Medical image model training method and device, electronic equipment and storage medium

Abstract

The application discloses a medical image model training method and device, an electronic device and a storage medium. The method comprises the following steps: acquiring a target image sequence set, wherein the target image sequence set comprises M types of scanning sequences, and M is greater than or equal to 2; performing position registration on images in the M types of scanning sequences to obtain a first training set, wherein each first sample in the first training set comprises M images, the positions of the M images of the first sample correspond to each other and belong to different types of scanning sequences; dividing the first samples in the first training set into N batches, and performing training on a target image model based on a loss function, wherein the N batches respectively comprise a plurality of first samples corresponding in position. In this way, the influence of the position classification factor on the model can be excluded in the model training process, the target image model is driven to converge in the direction of the expected classification ability, and thus the expected training effect can be obtained under the condition of a small sample data amount.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of computers, and particularly relates to a medical image model training method and device, an electronic device and a storage medium. BACKGROUND

[0002] In the field of medical imaging, scanning imaging devices such as MR or CT are often used to scan target objects. The image sequences obtained by scanning may include multiple different types, and thus there is a need to identify the types of images. One possible method is to classify images based on deep learning of artificial intelligence. However, due to the privacy of data in the field of medical imaging, the application of deep learning usually has a large data bottleneck, which makes the model unable to obtain sufficient training data and affects the actual training effect.

[0003] The information disclosed in this BACKGROUND section is only for the purpose of increasing the understanding of the background of the present application and should not be taken as an acknowledgment or any form of suggestion that this information forms prior art that is publicly known. SUMMARY

[0004] The present application aims to provide a medical image model training method, which is used to solve the problem of poor training effect caused by the data bottleneck of medical image model training in the prior art.

[0005] To achieve the above-mentioned purpose, the present application provides a medical image model training method, which comprises:

[0006] obtaining a target image sequence set, wherein the target image sequence set comprises M types of scanning sequences, and M is greater than or equal to 2;

[0007] performing position registration on the images in the M types of scanning sequences to obtain a first training set, wherein each first sample in the first training set comprises M images, and the positions of the M images of the first sample correspond and belong to different types of scanning sequences;

[0008] dividing the first samples in the first training set into N batches, and performing training on a target image model based on a loss function, wherein the N batches respectively comprise a plurality of first samples corresponding in position.

[0009] In an embodiment, the positions of the images in the M types of scanning sequences are registered to obtain a first training set, specifically comprising:

[0010] respectively taking one image from the M types of scanning sequences to construct an image pair;

[0011] determining whether the position information of the M images in the image pair corresponds; if yes,

[0012] the first training set.

[0013] In an embodiment, the images in the M types of scan sequences are positionally registered to obtain the first training set, specifically comprising:

[0014] The image pairs belonging to the first training set are subjected to gray scale normalization, and the normalized image pairs are classified into the first training set.

[0015] In an embodiment, after the training of the target image model using the N batches of first samples reaches a preset training target, the method further comprises:

[0016] A second training set is obtained, each second sample in the second training set including M images, and the M images of the second sample belonging to different types of scan sequences.

[0017] The target image model is trained based on the second training set and a loss function.

[0018] In an embodiment, the method specifically comprises:

[0019] An image pair is constructed by taking one image from each of the M types of scan sequences.

[0020] It is determined whether the position information of the M images in the image pair is consistent; if not,

[0021] The image pair is classified into the second training set.

[0022] In an embodiment, the second sample includes images belonging to different visits, or different examinations, or different subjects.

[0023] In an embodiment, the images in the first sample belong to the same visit or the same examination.

[0024] The present application also provides a medical image model training device, comprising:

[0025] An acquisition module is configured to acquire a target image sequence set, wherein the target image sequence set includes M types of scan sequences, and M is greater than or equal to 2.

[0026] A registration module is configured to positionally register the images in the M types of scan sequences to obtain a first training set, wherein each first sample in the first training set includes M images, and the positions of the M images of the first sample correspond and belong to different types of scan sequences.

[0027] The training module is configured to divide the first samples in the first training set into N batches, each of which includes a plurality of first samples corresponding to a position, and perform training on the target image model based on a loss function.

