Target region determination method and medical imaging system

Abstract

The application provides a target region determination method and a medical imaging system, in particular, a target region determination method and device, a medical imaging system and a non-transitory computer readable storage medium. The target region determination method comprises preprocessing an original image to obtain a preprocessed medical image, obtaining a target point in the medical image, and obtaining a plurality of cross-sectional images of the medical image according to the target point, and obtaining a target region corresponding to the target point in the plurality of cross-sectional images based on a first model.

Description

Technical Field

[0001] This invention relates to medical image processing, and more particularly to a method for determining a target region, a medical imaging system, and a non-transitory computer-readable storage medium. Background Technology

[0002] During computed tomography (CT) scanning, detectors acquire X-ray data as it passes through the object being examined. This acquired X-ray data is then processed to obtain projection data. This projection data can be used to reconstruct CT images. Complete projection data allows for the reconstruction of accurate CT images for diagnostic purposes.

[0003] Typically, after acquiring CT images, they need to be segmented to obtain images of target regions for further diagnosis. For example, the heart region needs to be acquired or segmented from the CT image first, and then the heart region image is further processed or calculated to diagnose coronary artery stenosis.

[0004] In the process of acquiring images of the heart region, processing is usually based on 3D data. However, processing 3D data requires a large amount of computation and a large set of images. If 2D data is used, a lot of information will be lost. Summary of the Invention

[0005] This invention provides a method for determining a target region, a medical imaging system, and a non-transitory computer-readable storage medium.

[0006] An exemplary embodiment of the present invention provides a method for determining a target region. The method includes preprocessing an original image to obtain a preprocessed medical image; obtaining target points in the medical image and obtaining multiple cross-sectional views of the medical image based on the target points; and obtaining the target region corresponding to the target points in the multiple cross-sectional views based on a first model.

[0007] Specifically, the preprocessing includes at least one of size adjustment and pixel filling.

[0008] Specifically, the target point in the medical image is the center point of the original image.

[0009] Specifically, the plurality of cross-sectional views include transverse views, coronal views, and sagittal views.

[0010] Specifically, the training data for the first model includes a first dataset and a second dataset. The first dataset includes multiple cross-sectional images with preset target points. The second dataset includes the distances from the preset target points in each cross-sectional image to each edge of the reference target region. The preset target points include initial target points or offset target points. The initial target point is the center point of the original image. The offset target point is a target point with a preset offset from the reference target point. The reference target point is the center of the reference target region.

[0011] Specifically, obtaining the target region corresponding to the target point in the plurality of cross-sectional images includes inputting the plurality of cross-sectional images containing the target point into the first model to output the distance from the target point in each cross-sectional image to each edge of the reference target region; and obtaining the target region based on the distance.

[0012] Specifically, it further includes correcting the target point in each cross section to obtain the corrected target point, and obtaining multiple corrected cross sections based on the corrected target point. Obtaining the target region corresponding to the target point in the multiple cross sections includes obtaining the target region corresponding to the corrected target point in the multiple corrected cross sections.

[0013] Specifically, correcting the target points in each cross-sectional image includes obtaining the corrected target points based on the second model.

[0014] Specifically, the training data for the second model includes a third dataset and a fourth dataset. The third dataset includes multiple cross-sectional images with initial target points, and the fourth dataset includes the offsets between the initial target points and reference target points in each cross-sectional image. The initial target point is the location center point of the original image, and the reference target point is the center of the reference target region.

[0015] An exemplary embodiment of the present invention also provides a non-transitory computer-readable storage medium for storing a computer program that, when executed by a computer, causes the computer to perform the instructions described above for the target region determination method.

[0016] An exemplary embodiment of the present invention also provides a medical imaging system, the system including a processor that performs the target region determination method described above.

[0017] An exemplary embodiment of the present invention also provides a medical imaging system, the medical imaging system including a target region determination device, the device including a preprocessing module, a target point acquisition module, and a region determination module, wherein the preprocessing module is used to preprocess an original image to obtain a preprocessed medical image, the target point acquisition module is used to acquire target points in the medical image and acquire multiple cross-sectional views of the medical image based on the target points, and the region determination module is used to acquire the target region corresponding to the target point in the multiple cross-sectional views based on a first model.

[0018] Specifically, the target point in the medical image is the center point of the original image.

[0019] Specifically, the preprocessing includes at least one of size adjustment and pixel filling.

[0020] Specifically, the plurality of cross-sectional views include transverse views, coronal views, and sagittal views.

[0021] Specifically, the training data for the first model includes a first dataset and a second dataset. The first dataset includes multiple cross-sectional images with preset target points. The second dataset includes the distances from the preset target points in each cross-sectional image to each edge of the reference target region. The preset target points include initial target points or offset target points. The initial target point is the center point of the original image. The offset target point is a target point with a preset offset from the reference target point. The reference target point is the center of the reference target region.

[0022] Specifically, the region determination module is further configured to input multiple cross-sectional views containing the target point into the first model to output the distance from the target point in each cross-sectional view to each edge of the reference target region; and to obtain the target region based on the distance.

[0023] Specifically, the target region determination device further includes a target point correction module, which is used to correct the target points in each cross-sectional image to obtain corrected target points, and obtain multiple corrected cross-sectional images based on the corrected target points. The region determination module is used to obtain the target region corresponding to the corrected target points in the multiple corrected cross-sectional images.

