Method and apparatus for detecting MRI images

Abstract

Provided is a method and apparatus for detecting MRI images. The method is executed by an MRI image detection model, which is pre-trained based on a deep learning method. The MRI image detection method includes: obtaining an MRI image to be detected, where the MRI image to be detected includes feature maps of multiple sequences of MRI images at different scales; fusing the feature maps of each scale of the multiple sequences of MRI images respectively to obtain a first fused feature map; and identifying a lesion area of the MRI image according to the first fused feature map. The method in the embodiments of the present application can improve the recognition efficiency and accuracy of the CVST lesion area.

Description

Technical Field

[0001] This application relates to the technical field of medical devices, and particularly to a method and device for detecting MRI images. Background Art

[0002] Cerebral Venous Sinus Thrombosis (CVST) is a rare stroke with an increasing incidence rate year by year. Due to various risk factors and non-specific clinical manifestations, it is easily misdiagnosed and missed diagnosed. Therefore, rapid and accurate diagnosis helps to carry out targeted clinical intervention for CVST.

[0003] Magnetic Resonance Imaging (MRI) is currently the most commonly used and effective method for identifying abnormal signals of CVST. Doctors judge whether there is a thrombus and its location by observing MRI images. However, in existing MRI image detection methods, mostly MRI images obtained by single-sequence scanning are used for diagnosis. Since the MRI images obtained by different-sequence scanning have different focuses, and the presented lesion areas are small and not obvious, it affects the recognition efficiency and accuracy of doctors for CVST lesion areas. Summary of the Invention

[0004] To solve the above problems existing in the prior art, embodiments of this application provide a method and device for detecting MRI images. The method and device can fuse multi-sequence MRI images, enhance the signal characteristics of the lesion area, and solve the problems of low recognition efficiency and low accuracy caused by single-sequence MRI images.

[0005] According to the first aspect of this application, a method for detecting MRI images is provided. The method is executed by an MRI image detection model, and the MRI image detection model is pre-trained based on a deep learning method. The MRI image detection method includes: obtaining an MRI image to be detected, where the MRI image to be detected includes feature maps of multiple sequences of MRI images at different scales; fusing the feature maps of each scale of the multiple sequences of MRI images respectively to obtain a first fused feature map; and identifying the lesion area of the MRI image according to the first fused feature map.

[0006] In one embodiment of the present application, fusing the feature maps of the MRI images of the multiple sequences at each scale to obtain a first fused feature map includes: acquiring the MRI images of different sequences, respectively constructing prototype feature libraries corresponding to the MRI images of different sequences, the prototype feature libraries classifying and storing the MRI images of different sequences, the size of the prototype feature library being K, where K is an integer and K≥2; calculating the similarity between the feature maps of the MRI images of the multiple sequences at different scales and each prototype feature map in the prototype feature library to obtain N similar prototype feature maps, where N is an integer and N≥1; reorganizing the N prototype feature maps to obtain a first reorganized feature map; and merging the first fused feature map and the first reorganized feature map to obtain a second fused feature map.

[0007] In one embodiment of the present application, the acquiring the MRI images of different sequences and respectively constructing the prototype feature libraries corresponding to the MRI images of different sequences includes: performing similarity training on the prototype feature maps in the prototype feature library, the similarity between each prototype feature map and other prototype feature maps being Ri, where R 1 >R 2 >……>R K , 1≤i≤K, when R 1 >R 2 + similarity distance D, the prototype feature library will reduce the similarity between the prototype feature maps.

[0008] In one embodiment of the present application, identifying the lesion area of the MRI image according to the first fused feature map includes: enhancing and optimizing the feature signals of the lesion area of the MRI image in the first fused feature map to identify the lesion area of the MRI image.

[0009] According to a second aspect of the present application, a method for training a learning model is provided, which is characterized by including: determining training sample images, the training sample images including MRI images of multiple sequences; performing data annotation on the training sample images to obtain first annotation data; performing preprocessing on the training sample images to obtain first training data; and training the learning model based on the first annotation data and the first training data to generate a detection result for identifying the lesion area of the MRI images of different sequences.