[0028] The application also provides an electronic device, comprising:

[0029] at least one processor; and

[0030] a memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the medical image model training method as described above.

[0031] The application also provides a machine-readable storage medium storing executable instructions that, when executed, cause the machine to perform the medical image model training method as described above.

[0032] Compared with the prior art, the medical image model training method according to the application can exclude the influence of the position classification factor on the model in the model training process by constructing the first training set through the position registration operation, drive the target image model to converge in the direction of the expected classification ability, and thus obtain the expected training effect under the condition of a small amount of sample data.

[0033] In another aspect, when the target image model is trained by using the first training set, the positions of the first samples in each batch correspond to each other, and at each iteration, the target image model is also not affected by the influence of the position classification factor on the model, further accelerating the convergence speed of the model.

[0034] In another aspect, by training the target image model in two stages by using the first training set and the second training set in sequence, even under the condition of limited data samples, only by organizing the input mode of the samples, the model can have a faster convergence speed than the direct input of the image samples, and at the same time, the problems of possible model overfitting and weak generalization ability can be prevented. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is an application scenario diagram of the medical image model training method according to an embodiment of the application;

[0036] Figure 2 is a flowchart of the medical image model training method according to an embodiment of the application;

[0037] Figure 3 is an architecture diagram of the target image model backbone network in the medical image model training method according to an embodiment of the application;

[0038] Figure 4is a module diagram of a medical image model training apparatus according to an embodiment of the application;

[0039] Figure 5 is a hardware structure diagram of an electronic device according to an embodiment of the application. DETAILED DESCRIPTION

[0040] The application will be described in detail below with reference to the embodiments shown in the drawings. However, the embodiments do not limit the application, and the changes made by those skilled in the art to the structure, method, or function of the embodiments are included in the protection scope of the application.

[0041] The terms "first", "second", "third", "fourth" and the like used in the description and claims of the application and above-described drawings, if any, are used to distinguish similar objects, and are not necessarily used to describe a particular sequential or chronological order. It should be understood that the data thus used can be interchanged, where appropriate, so that the embodiments of the application described herein can be carried out in a different order than that depicted or described herein. In addition, the terms "comprising" and "comprising" and any variations thereof, are intended to cover not exclusively including, for example, a process, method, system, product or device that includes a series of steps or units, which does not necessarily limit to those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0042] Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, both hardware and software level technology. Artificial intelligence basic technology generally includes, such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology and machine learning / deep learning and other major directions.

[0043] Computer Vision (CV), Computer Vision is a science that studies how to make machines "see", further, it refers to using cameras and computers to replace human eyes to identify and measure targets, and further do image processing, so that the computer processing becomes more suitable for human eye observation or image transmission to instrument detection. As a scientific discipline, computer vision researches related theories and technologies, trying to establish artificial intelligence systems that can obtain information from images or multidimensional data. Computer vision technology usually includes image processing, image recognition, image semantic understanding, image retrieval, OCR, video processing, video semantic understanding, video content / behavior recognition, three-dimensional object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous localization and mapping, and other technologies, as well as common face recognition, fingerprint recognition and other biometric identification technologies.

[0044] With the research and progress of artificial intelligence technology, artificial intelligence technology is researched and applied in multiple fields, such as common smart home, smart wearable device, virtual assistant, smart speaker, smart marketing, unmanned vehicle, autonomous driving, unmanned aerial vehicle, robot, smart medical treatment, smart customer service, etc. In addition, artificial intelligence technology can also be applied in other fields, for example, in various embodiments of the present application, computer vision technology is used to input the position-registered first training set into the target image model for training by batch, and by excluding common position differences and other classification interference factors, the target image model can converge faster under fewer sample training, and finally serve as a usable medical image model.