[0024] Specifically, the target point correction module is further used to obtain the correction target point based on the second model.

[0025] Specifically, the training data for the second model includes a third dataset and a fourth dataset. The third dataset includes multiple cross-sectional images with initial target points, and the fourth dataset includes the offsets between the initial target points and reference target points in each cross-sectional image. The initial target point is the location center point of the original image, and the reference target point is the center of the reference target region.

[0026] Other features and aspects will become clear from the following detailed description, accompanying drawings, and claims. Attached Figure Description

[0027] The invention can be better understood by describing exemplary embodiments of the invention in conjunction with the accompanying drawings, in which:

[0028] Figure 1 This is a schematic diagram of a CT system according to some embodiments of the present invention;

[0029] Figure 2 This is a schematic diagram of a target area determination device according to some embodiments of the present invention;

[0030] Figure 3 This is a schematic diagram of a medical image and its corresponding multiple cross-sectional views;

[0031] Figure 4 This is a schematic diagram of the reference target area in medical images and cross-sectional views;

[0032] Figure 5 This is a schematic diagram of obtaining the target region using the first model;

[0033] Figure 6 This is a schematic diagram of a target area determination device according to other embodiments of the present invention;

[0034] Figure 7 This is a flowchart of a target region determination method according to some embodiments of the present invention; and

[0035] Figure 8 This is a flowchart of a target region determination method according to other embodiments of the present invention. Detailed Implementation

[0036] The following describes specific embodiments of the present invention. It should be noted that, in order to provide a concise description, this specification cannot exhaustively describe all features of the actual embodiments. It should be understood that, in the actual implementation of any embodiment, just as in any engineering or design project, various specific decisions are often made to achieve the developer's specific goals and to meet system-related or business-related constraints, and this can change from one embodiment to another. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this invention, some design, manufacturing, or production modifications based on the technical content disclosed herein are merely conventional technical means and should not be construed as insufficient content of this disclosure.

[0037] Unless otherwise defined, the technical or scientific terms used in the claims and description shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in the patent application description and claims of this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. The terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "comprising" or "including" and similar terms mean that the element or object preceding "comprising" or "including" encompasses the element or object listed following "comprising" or "including" and its equivalents, and do not exclude other elements or objects. The terms "connected" or "linked" and similar terms are not limited to physical or mechanical connections, nor are they limited to direct or indirect connections.

[0038] As used in this invention, the term "object being detected" may include any object being imaged.

[0039] It should be noted that, from the perspective of those skilled in the art or related fields, such description should not be construed as limiting the present invention only to CT systems. In fact, the target area determination method and apparatus described herein can be reasonably applied to other imaging fields in the medical and non-medical fields, such as X-ray systems, PET systems, SPECT systems, MR systems, or any combination thereof.

[0040] Figure 1 A schematic diagram of a CT system 10 according to some embodiments of the present invention is shown. Figure 1As shown, system 10 includes a rack 12 on which an X-ray source 14 and a detector array 18 are disposed opposite each other. The detector array 18 consists of multiple detectors 20 and a data acquisition system (DAS) 26. The DAS 26 is used to convert the sampled analog attenuated data received by the multiple detectors 20 into digital signals for subsequent processing. In some embodiments, system 10 is used to acquire projection data of the object under test at different angles. Therefore, the components on rack 12 are used to rotate around a rotation center 24 to acquire projection data. During rotation, the X-ray radiation source 14 projects X-rays 16 that penetrate the object under test toward the detector array 18. The attenuated X-ray beam data is preprocessed and used as projection data of the target volume of the object. Based on this projection data, an image of the object under test can be reconstructed. The reconstructed image can display the internal features of the object under test, including, for example, lesions, size, and shape of body tissue structures. The rotation center 24 of the rack also defines the center of the scanning domain 80.

[0041] System 10 further includes an image reconstruction module 50. As described above, the DAS 26 samples and digitizes projection data acquired by multiple detectors 20. The image reconstruction module 50 then performs high-speed image reconstruction based on the sampled and digitized projection data. In some embodiments, the image reconstruction module 50 stores the reconstructed image in a storage device or mass storage device 46. Alternatively, the image reconstruction module 50 transmits the reconstructed image to a computer 40 to generate patient information for diagnosis and evaluation.

[0042] although Figure 1 The image reconstruction module 50 is illustrated as a separate entity, but in some embodiments, the image reconstruction module 50 may be part of the computer 40. Alternatively, the image reconstruction module 50 may not be present in the system 10, or the computer 40 may perform one or more functions of the image reconstruction module 50. Furthermore, the image reconstruction module 50 may be located locally or remotely and may be connected to the system 10 using a wired or wireless communication network. In some embodiments, computing resources centralized in a cloud communication network may be used for the image reconstruction module 50.

[0043] In some embodiments, system 10 includes a control mechanism 30. The control mechanism 30 may include an X-ray controller 34 for providing power and timing signals to the X-ray radiation source 14. The control mechanism 30 may also include a gantry controller 32 for controlling the rotational speed and / or position of the gantry 12 based on imaging requirements. The control mechanism 30 may also include a bed controller 36 for driving a bed 28 to a suitable position to position the object to be inspected within the gantry 12 to acquire projection data of the target volume of the object. Further, the bed 28 includes a drive mechanism, and the bed controller 36 can control the bed 28 by controlling the drive mechanism.