[0010] In one embodiment of the present application, the performing preprocessing on the training sample images includes: performing image registration, intensity bias correction, intensity normalization, central cropping, etc. on the training sample images.

[0011] According to a third aspect of the present application, there is provided a device for detecting MRI images, including: an extraction module configured to obtain the MRI images to be detected, where the MRI images to be detected include feature maps of multiple sequences of MRI images at different scales; a fusion module configured to fuse the feature maps of each scale of the multiple sequences of MRI images respectively to obtain a first fused feature map; and an identification module configured to identify the lesion area of the MRI images based on the first fused feature map.

[0012] According to a fourth aspect of the present application, there is provided a device for training a learning model, characterized by including: a determination module configured to determine training sample images, where the training sample images include multiple sequences of MRI images; a labeling module configured to perform data labeling on the training sample images to obtain first labeled data; a preprocessing module configured to preprocess the training sample images to obtain first training data; and a training module configured to train the learning model based on the first labeled data and the first training data to generate a detection result for identifying the lesion area of the MRI images of different sequences.

[0013] According to a fifth aspect of the present application, there is provided a chip, where the chip includes a processor and a data interface, and the processor reads instructions stored on a memory through the data interface and executes the method in the first aspect or any possible implementation manner of the first aspect.

[0014] Optionally, as an implementation manner, the chip may further include a memory, where instructions are stored in the memory, and the processor is configured to execute the instructions stored on the memory. When the instructions are executed, the processor is configured to execute the method in the first aspect or any possible implementation manner of the first aspect.

[0015] In a fifth aspect, there is provided a computer-readable storage medium, where the computer-readable medium stores program code for a device to execute, and the program code includes instructions for executing the method in the first aspect or any possible implementation manner of the first aspect.

[0016] In the embodiments of the present application, a deep learning method is used to fuse the feature maps of multiple sequences of MRI images at different scales, and the obtained fused feature map has complementary information between different sequences, so that it is easier to identify the lesion area from the MRI images, improving the identification efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 An exemplary block diagram for an application scenario applicable to the embodiments of the present application.

[0018] Figure 2Schematic block diagram of a method for detecting MRI images in an embodiment of the present application.

[0019] Figure 3 Exemplary block diagram of a method for detecting MRI images in another embodiment of the present application.

[0020] Figure 4 Exemplary block diagram of a feature fusion module in an embodiment of the present application.

[0021] Figure 5 Exemplary block diagram of a feature enhancement module in an embodiment of the present application.

[0022] Figure 6 Schematic block diagram of a method for training a learning model in an embodiment of the present application.

[0023] Figure 7 Exemplary block diagram of a method for training a learning model in another embodiment of the present application.

[0024] Figure 8 Schematic block diagram of a device for detecting MRI images provided in an embodiment of the present application.

[0025] Figure 9 Schematic block diagram of a device for training a learning model provided in an embodiment of the present application.

[0026] Figure 10 Schematic block diagram of a device for detecting MRI images provided in another embodiment of the present application. Detailed implementation manners

[0027] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0028] Magnetic Resonance Imaging (MRI)

[0029] Magnetic Resonance Imaging (MRI) applies a radiofrequency pulse of a specific frequency to the human body in a static magnetic field, causing the hydrogen protons in the human body to be excited and undergo magnetic resonance phenomena. After the pulse is stopped, the protons generate MRI signals during the relaxation process. Through processes such as the reception, spatial encoding, and image reconstruction of MRI signals, MRI signals are generated. Simply put, the principle of MRI is equivalent to applying a magnetic field to make the atomic nuclei move in rhythm with the magnetic field (i.e., resonance). When the magnetic field stops, the atomic nuclei return to their normal state, and this recovery process releases energy in the form of electromagnetic waves. The probe detects this energy and uses it for imaging to form an MRI image (also known as MRI).

[0030] Different from radiation imaging, in contrast imaging, the contrast depends on the attenuation rate of the imaged structure. In MRI, the contrast depends on the magnetism and the number of hydrogen nuclei in the imaged area. By running different sequences with different weights, different contrasts in the area to be imaged can be selected. The scanning sequence is a different combination of different radiofrequency pulses at different gradients and times, and the emphases reflected by MRIs of different sequences are not the same. Common scanning sequences include Spin Echo (SE), Inversion Recovery sequences, etc. Among them, the SE sequence includes, for example, T1 sequence, T2 sequence, etc., and the Inversion Recovery sequence includes Short TI Inversion Recovery (Stir) and Fluid Attenuated Inversion Recovery (Flair).

