the loss function of the i-th sub-network, λi is the weight of the loss function of the i-th sub-network, and the λ weight can also be set according to the same method as the weight setting method of the accuracy loss function described in the first embodiment.
[0115] The loss function setting unit 20 a sets the comprehensive loss function of the plurality of tasks as described above, so that the learning unit 30 a can train the multi-task detection module by linking the plurality of tasks using the comprehensive loss function.
[0116] The learning unit 30a outputs the multi-task detection module trained by the multi-task detection module training unit 31a as a multi-task detection model that takes an image including a region of interest as input and outputs parameters required for MR scan planning.
[0117] Thus, the model training device 100a generates a multi-task detection model and outputs it as a product. The overall processing of the model training device 100a according to the second embodiment will be described below. Fig.13 It is a flowchart of the process of creating a multi-task detection model in the model training device involved in the second embodiment.
[0118] First, after starting to create a multi-task detection model, the training data acquisition unit 10 acquires training data including medical image data and label data from the outside (step S1301). Then, in step S1302, the loss function setting unit 20a sets the comprehensive loss function Loss used in the deep learning of the multi-task detection module based on the neural network. total , for example, the comprehensive loss function Loss total Set it as the weighted sum of the loss function loss(Loss1, Loss2) of the output of the first two detection tasks and the loss function loss(Loss3) of the final output result.
[0119] Next, in step S1303, the multi-task detection module training unit 31a uses the comprehensive loss function Loss total , train the multi-task detection module in the deep neural network structure of the multi-task detection module, and obtain the trained multi-task detection module.
[0120] Next, the process proceeds to step S1304, where the trained multi-task detection module is output as a multi-task detection model.
[0121] The MR scan planning device 200a can obtain the multi-task detection model by purchasing products or other means, and use the multi-task detection model to perform MR scan planning.
[0122] return Fig.11 As described above, the MR scan planning device 200a comprises: an acquisition unit 40 for acquiring an MR positioning image; a positioning unit 50 for determining a region of interest including a target structure for MR scanning in the MR positioning image acquired by the acquisition unit 40; a multi-task detection unit 60a for calculating the parameters required for MR scan planning using a multi-task detection model that takes an image including the region of interest as input and outputs the parameters required for MR scan planning; and an MR scan planning unit 70 for performing MR scan planning using the parameters calculated by the multi-task detection unit 60a.
[0123] The multi-task detection unit 60a has a multi-task detection module 61a. In the multi-task detection module 61a, each of the multiple tasks is assigned an execution order, the task of feature point detection and image segmentation is assigned a first order, and the task of direction prediction is assigned a second order. The task of direction prediction is performed by executing the task of feature point detection and image segmentation in parallel and using the prediction results of feature point detection and image segmentation as inputs of direction prediction, thereby obtaining the prediction results of direction prediction as parameters.
[0124] Fig.14 FIG. 2 is a schematic diagram of the use of a multi-task detection module in the MR scan planning device according to the second embodiment. Fig.14 As shown, the cropped image including the optic nerve is input to the multi-task detection module 61a, so that the multi-task detection module 61a outputs the direction vector Dir as the optic nerve scanning direction. Or voting map. Thus, the multi-task detection module 61a takes as input an image containing a region of interest and outputs corresponding parameters used in MR scan planning.
[0125] The MR scan planning unit 70 uses the optic nerve scanning direction dir output by the multi-task detection module 61a to perform MR scan planning. In addition to using the detection parameters detected by the multi-task detection unit 60a in MR scan planning, the MR scan planning unit 70 can use various existing MR scan planning methods, so detailed description is omitted.
[0126] Fig.15 FIG. 4 is a flowchart showing the MR scan planning process in the MR scan planning device according to the second embodiment. Fig.15 As shown, the acquisition unit 40 acquires an MR positioning image (step S1501), then the positioning unit 50 determines a region of interest including a target structure of the MR scan in the MR positioning image acquired by the acquisition unit 40 (step S1502), inputs the image including the region of interest into the multi-task detection module in the multi-task detection model, and acquires corresponding parameters (step S1503). The MR scan planning unit 70 performs MR scan planning using the parameters output by the multi-task detection unit 60a (step S1504).
[0127] According to the second embodiment, by performing MR scan planning using a multi-task detection model that uses multiple prediction results of multiple tasks, the multi-task detection results can be combined to make the MR scan planning have better generalization and stability.