[0044] In some embodiments, system 10 further includes computer 40, to which images reconstructed by data sampled and digitized by DAS 26 and / or image reconstruction module 50 are transmitted for processing. In some embodiments, computer 40 stores data and / or images in a storage device such as mass storage 46. Mass storage 46 may include hard disk drives, floppy disk drives, optical disc read / write (CD-R / W) drives, digital universal disk (DVD) drives, flash memory drives, and / or solid-state storage devices, etc. In some embodiments, computer 40 transmits reconstructed images and / or other information to display 42, which is communicatively connected to computer 40 and / or image reconstruction module 50. In some embodiments, computer 40 may connect to local or remote displays, printers, workstations, and / or similar devices, for example, such devices in medical institutions or hospitals, or remote devices via one or more configured wires or wireless communication networks such as the Internet and / or VPNs.

[0045] Furthermore, computer 40 can provide commands and parameters to DAS 26 and control mechanisms 30 (including rack controller 32, X-ray controller 34, and bed controller 36) based on user-provided and / or system-defined commands to control system operations, such as data acquisition and / or processing. In some embodiments, computer 40 controls system operations based on user input; for example, computer 40 can receive user input, including commands, scan protocols, and / or scan parameters, via an operator console 48 connected to it. Operator console 48 may include a keyboard (not shown) and / or a touchscreen to allow users to input / select commands, scan protocols, and / or scan parameters. Although Figure 1 Only one operator console 48 is shown as an example, but the computer 40 can connect to more consoles, for example, for inputting or outputting system parameters, requesting medical examinations, and / or viewing images.

[0046] In some embodiments, system 10 may include or be connected to an image storage and transmission system (PACS) (not shown). In some embodiments, the PACS is further connected to remote systems such as radiology information systems, hospital information systems, and / or internal or external communication networks (not shown) to allow operators located in different locations to provide commands and parameters, and / or access image data.

[0047] The methods or processes further described below may be stored as non-volatile memory on the computing device of the executable instruction storage system 10. For example, computer 40 may include executable instructions in non-volatile memory and may automatically execute part or all of the scanning process using the methods described herein, such as selecting a suitable protocol, determining suitable parameters, etc. Similarly, image reconstruction module 50 may include executable instructions in non-volatile memory and may perform image reconstruction tasks using the methods described herein.

[0048] Computer 40 can be configured and / or arranged for use in different ways. For example, in some implementations, a single computer 40 may be used; in other implementations, multiple computers 40 are configured to work together (e.g., based on a distributed processing configuration) or individually, each computer 40 being configured to process specific aspects and / or functions, and / or process data for generating models that are only used for a specific medical imaging system 10. In some implementations, computer 40 may be local (e.g., in the same location as one or more medical imaging systems 10, such as within the same facility and / or the same local communications network); in other implementations, computer 40 may be remote and therefore accessible only via a remote connection (e.g., via the Internet or other available remote access technologies).

[0049] Figure 2 A target region determination device 100 in a medical imaging system according to some embodiments of the present invention is shown. For example... Figure 2 As shown, the target region determination device 100 includes a preprocessing module 110, a target point acquisition module 120, and a region determination module 130. The preprocessing module 110 is used to preprocess the original image to obtain a preprocessed medical image. The target point acquisition module 120 is used to acquire target points in the medical image and acquire multiple cross-sectional views of the medical image based on the target points. The region determination module 130 is used to acquire the target region corresponding to the target point in the multiple cross-sectional views based on a first model.

[0050] In some embodiments, the target area determination device 100 may be as follows: Figure 1 This is part of the computer 40 (or processor) shown, or it may be stored and controlled in the form of a cloud.

[0051] Specifically, the original image refers to the image obtained after reconstructing the original data acquired after scanning. This original image can be directly acquired from a medical imaging system (e.g., CT system, MRI system, PET system, PET-CT system, etc.) or from a workstation or PACS. In some non-limiting embodiments, the size of the image acquired from the CT system can be 512×512×n, where n is the number of scan layers, for example, 512×512×320.

[0052] Preprocessing includes at least one of resizing and padding. Resizing refers to manipulating pixel values, such as averaging the pixel values ​​of two or more pixels. Padding involves filling certain pixel locations with zero pixel values. By preprocessing the original image, a medical image of a preset size can be obtained, for example, 256×256×256. Preprocessing the original image not only facilitates subsequent acquisition of cross-sectional images, determination of target points and target regions, but also benefits the training and learning of deep learning networks.

[0053] In some embodiments, preprocessing may also include conventional normalization and / or noise reduction.

[0054] Although the above-mentioned preset size for medical images is 256×256×256, those skilled in the art should understand that this size is merely exemplary and any other preset size is possible.

[0055] In some embodiments, the target point acquisition module 120 can acquire target points in a medical image, wherein the target point in the medical image is the center point of the original image. Specifically, for example, for an original image with a size of 512×512×320, the coordinates of its center point are (256, 256, 160). After the original image is resized, the image size can become 256×256×160, and the coordinates of the center point at this time are (128, 128, 80). Subsequently, the resized image is filled with pixels to obtain a medical image with a size of 256×256×256. The position of the target point in the medical image is still (128, 128, 80), which corresponds to the center point in the original image, but not the center point (128, 128, 128) of the medical image.