[0045] Referring to Figure 1 In an implementation environment scenario example, a server and a terminal are connected through a network. A Medical Imaging Reading System (MIRS) is run on the terminal, and the MIRS is used for image management and reading management. In terms of image management, the MIRS supports multi-center image uploading and image query, and performs auditing and quality control management on the uploaded images. In terms of reading management, the MIRS supports reading process design, multi-level reading allocation, tracking and query, and supports multiple readings. The server obtains a medical image model that can be used for image classification by running the medical image model training method provided in the embodiments of the present application. When a user reads through the MIRS on the terminal, the medical image model can be called to determine the specific type of the image.

[0046] Alternatively, in other implementation scenarios, the medical image model training method provided by the embodiment can also be run by the server and the terminal together. For example, after the terminal obtains the target image sequence set, the terminal can first perform position registration on the images in the target image sequence set to obtain a first training set, and the server performs training on the target image model based on the first training set. The medical image model obtained by training can be deployed on the server or the terminal, and the present application does not limit this.

[0047] In the above implementation environment, the terminal and the server perform data communication through a communication network. The communication network can be a wired network or a wireless network, and can be at least one of a local area network, a metropolitan area network, and a wide area network. The server can be a standalone physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms, etc. The server and the terminal device can be independent devices or integrated into the same system, and are not limited herein.

[0048] Based on the above description of the model training process, it can be seen that the medical image model training method provided by the embodiment can be executed by any suitable computer device (terminal or server). Alternatively, the medical image model training method can be executed by the terminal and the server together. For ease of description, the medical image model training method executed by the computer device will be described hereinafter.

[0049] Referring to Figure 2 An embodiment of the medical image model training method provided by the present application is introduced. In this embodiment, the method comprises:

[0050] S11, obtaining a target image sequence set.

[0051] The image sequence in the target image sequence set can be generated by checking the target object through various technical means such as CT (Computed Tomography), MRI (Magnetic Resonance Imaging), etc. according to an image collection scheme. The target object can be a human or animal body, etc.

[0052] Each image sequence in the image sequence set can consist of multiple consecutive medical images. Taking the human body as an example, the image sequence can include at least one of the following: head, neck, chest, abdomen, pelvis, head and neck, head and neck chest, head and neck chest and abdomen, head and neck chest and abdomen pelvis, neck and chest, neck and chest and abdomen, neck and chest and abdomen pelvis, chest and abdomen, chest and abdomen, and abdominal pelvis. Head, neck, chest, abdomen, and pelvis can be understood as single body parts, while head and neck, head and neck chest, head and neck chest and abdomen, head and neck chest and abdomen pelvis, neck and chest, neck and chest and abdomen, neck and chest and abdomen pelvis, chest and abdomen, chest and abdomen, and abdominal pelvis can be understood as composite body parts consisting of multiple body parts.

[0053] Depending on the imaging method of the target object, the desired classification criteria, etc., the scan sequences in the icon image sequence set can be distinguished into different types.

[0054] In a demonstration scenario, the target image sequence is obtained through enhanced CT scans. Enhanced CT involves injecting a contrast agent (a liquid with a stronger ability to absorb X-rays than human tissue) intravenously. During the examination, the contrast agent circulates through the bloodstream, entering arteries, veins, hepatic artery, hepatic vein, renal artery, renal vein, and later, the ureters. Because cancerous or diseased tissues have a very active blood supply, the contrast agent follows the blood circulation to reach the lesion area. Therefore, the use of contrast agents enhances the contrast of the lesion, making it easier to accurately observe the location and size of the lesion on CT images.

[0055] In contrast-enhanced CT scans, the scanning phase corresponds to the period when the contrast agent reaches different parts of the body, facilitating the observation of lesions in different locations. For example, in the arterial phase of the scanning phase, the contrast agent fills the arterial system, enhancing arterial vessels and organs with rich blood supply; in the venous phase, which is later than the arterial phase, the portal vein is enhanced by the contrast agent, making it very useful for observing lesions in the liver; in the equilibrium phase, the contrast agent redistributes to the portal vein, inferior vena cava, and abdominal aorta, allowing the filling of the contrast agent to be seen in both the venous and arterial systems; in the delayed phase, the contrast agent in the blood vessels decreases, but if there are highly vascularized tissues such as tumors, the contrast agent decays more slowly, making it useful for observing these highly vascularized tissues.