[0031] The T1 sequence in the SE sequence is also called T1-weight imaging (T1WI). The so-called weighting means highlighting. The T1 sequence is an MRI sequence weighted according to the T1 value and can be considered a modality of MRI. The signal amplitude collected by the T1 sequence mainly reflects the differences in T1 relaxation (longitudinal relaxation) of each tissue, and the T1 sequence is better for observing anatomical structures.

[0032] The T2 sequence in the SE sequence is also called T2-weight imaging (T2WI). The T2 sequence is an MRI sequence weighted according to T2 and can be considered a modality of MRI. The signal amplitude collected by the T2 sequence mainly reflects the differences in T2 relaxation (transverse relaxation) of each tissue, and the T2 sequence is better for showing tissue lesions.

[0033] The T1 sequence and T2 sequence in the above SE sequence have the characteristics of simple structure and high signal-to-noise ratio of the acquired images. However, the SE sequence has some disadvantages, such as low sensitivity to magnetic fields, long acquisition time, easy artifacts in the acquired images, and difficulty in dynamic enhancement.

[0034] The Flair sequence is also a commonly used sequence technology in MRI technology and can be considered as a modality of MRI. The Flair sequence can effectively suppress the signal of cerebrospinal fluid, thereby avoiding the missed diagnosis of lesions caused by the partial volume effect of cerebrospinal fluid. The Flair sequence can enhance the contrast of the T1 sequence and selectively suppress tissue signals with a certain T1 value. However, the Flair sequence also has its own shortcomings. For example, compared with the SE sequence, the signal-to-noise ratio is reduced, the scanning time is long, and there are more artifacts in the collected signals, which also affects its clinical application to a certain extent.

[0035] Of course, the MRI scanning sequence mentioned in this application also includes other sequences, such as proton density weighting in SE sequence, spin echo sequence (Fast Spin Echo, FSE), gradient echo sequence (Gradient Echo Pulse Sequence, GRE), echo planar imaging (Echo planar imaging, EPI), etc. This application does not limit the specific type of MRI scanning sequence. As mentioned above, no matter which sequence is scanned, the MRI image has certain disadvantages. According to research, the sensitivity of using different MRI sequences in diagnosing venous sinus thrombosis is between 34.4% and 83.5%. That is to say, the diagnostic accuracy of MRI images using different sequences fluctuates, which also affects the accuracy of MRI image detection.

[0036] Cerebral Venous Sinus Thrombosis (CVST)

[0037] Cerebral venous sinus thrombosis (CVST) is a rare type of stroke with an increasing incidence rate year by year. CVST is a characteristic type of cerebrovascular disease caused by various causes, characterized by obstruction of cerebral venous return and cerebrospinal fluid absorption disorder, accounting for about 0.5% to 1% of all strokes.

[0038] CT (Computed Tomography) can be used as an imaging method to detect CVST, but the positive rate of head CT is low and the specificity is not strong. 20% to 30% of CVST patients have normal head CT, and this proportion can be as high as 50% in patients with simple intracranial hypertrophy.

[0039] MRI is an effective method for identifying abnormal signals of cerebral veins and venous sinuses thrombosis, and it is also the most commonly used first-line screening method at present. MRI can directly display intracranial vein and venous sinus thrombosis, which is more sensitive and accurate than CT. In most cases, MRI can accurately diagnose CVST. However, as mentioned above, for different MRI scanning sequences, the accuracy rate of the obtained test results fluctuates. In the prior art, multiple MRI images are mostly required to be scanned by multiple sequences, and then the MRI images of different sequences are compared to more accurately judge whether there is CVST and the specific location where the thrombosis occurs. Obviously, this reduces the doctor's recognition efficiency. In addition, when the following situations occur, doctors need to read MRI images more carefully to reduce misdiagnosis. For example, one is when there is venous stenosis or slow blood flow, the normal blood flow signal representation is very similar to the thrombosis representation. If the image is not carefully observed, it is easy to be diagnosed as normal. The other is that acute and chronic thrombi are also easily overlooked and misdiagnosed as normal.