[0128] The components of each device in each of the above-mentioned embodiments are conceptual functions and do not necessarily need to be physically configured as shown in the figure. That is, the specific form of the dispersion and integration of each device is not limited to the content of the figure, and all or part of it can be functionally or physically dispersed and integrated in any unit corresponding to various loads and usage conditions. Furthermore, all or any part of each processing function performed in each device can be realized by a CPU and a program analyzed and executed by the CPU, or can be realized as hardware based on wired logic.
[0129] In addition, the model training device and the MR scan planning device described in the above-mentioned embodiments can be implemented by executing a pre-prepared program by a computer such as a personal computer and a workstation. The program can be distributed via a network such as the Internet. In addition, the program can also be recorded in a computer-readable non-temporary recording medium such as a hard disk, a floppy disk (FD), a CD-ROM, an MO, a DVD, etc. and read out from the recording medium by a computer to be executed.
[0130] The model training device and the MR scanning planning device can communicate with each other by wired or wireless connection, so as to transmit the model and data. Alternatively, the model trained by the model training device can be recorded in a storage medium as a product for transmission, and the product can be loaded into the MR scanning planning device for use.
[0131] Several embodiments of the present invention have been described, but these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and changes can be made without departing from the scope of the subject matter of the invention. These embodiments and their variations are included in the scope and subject matter of the invention, and are also included in the invention described in the claims and their equivalents.
Claims
1. An MR scanning planning device, characterized in that: have: An acquisition unit, for acquiring an MR positioning image; A positioning unit, which determines a region of interest including a target structure of the MR scan in the MR positioning image; a multi-task detection unit, which calculates the parameters required in the MR scan planning by using a multi-task detection model that takes the image containing the region of interest as input and outputs the parameters required in the MR scan planning, wherein the multi-task detection model uses a plurality of prediction results of a plurality of tasks, and the plurality of tasks are different detection methods for obtaining prediction results related to the same parameter; and The MR scan planning unit performs MR scan planning using the parameters calculated by the multi-task detection unit.
2. The MR scan planning device according to claim 1, characterized in that: The multi-task detection unit executes each of the plurality of tasks in the multi-task detection model in parallel, and outputs the parameter using prediction results corresponding to each task.
3. The MR scan planning device according to claim 1, characterized in that: Each of the multiple tasks in the multi-task detection model is assigned an execution order, and the multi-task detection unit executes the tasks in sequence according to the execution order of each task, and uses the prediction results obtained from executing the previous level task in the execution of the next level task, and outputs the parameters using the prediction results corresponding to each task of the last level.
4. The MR scan planning device according to claim 1, characterized in that: The multiple tasks include feature point detection, image segmentation or direction prediction.
5. The MR scan planning device according to claim 4, characterized in that: The multi-task detection unit performs the feature point detection, the image segmentation and the direction prediction in parallel to obtain the target structure direction based on the feature point detection, the target structure direction based on the image segmentation and the target structure direction based on the direction prediction.
6. The MR scan planning device according to claim 4, characterized in that: The feature point detection and the image segmentation are assigned a first order, and the direction prediction is assigned a second order, The multi-task detection unit performs the feature point detection and the image segmentation in parallel, and uses the prediction results of the feature point detection and the image segmentation as the input of the direction prediction, performs the direction prediction, and uses the prediction result of the direction prediction as the parameter.
7. The MR scan planning device according to claim 1, characterized in that: The parameters are the target structure direction, the MR scanning center or the scanning target range.
8. The MR scan planning device according to any one of claims 1 to 7, characterized in that: The multi-task detection model includes: a multi-task detection module, which takes the image containing the region of interest as input and outputs a plurality of the prediction results of a plurality of tasks; and The parameter calculation unit determines the corresponding parameters according to the prediction results.
9. The MR scan planning device according to claim 8, characterized in that: The parameter calculation unit performs weighted averaging on the plurality of prediction results of each task to determine the corresponding parameter.
10. The MR scan planning device according to claim 9, characterized in that: The parameter calculation unit performs weighted averaging on the plurality of prediction results according to the respective confidence levels of the plurality of prediction results to determine the corresponding parameters.