[0056] If the center point of a medical image is chosen as the target point, due to pixel padding, this target point may not be within the target area (e.g., within the heart region), or it may be significantly offset from the center point of the heart region. This would introduce a large bias into the subsequent method of obtaining a cross-sectional image from the target point to determine the target area from the outset. Therefore, to ensure that the obtained target point is as close as possible to the center of the target area, the center point of the original image is mapped onto the preprocessed medical image to obtain the target point, instead of selecting the center point of the medical image as the target point. This ensures that the target point is located as close as possible to the center of the target area.

[0057] Figure 3 A schematic diagram of a medical image 200 and its corresponding multiple cross-sectional views 210 / 220 / 230 is shown. (See diagram for reference.) Figure 3 As shown, the target point acquisition module 120 can acquire multiple cross-sectional views of the medical image 200 based on the acquired target point 202. Specifically, the multiple cross-sectional views include a coronal plane 210, a sagittal plane 220, and a transverse plane 230. Specifically, the sagittal plane is the anatomical plane that divides the human body into left and right sides; the coronal plane is the plane that longitudinally cuts the human body into anterior and posterior parts along its long axis; and the transverse plane is the plane that transversely cuts the human body into upper and lower parts from the anterior-posterior direction. Therefore, multiple cross-sectional views can be obtained from the three sections passing through the acquired target point.

[0058] By acquiring 2.5D images corresponding to medical images, namely multiple cross-sectional views: transverse, coronal, and sagittal views, and performing subsequent target region determination and / or other processing or operations based on 2.5D images, it is possible not only to solve the complex computational burden caused by 3D image-based calculations and processing, but also to solve the problem of information or detail loss caused by 2D images. By combining the advantages of 2D and 3D image processing, the computational burden and complexity can be reduced while retaining more information and details.

[0059] The region determination module 130 can determine target regions in medical images, such as the heart region, based on a first model. Specifically, the region determination module 130 includes a first training unit (not shown in the figure), which is trained and learned based on training data to obtain a first model, which is a deep learning network model.

[0060] Specifically, the training data for the first model includes a first dataset and a second dataset. The first dataset includes multiple cross-sectional images with preset target points. The second dataset includes the distances from the preset target points to the edges of the reference target region in each cross-sectional image. The preset target points include initial target points, which are the center points of the original image, and reference target points, which are the centers of the reference target regions.

[0061] Specifically, for the first dataset, the original image is preprocessed to obtain a medical image, and the center point in the original image is mapped to (or the same preprocessing operation is performed, but the pixel position remains unchanged) the medical image. This point in the medical image is the initial target point. Based on three cross-sections of the initial target point, corresponding cross-sectional images are obtained. Multiple sets of cross-sectional images containing the initial target point constitute the first dataset.

[0062] In some embodiments, for the second dataset, the reference target region is manually selected by the user or operator, and the selection criteria for the reference target region can vary; for example, the ends of coronary arteries can be preserved. In some non-limiting embodiments, the reference target region can be selected and displayed in a preprocessed medical image (e.g., the heart region). The reference target point is the center point of the reference target region. In the process of acquiring the second dataset, firstly, the reference target region is selected on the medical image; secondly, the corresponding cross-sectional image is obtained through the initial target point, and a two-dimensional reference target region (or a cross-sectional view of the reference target region) is displayed in each cross-sectional image; furthermore, the distances from the initial target point to each edge of the two-dimensional reference target region are obtained. Therefore, the distances from the initial target point to each edge of the reference target region in each cross-sectional image constitute the second dataset.

[0063] Figure 4 A schematic diagram of a reference target region is shown in a medical image and cross-sectional view. For example... Figure 4 As shown, in the medical image 300, region 301 shows a defined reference target region, and multiple cross-sectional views 310, 320 and 330 can be obtained through the center point 302 of the reference target region 301, each cross-sectional view showing the corresponding reference target region.

[0064] Specifically, the first model can be trained using the Inception-Resnet neural network model or other well-known models. The first dataset is used as the known input, and the second dataset is used as the expected output. When the first model is trained, as long as the cross-sectional image with the target points is input into the first model, the distance from each target point in the cross-sectional image to each edge of the reference target region can be obtained. Based on this distance, the (three-dimensional) target region can be obtained.

[0065] Figure 5 A schematic diagram illustrating the acquisition of the target region using the first model is shown. (For example...) Figure 5 As shown, by inputting the coronal plane 410, sagittal plane 420, and cross-sectional plane 430 containing the target point 401 into the first model, the distances (y+, y-, z+, z-) of the target point 401 from the four sides of the reference target region 411 in the coronal plane 410, the distances (x+, x-, z+, z-) of the target point 401 from the four sides of the reference target region 421 in the sagittal plane 420, and the distances (x+, x-, y+, y-) of the target point 401 from the four sides of the reference target region 431 in the cross-sectional plane 430 can be obtained. The reference target regions 411, 421, and 431 are the two-dimensional regions of the three-dimensional reference target region in the coronal plane, sagittal plane, and cross-sectional plane, respectively.

[0066] By obtaining the aforementioned multiple distances, the final three-dimensional target region can be obtained based on the target point.