[0056] In such a scenario, scanning sequences can be distinguished based on the different types of scanning periods. That is, a scanning sequence may only include a portion of the images from a single enhanced CT scan; from another perspective, the embodiments of this application do not limit the scanning sequence to be divided only by the number of scans in the actual scanning imaging. For a scanning sequence obtained from a single enhanced CT scan, based on the number of scanning periods it contains, it can correspond to two or more types of scanning sequences in the embodiments of this application.

[0057] In another demonstration scenario, the target image sequence set is obtained through MRI scans. MRI images are displayed in grayscale, reflecting the intensity of the magnetic resonance signal or the relaxation time of the tissue. The stronger the magnetic resonance signal, the greater the brightness; the weaker the signal, the smaller the brightness. From high to low magnetic resonance signals, this is represented on the image as white, gray, and black. Different tissues exhibit different grayscale representations on MRI, and the same tissue will show different signal characteristics (grayscale) on different MRI sequences.

[0058] MRI scans can include plain scans and contrast-enhanced scans. Contrast agents used in contrast-enhanced MRI scans are chemical products injected (or ingested) into human tissues or organs to enhance image observation. These include magnetic substances such as iron and manganese, whose density is higher or lower than that of surrounding tissues. When these agents approach hydrogen atoms in resonance, they effectively alter the magnetic field of protons, significantly shortening T1 (longitudinal relaxation) and T2 (lateral relaxation) relaxation times. This changes the relaxation rate of water protons in local tissues, improving the imaging contrast and resolution between non-lesion and lesion areas, providing more information for lesion localization and diagnosis. The contrast agent concentration in tissues changes with blood flow. For example, in tissues without lesions, where blood flow is smooth, the contrast agent concentration decreases rapidly, and the brightness of the corresponding non-lesion area in the MRI image gradually increases. Conversely, in tissues containing lesions, where blood flow is poor, the contrast agent concentration decreases relatively slowly, and the brightness of the corresponding lesion area in the MRI image changes slowly, creating a clear contrast with the brightness of the non-lesion area.

[0059] Depending on the scan type and whether it is enhanced, MRI scan sequences also include various types, such as: T2-weighted imaging (T2WI), T1-weighted imaging (T1WI), diffusion-weighted imaging (DWI), fluid inversion recovery attenuation sequence (FLAIR), T1WI with contrast enhancement, susceptibility-weighted imaging (SWI), arterial spin labeling perfusion imaging (ASL) perfusion imaging (PWI), magnetic resonance spectroscopy (MRS), intracranial artery imaging (MRA), and intracranial vein imaging (MRV). In such scenarios, the scan sequences can be directly distinguished by their different MRI scan sequence types.

[0060] In this embodiment, the target image sequence set to be collected should include M ≥ 2 types of scan sequences. For example, MRI image sequences can include four types: T1W, T2W, T2FLAIR, and DWI.

[0061] S12. Perform position registration on the images in the M types of scan sequences to obtain the first training set.

[0062] Location registration can be performed using image location information or by combining machine learning. In this embodiment, it is expected that each first sample in the first training set includes M images, and the locations of the M images of the first sample correspond to and belong to different types of scan sequences.

[0063] In one embodiment, an image pair can be constructed by taking one image from each of the M types of scanning sequences, and it can be determined whether the position information of the M images in the image pair corresponds. If so, the image pair is included in the first training set.

[0064] Taking MRI scan sequences as an example, they can be saved as DICOM images (i.e., DICOM files). The saving method for DICOM files is as follows: one medical image from an MRI scan sequence is saved as one DICOM file. If an image series is acquired, such as a brain scan or a whole-body scan, it will be saved as a corresponding number of DICOM files. Here, a DICOM file refers to a single, independent file (e.g., a file with the *.dcm extension). Each DICOM file contains image data corresponding to an image slice within the scan sequence. Multiple scan sequences can correspond to the same study for a single subject. DICOM files also include header files. Computer devices can read the header file to obtain information such as the corresponding image's ID, size, and location. The ImagePosition(Patient) attribute contains the image's coordinates (x, y, z) on a coordinate axis centered at the top-left corner. The z-coordinate value represents the coordinate value perpendicular to the scanning instrument plane (i.e., the scanning direction). Computer devices can determine whether two images represent the same position of a target object based on the Image Position property; if so, the two images are considered to correspond in position.