11. The MR scan planning device according to claim 9, characterized in that: The parameter calculation unit performs weighted averaging on the multiple prediction results of each task according to the accuracy performance of the multiple tasks to determine the corresponding parameters.
12. The MR scan planning device according to claim 9, characterized in that: The parameter calculation unit determines the corresponding parameters by weighted averaging the multiple prediction results of each task based on the scoring results of the image classifier that scores the image state of the MR positioning image and the preset weights set for the performance of different tasks under different image states.
13. The MR scan planning device according to claim 9, characterized in that: The parameter calculation unit adjusts the weight of each task based on multiple classification results of whether the performance of the task on the MR positioning image is reliable obtained by multiple image classifiers corresponding to the multiple tasks, and performs weighted averaging on the multiple prediction results of each task with the adjusted weight to determine the corresponding parameter.
14. The MR scan planning device according to claim 9, characterized in that: The multi-task detection unit also has an outlier exclusion unit, which uses pre-set outlier judgment conditions to determine whether the multiple prediction results of the multi-task detection module are outliers. The parameter calculation unit determines the corresponding parameters based on the prediction results after excluding the prediction results judged as outliers.
15. A model training device, characterized in that: for training a multi-task detection model used in MR scan planning, which takes an image containing the region of interest as input and outputs parameters required in MR scan planning, wherein the multi-task detection model uses multiple prediction results of multiple tasks, and the multiple tasks are different detection methods for obtaining prediction results related to the same parameter, The model training device comprises: A training data acquisition unit that acquires training data including an MR positioning image; A loss function setting unit, which sets a comprehensive loss function for the multiple tasks as a loss function used when training the multi-task detection model; as well as The learning unit uses the comprehensive loss function set by the loss function setting unit to train the multi-task detection model, thereby outputting a trained multi-task detection model.
16. The model training device according to claim 15, characterized in that: The learning department includes: a multi-task detection module training unit, which uses the comprehensive loss function set by the loss function setting unit to train a multi-task detection module that takes an image containing the region of interest as an input and outputs a plurality of the prediction results of a plurality of tasks; and A parameter calculation unit determines the corresponding parameter according to the prediction result. The learning unit outputs the multi-task detection module together with the parameter calculation unit as a multi-task detection model.
17. The model training device according to claim 16, characterized in that: The parameter calculation unit performs weighted averaging on the plurality of prediction results of each task to determine the corresponding parameter.
18. The model training device according to claim 17, characterized in that: The learning unit also includes an image classifier training unit, which trains an image classifier for scoring image states and integrates the image classifier into the parameter calculation unit. The parameter calculation unit stores preset weights for the performance of different tasks under different image states, and performs weighted averaging on multiple prediction results of each task based on the scores of the image classifier and the preset weights to determine the corresponding parameters.
19. The model training device according to claim 17, characterized in that: The learning unit further includes an image classifier training unit, which trains a plurality of image classifiers for classifying the performance of the task on the image according to each task. The parameter calculation unit adjusts the weight of each task according to the classification results of the multiple image classifiers, and performs weighted averaging on the multiple prediction results of each task with the adjusted weight to determine the corresponding parameter.
20. The model training device according to any one of claims 15 to 19, characterized in that: The comprehensive loss function is an accuracy loss function obtained by weighted addition of the loss functions of at least two tasks among the multiple tasks.
21. The model training device according to any one of claims 15 to 19, characterized in that: The comprehensive loss function is a consistency loss function that evaluates the consistency of prediction results between at least two tasks among the multiple tasks.
22. The model training device according to any one of claims 15 to 19, characterized in that: The comprehensive loss function is a weighted addition of an accuracy loss function obtained by weighted addition of the loss functions of at least two of the multiple tasks and a weighted addition of a consistency loss function that evaluates the consistency of prediction results between at least two of the multiple tasks.
23. A MR scan planning method, characterized in that: include: An acquisition step of acquiring an MR positioning image; A positioning step of determining a region of interest including a target structure of the MR scan in the MR positioning image; a multi-task detection step, calculating the parameters required in the MR scan planning by using a multi-task detection model that takes the image containing the region of interest as input and outputs the parameters required in the MR scan planning, wherein the multi-task detection model uses multiple prediction results of multiple tasks, and the multiple tasks are different detection methods for obtaining prediction results related to the same parameter; as well as The MR scan planning step uses the parameters calculated in the multi-task detection step to perform MR scan planning.