[0067] Although the figure only shows an example of obtaining the target area of ​​the heart through CT images, those skilled in the art should understand that the target area is not limited to the heart region and can include any other region, such as brain structures or other tissue structures. Of course, it is not limited to obtaining the above-mentioned region through CT images, but can also be obtained based on MR images, PET-CT images or other images through processing.

[0068] Figure 6 A schematic diagram of a target area determination apparatus 500 according to other embodiments of the present invention is shown. For example... Figure 6 As shown, with Figure 2 Unlike the target region determination device 200 shown, the target region determination device 500 also includes a target point correction module 540. The target point correction module 540 corrects the target points in each cross-sectional image to obtain corrected target points, and obtains multiple corrected cross-sectional images based on the corrected target points. In addition, the region determination module 530 can obtain the target regions corresponding to the corrected target points in the multiple corrected cross-sectional images.

[0069] In some embodiments, since the position of the target point obtained based on the location center point in the original image may be too far from the actual target area or the center of the target area, resulting in a large error in obtaining the target area corresponding to the target point through the first model, the target area determination device 500 can further correct the target point so that the corrected target point is located as close as possible to the center of the target area.

[0070] In some embodiments, the target point correction module 540 is further configured to obtain a correction target point based on the second model.

[0071] The training data for the second model includes a third dataset and a fourth dataset. The third dataset includes multiple cross-sectional images with initial target points, and the fourth dataset includes the offsets between the initial target points and reference target points in each cross-sectional image. The initial target point is the location center point of the original image, and the reference target point is the center of the reference target region.

[0072] In some embodiments, the current third dataset is the same as the first dataset in the target region determination device 200, both including multiple cross-sectional views with initial target points.

[0073] The acquisition method of the fourth dataset is similar to that of the second dataset in the target region determination device 200. First, a reference target region is selected on the medical image. Second, corresponding cross-sectional views are obtained through the initial target points. In each cross-sectional view, a two-dimensional reference target region (or a cross-sectional view of the reference target region) is displayed. Third, the center point of the two-dimensional reference target region is determined, which is the reference target point. The center point of the two-dimensional reference target region is the mapping point of the center point of the three-dimensional reference target on that cross-section. Finally, the offset between the initial target point and the reference target point is obtained. This offset is the distance between the two target points along two axes (parallel to the two adjacent sides of the reference target region), for example, the distance between the two points along the two axes of the coordinate system. Therefore, the offset between the initial target point and the reference target point is the second dataset.

[0074] Specifically, the second model can be an Inception-ResNet neural network model or other well-known models. The third dataset serves as the known input, and the fourth dataset serves as the expected output. Once the second model is trained, simply inputting the cross-sectional image with the target point into the second model will allow us to obtain the offset between the target point and the reference target point. Based on this distance, the position of the corrected target point can be obtained.

[0075] In some embodiments, the training data of the first model in the region determination module 530 includes a first dataset and a second dataset. The first dataset includes multiple cross-sectional views with preset target points. The second dataset includes the distances from the preset target points in each cross-sectional view to each edge of the reference target region. The preset target points include offset target points, which are target points with a preset offset from the reference target point. The reference target point is the center of the reference target region.

[0076] Specifically, for the first dataset, for the determined reference target area and reference target point, manually select some offset target points with a preset offset from the reference target point and their corresponding cross-sectional images as the first dataset.

[0077] The second dataset is obtained in a manner similar to that of the second dataset in the target region determination device 200, by obtaining the distances from the offset target point to each edge of the reference target region in each cross-sectional view.

[0078] Once the first model is trained, simply inputting the calibration cross-section map with the calibration target points into the first model will allow us to obtain the distances from the calibration target points to each edge of the reference target region. Based on these distances, the target region can then be obtained.

[0079] Although the above-mentioned methods of using deep learning or artificial intelligence to correct the target point should be understood by those skilled in the art, other methods can also be used to correct the target point. When the initial target point is located within the target area and close to the center, it is not necessary to correct the target point.

[0080] The first and / or second models mentioned above are both deep learning networks. These deep learning networks may include an input layer, an output layer, and a processing layer (or hidden layer). The input layer is used to preprocess the input data or image, such as removing the mean, normalizing, or reducing dimensionality. The processing layer may include a convolutional layer for feature extraction, a batch normalization layer for standard normal distribution of the input, and an activation layer for nonlinear mapping of the output of the convolutional layer. In addition, it may include a fully connected layer for outputting the corresponding offset.

[0081] Each convolutional layer includes a number of neurons, and the number of neurons in each layer can be the same or different as needed. Based on a first or third dataset (known input) and a second or fourth dataset (desired output), the mathematical relationship between the known input and the desired output is identified and / or the mathematical relationship between the input and output of each layer is identified and characterized by setting the number of processing layers in the network and the number of neurons in each processing layer, and by estimating (or adjusting or calibrating) the weights and / or biases of the network.

[0082] Specifically, when the number of neurons in one layer is n, and the corresponding values ​​of these n neurons are X1, X2, ... X n The number of neurons in the next layer connected to one of the layers is m, and the corresponding values ​​of these m neurons are Y1, Y2, ... Y m Then the relationship between the two adjacent layers can be represented as:

[0083] The corresponding value, W ji B represents the weight. j This represents the deviation. In some embodiments, the function f is a corrected linear function.

[0084] Therefore, by adjusting the weight Wji and / or deviation B j This allows us to identify the mathematical relationship between the input and output of each layer, causing the loss function to converge, thus training the first model and / or the second model described above.