[0065] In a specific scenario, the METADATA (header file) of each DICOM image in a DICOM sequence can be obtained using PYDICOM or PYTHON DICOM loading libraries such as PYDICOM or SIMPLEITK. This allows access to the values ​​of the Study Instance UID and Series Instance UID fields, as well as the information indicating the sequence type (T1W, T2W, T2FLAIR, DWI, etc.). The image file organized according to the Study Instance UID, Series Instance UID, and the labeled sequence type has the following structure:

[0066] Study Instance UID 1:

[0067] T1W:

[0068] Series Instance UID-1

[0069] Series Instance UID-2

[0070] …

[0071] Series Instance UID-QT2W:

[0072] Series Instance UID-1

[0073] …

[0074] Series Instance UID-MT2FLAIR:

[0075] Series Instance UID 1-PDWI:

[0076] Series Instance UID 1-KStudy Instance UID 2:

[0077] T1W:

[0078] Series Instance UID-1

[0079] Series Instance UID-2

[0080] …

[0081] Series Instance UID-Q1

[0082] T2W:

[0083] Series Instance UID-1

[0084] …

[0085] Series Instance UID-M1

[0086] T2FLAIR:

[0087] Series Instance UID 1-P1

[0088] DWI:

[0089] Series Instance UID 1-K1…

[0090] Study Instance UID N:

[0091] T1W:

[0092] Series Instance UID-1

[0093] Series Instance UID-2

[0094] …

[0095] Series Instance UID-QNT2W:

[0096] Series Instance UID-1

[0097] …

[0098] Series Instance UID-MNT2FLAIR:

[0099] Series Instance UID 1-PNDWI:

[0100] Series Instance UID 1-KN

[0101] In this embodiment, the images in the first sample can belong to the same visit or the same examination. Taking the images in the first sample belonging to the same examination (i.e., corresponding to the same Study Instance UID) as an example, Q*M*P*K T1W-T2W-T2FLAIR-DWI image pairs can be constructed by permutation and combination under Study Instance UID1, Q1*M1*P1*K1 T1W-T2W-T2FLAIR-DWI image pairs can be constructed by permutation and combination under Study Instance UID N, ..., QN*MN*PN*KN T1W-T2W-T2FLAIR-DWI image pairs can be constructed by permutation and combination. For each T1W-T2W-T2FLAIR-DWI image pair, it is determined whether the position information of the four images in it corresponds, and the T1W-T2W-T2FLAIR-DWI image pair corresponding to the position information is included in the first training set.

[0102] In another embodiment, identification can also be performed using a pre-trained neural network model or a modified neural network model to determine whether the positions of the M images in an image pair are consistent. If so, the image pair is included in the first training set. For example, if it is desired to identify the scan locations "head," "neck," "chest," "abdomen," and "pelvis," a trained neural network model can receive image input and output using its built-in 5-class classifier. Of course, it is understood that if more types of scan locations need to be identified, the neural network model can be modified and a corresponding number of classifiers can be added, which will not be elaborated here.

[0103] By constructing the first training set through location registration, the influence of location classification factors on the model can be eliminated during model training, driving the model to converge toward the desired classification ability, and thus achieving the expected training effect with a small amount of sample data.

[0104] In this embodiment, the computer device can further perform grayscale normalization on the image pairs included in the first training set, and then include the normalized image pairs back into the first training set, thereby completing the sample amplification. For example, the grayscale value range of the original image is between 0 and 255. In the red-green-blue (RGB) color mode, when R=G=B=255, the image is white, and when R=G=B=0, the image is black. Different medical imaging factors cause inconsistencies in the grayscale information of tissues with the same properties. Grayscale normalization can reduce or even eliminate grayscale inconsistencies in the image while preserving grayscale differences with diagnostic value, so that the computer device can automatically analyze and process the image.

[0105] S13. Divide the first sample in the first training set into N batches, and train the target image model based on the loss function.

[0106] The target image model can be any classification model including neural network architectures, such as AlexNet (a convolutional neural network based on LeNet), VGG (Visual GeometryGroup Network), ResNet (residual convolutional neural network), or GoogleLeNet (a convolutional neural network based on Inception). Taking ResNet as the backbone network of the target image model as an example, this backbone network includes at least several convolutional layers and one fully connected layer. Figure 3 As shown, for example, the computer device uses a ResNet-18 network as the backbone network, which includes several convolutional layers and a fully connected layer. The residuals calculated between every two convolutional layers are input into the next convolutional layer for computation. The ResNet-18 network can achieve good compatibility with various services.

[0107] The batch size for the first sample can be determined based on the actual number of samples in the first training set. For example, if there are 24,000 samples in the first training set, the batch size can be set to 16, and there will be 24,000 / 16 = 1,500 iterations in one training cycle.

[0108] In this embodiment, each batch includes a first sample that corresponds to a specific location. Since the images in each first sample in the first training set have been pre-registered, each first sample has a corresponding location. For example, if all four images in first sample A correspond to the head location, then first sample A also corresponds to the head location; if first sample B corresponds to the head and neck location, the computer determines that the locations of first sample A and first sample B do not correspond; if first sample C corresponds to the head location, the computer determines that the locations of first sample A and first sample C correspond.

[0109] As can be seen, since the position of the first sample in each batch corresponds, the target image model will not be affected by the position classification factor in each iteration, which further accelerates the convergence speed of the model.

[0110] After training the target image model using the first training set to achieve the preset training objective, this application embodiment further proposes to continue training the target image model at this point in the next stage. The preset training objective can be determined according to the needs of different scenarios. For example, when training the target image model using the first training set, the degree of decrease in the objective function can be monitored. When the value of the objective function decreases to a certain level and no longer decreases, it is considered that the preset training objective has been achieved. Alternatively, the number of training cycles for the target image model can be calculated. When the number of training cycles reaches a preset value, it is considered that the preset training objective has been achieved. This application does not impose any limitations on this.

[0111] In the next stage of training, the target image model can be trained again using the acquired second training set and the aforementioned loss function. In one embodiment, the second training set includes several second samples, each of which also includes M images, and the M images of the second sample belong to different types of scan sequences.

[0112] Since the "location correspondence" restriction on the images included in the second samples of the second training set is no longer applied, the computer device can acquire more qualified second samples to construct the second training set. For example, the second samples may include images belonging to different visits, different examinations, or different subjects.

[0113] In this embodiment, when constructing the second training set, the part of the samples that do not meet the requirements of the first sample when constructing the first training set can also be fully utilized. That is, for the image pairs constructed from the images in the target image sequence set, it is determined whether the position information of the M images in the image pair is consistent; if not, the image pair is classified into the second training set.

[0114] In another embodiment, some or all of the M images of the second sample in the second training set can also belong to the same type of scan sequence. That is, there are no longer restrictions on "position correspondence" and "different types" for the images included in the second sample in the second training set, which will not be elaborated here.

[0115] By performing the above two-stage training on the target image model, in the first stage, the target image model is quickly converged using the first training set after position registration. In the second stage, the convergence degree of the target image model is further improved using the second training set that does not require position registration. It can be seen that even if the images in the target image sequence set are used for model training, the embodiments of the present application can have a faster model convergence speed through the organized two-stage sample input method, preventing problems such as model overfitting and weak generalization ability that may occur when directly inputting a limited number of images in the target image sequence set, and obtaining better test results on the validation dataset.

[0116] In a possible implementation manner, the medical image model provided by the embodiments of the present application can be created using the Pytorch framework and run on an Nvidia Tesla P40 graphics card, etc. The present application does not limit this.

[0117] Refer Figure 4 , an embodiment of the medical image model training device of the present application is introduced. In this embodiment, the medical image model training device includes an acquisition module 21, a registration module 22, and a training module 23.