[0085] In one embodiment, while the configuration of the deep learning network is guided by prior knowledge of the estimation problem, the dimensions of the inputs, outputs, etc., it relies on or specifically on achieving the best approximation of the desired output data based on the input data. In various alternative implementations, certain aspects and / or features of the data, imaging geometry, reconstruction algorithms, etc., can be leveraged to give explicit meaning to certain data representations in the deep learning network, which can help accelerate training. This is because it creates the opportunity to train (or pre-train) or define certain layers separately in the deep learning network.

[0086] In some embodiments, the network trained described above is trained on a training module on an external carrier (e.g., a device other than a medical imaging system). In some embodiments, the training system may include a first module for storing the training dataset, a second module for training and / or updating the model, and a communication network for connecting the first and second modules. In some embodiments, the first module includes a first processing unit and a first storage unit, wherein the first storage unit stores the training dataset, and the first processing unit receives relevant instructions (e.g., retrieve training dataset) and sends the training dataset according to the instructions. Furthermore, the second module includes a second processing unit and a second storage unit, wherein the second storage unit stores the training model, and the second processing unit receives relevant instructions, trains and / or updates the network, etc. In other embodiments, the training dataset may also be stored in the second storage unit of the second module, and the training system may not include the first module. In some embodiments, the communication network may include various connection types, such as wired, wireless communication links, or fiber optic cables, etc.

[0087] Once data (e.g., a trained model) is generated and / or configured, it can be copied and / or loaded into the medical imaging system 10, which can be done in various ways. For example, the model can be loaded via a directed connection or link between the medical imaging system 10 and the computer 40. In this regard, communication between different components can be accomplished using available wired and / or wireless connections and / or according to any suitable communication (and / or network) standards or protocols. Alternatively or additionally, data can be loaded into the medical imaging system 10 indirectly. For example, data can be stored on a suitable machine-readable medium (e.g., a flash memory card, etc.) and then loaded into the medical imaging system 10 (on-site, such as by the system's user or authorized personnel) using that medium, or data can be downloaded to an electronic device capable of local communication (e.g., a laptop computer, etc.) and then used on-site (e.g., by the system's user or authorized personnel) to upload the data to the medical imaging system 10 via a direct connection (e.g., a USB connector, etc.).

[0088] As discussed in this paper, deep learning techniques (also known as deep machine learning, hierarchical learning, or deep structured learning) employ artificial neural networks for learning. Deep learning methods are characterized by using one or more network architectures to extract or model data of interest. Deep learning methods can be accomplished using one or more processing layers (e.g., input layers, output layers, convolutional layers, normalization layers, sampling layers, etc., with varying numbers and functions depending on the deep network model). The configuration and number of layers allow deep networks to handle complex information extraction and modeling tasks. Specific parameters (also called “weights” or “biases”) of the network are typically estimated through a so-called learning process (or training process). The parameters learned or trained usually result in (or output) a network corresponding to different levels of layers; therefore, extracting or modeling different aspects of the initial data or the output of a previous layer can often represent the hierarchical structure or cascade of layers. In image processing or reconstruction, this can be characterized as different layers relative to different levels of features in the data. Thus, processing can be layered; that is, earlier or higher-level layers may correspond to extracting “simple” features from the input data, followed by layers that combine these simple features into features exhibiting higher complexity. In practice, each layer (or more specifically, each "neuron" within a layer) can employ one or more linear and / or nonlinear transformations (so-called activation functions) to process the input data into an output data representation. The number of "neurons" can be constant across multiple layers or can vary from layer to layer. As discussed in this paper, as part of the initial training of a deep learning process to solve a specific problem, the training dataset includes known input values ​​(e.g., sample images or pixel matrices of images after coordinate transformation) and the expected (target) output value of the final output of the deep learning process (e.g., an image or recognition judgment result). In this way, the deep learning algorithm can process this training dataset (in a supervised or guided manner or in an unsupervised or unguided manner) until it identifies the mathematical relationship between the known inputs and the expected output and / or identifies and characterizes the mathematical relationship between the inputs and outputs of each layer. The learning process typically utilizes (partial) input data and creates a network output for that input data, then compares the created network output with the expected output of the dataset, and then uses the difference between the created and expected outputs to iteratively update the network parameters (weights and / or biases). The parameters of a network can typically be updated using the stochastic gradient descent (SGD) method; however, those skilled in the art will understand that other methods known in the art can also be used to update the network parameters.Similarly, a separate validation dataset can be used to validate the trained network, where both the known input and the desired output are known. The network output can be obtained by feeding the known input to the trained network, and then the network output is compared with the (known) desired output to validate previous training and / or prevent overtraining.

[0089] In some embodiments, a medical imaging system (or a computer therein) includes a processor capable of performing preprocessing on a raw image to obtain a preprocessed medical image; obtaining target points in the medical image and, based on the target points, obtaining a plurality of cross-sectional views of the medical image; and, based on a first model, obtaining a target region corresponding to the target points in the plurality of cross-sectional views.

[0090] Figure 7 A flowchart of a target region determination method 600 according to some embodiments of the present invention is shown. Figure 7 As shown, the target area determination method 600 includes steps 610, 620 and 630.

[0091] In step 610, the original image is preprocessed to obtain a preprocessed medical image.