[0118] The acquisition module 21 is used to acquire a target image sequence set, where the target image sequence set includes M types of scan sequences, M≥2; the registration module 22 is used to perform position registration on the images in the M types of scan sequences to obtain a first training set, where each first sample in the first training set includes M images, and the M images of the first sample correspond in position and belong to different types of scan sequences; the training module 23 is used to divide the first samples in the first training set into N batches and perform training on the target image model based on a loss function, where each of the N batches includes a number of first samples corresponding in position.

[0119] In one embodiment, the registration module 22 is specifically used to take one image from each of the M types of scanning sequences to construct an image pair; determine whether the position information of the M images in the image pair corresponds; if so, classify the image pair into the first training set.

[0120] In one embodiment, the registration module 22 is specifically used to perform grayscale normalization on the image pairs included in the first training set, and to include the normalized image pairs in the first training set.

[0121] In one embodiment, after the training module 23 uses the first samples of the N batches to train the target image model to achieve the preset training target, the training module 23 is further used to obtain a second training set, each second sample in the second training set includes M images, and the M images of the second sample belong to different types of scanning sequences; and the target image model is trained based on the second training set and the loss function.

[0122] In one embodiment, the registration module 22 is specifically used to take one image from each of the M types of scanning sequences to construct an image pair; determine whether the position information of the M images in the image pair is consistent; if not, classify the image pair into the second training set.

[0123] In one embodiment, the second sample includes images belonging to different visits, different examinations, or different subjects.

[0124] In one embodiment, the images in the first sample belong to the same visit or the same inspection.

[0125] As referred above Figures 1 to 3 This specification describes a method for training a medical image model according to embodiments thereof. The details mentioned in the above description of the method embodiments also apply to the medical image model training apparatus of embodiments thereof. The above-described medical image model training apparatus can be implemented in hardware, software, or a combination of hardware and software.

[0126] Figure 5 A hardware structure diagram of an electronic device according to an embodiment of this specification is shown. Figure 5 As shown, the electronic device 30 may include at least one processor 31, a memory 32 (e.g., non-volatile memory), a RAM 33, and a communication interface 34, and the at least one processor 31, memory 32, RAM 33, and communication interface 34 are connected together via a bus 35. The at least one processor 31 executes at least one computer-readable instruction stored or encoded in the memory 32.

[0127] It should be understood that the computer-executable instructions stored in memory 32, when executed, cause at least one processor 31 to perform the above-described combinations in the various embodiments of this specification. Figures 1 to 3 The description includes various operations and functions.

[0128] In the embodiments of this specification, electronic device 30 may include, but is not limited to: personal computer, server computer, workstation, desktop computer, laptop computer, notebook computer, mobile electronic device, smartphone, tablet computer, cellular phone, personal digital assistant (PDA), handheld device, messaging device, wearable electronic device, consumer electronic device, etc.

[0129] According to one embodiment, a program product, such as a machine-readable medium, is provided. The machine-readable medium may have instructions (i.e., the elements implemented in software as described above), which, when executed by a machine, cause the machine to perform the above-described combinations of the various embodiments of this specification. Figures 1 to 5 The various operations and functions described. Specifically, a system or apparatus equipped with a readable storage medium storing software program code that implements the functions of any of the embodiments described above, and enabling the computer or processor of the system or apparatus to read and execute the instructions stored in the readable storage medium.

[0130] In this case, the program code read from the readable medium itself can perform the functions of any of the above embodiments, and therefore the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of this specification.

[0131] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer or the cloud via a communication network.

[0132] Those skilled in the art will understand that the various embodiments disclosed above can be modified and varied without departing from the spirit of the invention. Therefore, the scope of protection of this specification should be defined by the appended claims.

[0133] It should be noted that not all steps and units in the above process and system structure diagrams are mandatory; some steps or units can be omitted according to actual needs. The execution order of each step is not fixed and can be determined as needed. The device structure described in the above embodiments can be a physical structure or a logical structure. That is, some units may be implemented by the same physical client, or some units may be implemented by multiple physical clients, or they may be jointly implemented by certain components in multiple independent devices.

[0134] In the above embodiments, the hardware units or modules can be implemented mechanically or electrically. For example, a hardware unit, module, or processor may include permanent dedicated circuitry or logic (such as a dedicated processor, FPGA, or ASIC) to perform the corresponding operation. The hardware unit or processor may also include programmable logic or circuitry (such as a general-purpose processor or other programmable processor), which can be temporarily configured by software to perform the corresponding operation. The specific implementation method (mechanical, dedicated permanent circuitry, or temporarily configured circuitry) can be determined based on cost and time considerations.

[0135] The specific embodiments described above with reference to the accompanying drawings are exemplary embodiments, but do not represent all embodiments that can be implemented or fall within the scope of the claims. The term "exemplary" as used throughout this specification means "serving as an example, instance, or illustration" and does not imply that it is "preferred" or "advantageous" compared to other embodiments. Specific details are included to provide an understanding of the described techniques. However, these techniques can be practiced without these specific details. In some instances, well-known structures and apparatuses are shown in block diagram form to avoid obscuring the concepts of the described embodiments.

[0136] The foregoing description of this disclosure is provided to enable any person skilled in the art to implement or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles applicable herein can be applied to other variations without departing from the scope of this disclosure. Therefore, this disclosure is not limited to the examples and designs described herein, but is consistent with the widest scope of the principles and novel features disclosed herein.

Claims

1. A method for training a medical image model, characterized in that, The method includes: Obtain a target image sequence set, wherein the target image sequence set includes M types of scan sequences, M≥2; The images in the M types of scanning sequences are registered to obtain a first training set. Each first sample in the first training set includes M images. The M images in the first sample are located in the same position and belong to different types of scanning sequences. The images in the first sample belong to the same visit or the same examination. The first samples in the first training set are divided into N batches, and the target image model is trained based on the loss function. Each of the N batches includes several first samples corresponding to a position. After training the target image model using the first samples of the N batches to achieve the preset training objective, a second training set is obtained. Each second sample in the second training set includes M images. The M images of the second sample belong to different types of scanning sequences. The second sample includes images belonging to different visits or different examinations. The target image model is trained based on the second training set and the loss function.

2. The medical image model training method according to claim 1, characterized in that, Position registration is performed on the images in the M types of scan sequences to obtain the first training set, specifically including: Take one image from each of the M types of scan sequences to construct an image pair; Determine whether the positional information of the M images in the image pair corresponds; if yes, The image pairs are assigned to the first training set.

3. The medical image model training method according to claim 2, characterized in that, Position registration is performed on the images in the M types of scan sequences to obtain the first training set, specifically including: The image pairs included in the first training set are grayscale normalized, and the normalized image pairs are then included in the first training set.

4. The medical image model training method according to claim 1, characterized in that, The method specifically includes: Take one image from each of the M types of scan sequences to construct an image pair; Determine whether the positional information of the M images in the image pair is consistent; if not, The image pairs are then assigned to the second training set.

5. A medical image model training device, characterized in that, include: An acquisition module is used to acquire a target image sequence set, wherein the target image sequence set includes M types of scan sequences, M≥2; The registration module is used to perform position registration on the images in the M types of scanning sequences to obtain a first training set. Each first sample in the first training set includes M images. The positions of the M images in the first sample correspond and belong to different types of scanning sequences. The images in the first sample belong to the same visit or the same examination. The training module is used to divide the first samples in the first training set into N batches and train the target image model based on a loss function. Each of the N batches includes several first samples corresponding to a location. After training the target image model using the first samples in the N batches to achieve a preset training target, a second training set is obtained. Each second sample in the second training set includes M images. The M images of the second sample belong to different types of scanning sequences. The second sample includes images belonging to different visits or different examinations. The target image model is trained based on the second training set and the loss function.

6. An electronic device, comprising: At least one processor; as well as A memory that stores instructions, which, when executed by the at least one processor, cause the at least one processor to perform the medical image model training method as described in any one of claims 1 to 4.

7. A machine-readable storage medium storing executable instructions that, when executed, cause the machine to perform the medical image model training method as described in any one of claims 1 to 4.

Images