[0092] Specifically, the original image refers to the image reconstructed from the raw data acquired after scanning. This original image can be obtained directly from a medical imaging system (e.g., CT system, MRI system, PET system, PET-CT system, etc.) or from a workstation or PACS. In some non-limiting embodiments, the size of the image acquired from the CT system can be 512×512×n, where n is the number of scan layers, for example, 512×512×320.

[0093] Preprocessing includes at least one of resizing and pixel padding. Resizing refers to further processing of pixel values, while pixel padding refers to filling certain pixel positions with 0 pixel values. By preprocessing the original image, a medical image of a preset size can be obtained. Preprocessing the original image not only facilitates the subsequent acquisition of cross-sectional images, the determination of target points and target regions, but also benefits the training and learning of deep learning networks.

[0094] In step 620, target points in the medical image are obtained, and multiple cross-sectional views of the medical image are obtained based on the target points.

[0095] The target point in a medical image is the center point of the original image. For example, for an original image with a size of 512×512×320, its center point has coordinates (256, 256, 160). After resizing the original image, the image size becomes 256×256×160, and the coordinates of the center point at this time are (128, 128, 80). Subsequently, the resized image is filled with pixels to fill the 256×256×256 medical image (i.e., the medical image obtained after preprocessing). The position of the target point in the medical image corresponds to the position of the center point (128, 128, 80) in the original image.

[0096] Multiple cross-sectional views include transverse, coronal, and sagittal views. These multiple cross-sectional views can be obtained from the three sections passing through the target point.

[0097] By acquiring 2.5D images corresponding to medical images, i.e., multiple cross-sectional views, and performing subsequent target region determination and / or other processing or operations based on 2.5D images, it is possible not only to solve the complex computational load brought about by 3D image calculation and processing, but also to solve the problem of information or detail loss brought about by 2D images. By combining the advantages of 2D image processing and 3D image processing, the computational load and complexity can be reduced while retaining more information and details.

[0098] In step 630, based on the first model, the target regions corresponding to the target points in the plurality of cross-sectional images are obtained.

[0099] Specifically, step 630 further includes inputting multiple cross-sectional images containing target points into the first model to output the distance from the target point in each cross-sectional image to each edge of the reference target region, and obtaining the target region based on the distance.

[0100] The training data for the first model includes a first dataset and a second dataset. The first dataset includes multiple cross-sectional images with preset target points. The second dataset includes the distances from the preset target points to the edges of the reference target region in each cross-sectional image. The preset target points include initial target points, which are the center points of the original image, and reference target points, which are the centers of the reference target regions.

[0101] Specifically, the first model can be an Inception-ResNet neural network model or other well-known models. The first dataset is used as the known input, and the second dataset is used as the expected output. When the first model is trained, as long as the cross-sectional image with the target point is input into the first model, the distance from the target point to each edge of the reference target region can be obtained. Based on this distance, the target region can be obtained.

[0102] For example, by obtaining the distances (y+, y-, z+, z-) of the target point from the four sides of the reference target region in the coronal plane, the distances (x+, x-, z+, z-) of the target point from the four sides of the reference target region in the sagittal plane, and the distances (x+, x-, y+, y-) of the target point from the four sides of the reference target region in the cross-sectional plane, the position of the entire target region can be obtained.

[0103] Figure 8 A flowchart of a target region determination method 700 according to other embodiments of the present invention is shown. Figure 7 The target region determination method 700, as shown in the target region determination method 600, further includes step 740.

[0104] In step 740, the target points in each cross-section are corrected to obtain corrected target points, and multiple corrected cross-sections are obtained based on the corrected target points.

[0105] Specifically, the correction of target points in each cross-sectional image includes obtaining correction target points based on the second model.

[0106] The training data for the second model includes a third dataset and a fourth dataset. The third dataset includes multiple cross-sectional images with initial target points, and the fourth dataset includes the offsets between the initial target points and reference target points in each cross-sectional image. The initial target point is the location center point of the original image, and the reference target point is the center of the reference target region.

[0107] Specifically, the second model can be trained using an Inception-ResNet neural network model or other well-known models. The third dataset serves as the known input, and the fourth dataset serves as the expected output. Once the second model is trained, simply inputting the cross-sectional image with the target point into the second model will yield the offset between the target point and the reference target point. Based on this offset, the corrected target point can be obtained.

[0108] Furthermore, in step 730, the target region corresponding to the correction target point in the plurality of correction cross-sectional images is obtained based on the first model.

[0109] The training data for the first model at this time includes a first dataset and a second dataset. The first dataset includes multiple cross-sectional images with preset target points, and the second dataset includes multiple cross-sectional images with reference target regions. The preset target points include offset target points, which are target points with a preset offset from the reference target points. The reference target points are the centers of the reference target regions.

[0110] Once the first model is trained, simply inputting the calibration cross-section map with the calibration target points into the first model will allow us to obtain the distances from the calibration target points to each edge of the reference target region. Based on these distances, the target region can then be obtained.

[0111] The target region determination method proposed in this invention firstly uses cross-sectional images (cross-sectional, coronal, and sagittal planes) to determine the target region, effectively avoiding the drawbacks of both 3D and 2D image processing. This combines the advantages of both, reducing computational and processing complexity while preserving image details. Secondly, it uses a deep learning network to determine the target region from the cross-sectional images, eliminating the need for manual determination by the user and thus improving work efficiency and user experience. Thirdly, by correcting the initial target points before target region determination, it ensures that the target points are located within the target region and as close as possible to its center, providing a foundation for accurate subsequent target region determination.

[0112] The present invention may also provide a non-transitory computer-readable storage medium for storing an instruction set and / or a computer program that, when executed by a computer, causes the computer to perform the target area determination method described above. The computer executing the instruction set and / or computer program may be a computer of a medical imaging system or other devices / modules of a medical imaging system. In one embodiment, the instruction set and / or computer program may be programmed into the processor / controller of the computer.

[0113] Specifically, when this instruction set and / or computer program is executed by the computer, it causes the computer to:

[0114] The original image is preprocessed to obtain a preprocessed medical image;

[0115] Obtain target points in the medical image, and based on the target points, obtain multiple cross-sectional views of the medical image; and

[0116] Based on the first model, the target region corresponding to the target point in the plurality of cross-sectional images is obtained.

[0117] As described above, instructions can be combined into a single instruction for execution, or any instruction can be split into multiple instructions for execution. Furthermore, the execution order of instructions is not limited to that described above.

[0118] As used herein, the term "computer" can include any processor-based or microprocessor-based system, including systems that use microcontrollers, reduced instruction set computers (RISC), application-specific integrated circuits (ASICs), logic circuits, and any other circuitry or processors capable of performing the functions described herein. The examples above are merely illustrative and are not intended to limit the definition and / or meaning of the term "computer" in any way.

[0119] The instruction set may include various commands that instruct a computer or processor, acting as a processor, to perform specific operations, such as methods and processes according to various embodiments. The instruction set may take the form of a software program, which may be part of one or more tangible, non-transitory computer-readable media. The software may take various forms, such as system software or application software. Furthermore, the software may take the form of a collection of independent programs or modules, a program module within a larger program, or part of a program module. The software may also include modular programming in the form of object-oriented programming. Input data processing by the processor may be in response to operator commands, previous processing results, or requests made by another processor.

[0120] Some exemplary embodiments have been described above; however, it should be understood that various modifications can be made. For example, suitable results may be achieved if the described techniques are performed in a different order and / or if components in the described system, architecture, device, or circuit are combined in a different manner and / or replaced or supplemented by other components or their equivalents. Accordingly, other embodiments also fall within the scope of the claims.

Claims

1. A method for determining a target region in a medical image, comprising: The original image is preprocessed to obtain a preprocessed medical image; The target point in the medical image is obtained, and multiple cross-sectional views of the medical image are obtained based on the target point. The multiple cross-sectional views include a cross-sectional view, a coronal view, and a sagittal view passing through the target point. The target points in each cross-section are corrected to obtain the corrected target points. Multiple corrected cross-sectional images are obtained based on the corrected target points; as well as Based on the first model, the distances from the correction target points in each correction section map to each edge of the reference target region are obtained, and the target regions corresponding to the correction target points in the plurality of correction section maps are obtained based on the distances. The correction target point is located as close as possible to the center of the target area.

2. The target region determination method as described in claim 1, wherein, The preprocessing includes at least one of resizing and pixel filling.

3. The target region determination method as described in claim 1, wherein, The target point in the medical image is the center point of the original image.

4. The target region determination method as described in claim 1, wherein, The training data for the first model includes a first dataset and a second dataset. The first dataset includes multiple cross-sectional images with preset target points. The second dataset includes the distances from the preset target points in each cross-sectional image to each edge of the reference target region. The preset target points include initial target points or offset target points. The initial target point is the center point of the original image. The offset target point is a target point with a preset offset from the reference target point. The reference target point is the center of the reference target region.

5. The target region determination method as described in claim 1, wherein, Correcting the target points in each cross-sectional view includes obtaining the corrected target points based on the second model.

6. The target region determination method as described in claim 5, wherein, The training data for the second model includes a third dataset and a fourth dataset. The third dataset includes multiple cross-sectional images with initial target points, and the fourth dataset includes the offsets between the initial target points and reference target points in each cross-sectional image. The initial target point is the location center point of the original image, and the reference target point is the center of the reference target region.

7. A non-transitory computer-readable storage medium for storing a computer program, which, when executed by a computer, causes the computer to perform the target region determination method according to any one of claims 1-6.

8. A medical imaging system, wherein the processor is configured to perform the target region determination method as described in any one of claims 1-6.

9. A medical imaging system comprising a target region determination device, the target region determination device comprising: The preprocessing module is used to preprocess the original image to obtain a preprocessed medical image; The target point acquisition module is used to acquire target points in the medical image and acquire multiple cross-sectional views of the medical image based on the target points. The multiple cross-sectional views include a cross-sectional view, a coronal view, and a sagittal view passing through the target points. The target point correction module is used to correct the target points in each cross-sectional image to obtain corrected target points, and to obtain multiple corrected cross-sectional images based on the corrected target points. as well as The region determination module is used to obtain, based on the first model, the distances from the correction target points in each correction section map to each edge of the reference target region, and to obtain the target regions corresponding to the correction target points in the plurality of correction section maps based on the distances. The correction target point is located as close as possible to the center of the target area.

10. The medical imaging system of claim 9, wherein, The target point in the medical image is the center point of the original image.

Images