[0040] The above difficulties increase the time required to read MRI images, significantly reducing the diagnostic efficiency and accuracy of doctors (especially junior doctors with less experience in cerebral vein diseases). In addition, CVST is a relatively rare disease, and it is difficult to efficiently train a large number of doctors with relevant experience.

[0041] To solve the above existing problems, the present application aims to propose a method for detecting CVST based on multi-sequence MRI images, and this detection method can be implemented through a detection model. Compared with the MRI image diagnosis method of a single sequence, the embodiments of the present application fuse multi-sequence MRI images by extracting and complementing different sequence information, enhancing the signal representation ability of the CVST lesion area, obtaining a more obvious MRI image of the lesion area, thereby improving the diagnostic efficiency and accuracy of CVST. In addition, the present application also proposes a deep learning model, which can assist doctors in diagnosing CVST and further improve the diagnostic efficiency.

[0042] It should be understood that this method can be specifically executed by the processor of a local electronic device (such as a local medical device or other computer devices), or by a server in the cloud. The local electronic device communicates with the server in the cloud to obtain the lesion analysis result. The present application does not strictly limit the specific application hardware scenario of this lesion analysis method based on medical images.

[0043] After introducing the basic principle of the present application, the various non-limiting embodiments of the present application will be specifically introduced below with reference to the accompanying drawings.

[0044] Figure 1 It is an application scenario diagram applicable to the embodiments of the present application.

[0045] Figure 1 The application scenario 100 therein may include a user 110, an MRI image detection device 120, and an MRI image detection result 130. Among them, the user 110 may be, for example, a certain patient, or the brain MRI image of a certain patient. The MRI image detection device 120 may be any device or equipment capable of detecting the lesion area in the MRI image. For example, the MRI image detection device 120 may also be a server. For example, the MRI detection device 120 may be in the cloud (or may also be referred to as a cloud server). This is not limited in the embodiments of the present application.

[0046] In a possible implementation manner of the present application, the application scenario 100 may further include an MRI image scanner 140, and the MRI image scanner 140 may be, for example, a nuclear magnetic resonance imaging instrument (or NMR imaging instrument). At this time, the user 110 is a certain patient. The user 110 generates the brain MRI image through the MRI image scanner 140, and then inputs the generated MRI image into the MRI image detection device 120 for detection, and outputs the detection result 130. The detection result 130 can show the corresponding lesion area in the MRI image.

[0047] Figure 2 It is a schematic block diagram of a method 200 for detecting an MRI image provided by an embodiment of the present application. The method 200 may be executed by Figure 1 the MRI detection device 120 therein. For example, the MRI detection device 120 may include an MRI detection model, and the MRI detection model may be obtained by pre-training based on a deep learning method.

[0048] It should be understood that Figure 2 shows the steps or operations of the method 200, but these steps or operations are only examples. The embodiments of the present application may also perform other operations or Figure 2 variations of each operation of the method 200 therein, or not all steps need to be executed, or these steps may be executed in other orders.

[0049] Optionally, for the MRI detection model that executes the method 200, the network backbone of this detection model may be implemented based on the network structure of Mask-RCNN, or other forms of network architectures may also be used. The present application does not limit the network structure adopted.

[0050] Such as Figure 3The figure shown is an example diagram of an MRI detection model provided by an embodiment of the present application. The network backbone of this MRI detection is based on the network structure of Mask-RCNN. For example, the left end of the MRI detection model may include a feature extraction stage and a feature fusion stage, and the right end of the MRI detection model may include a feature recognition stage. For example, the feature recognition stage may include a feature enhancement module stage and subsequent stages of the Mask-RCNN network structure, such as a feature pyramid network, a region proposal network, region of interest alignment, region of interest recognition, non-maximum suppression, etc. The MRI detection model finally outputs a detection result, which may be, for example, the possible lesion area of CVST and the corresponding probability.

[0051] The following combines Figure 3 the MRI detection model in Figure 2 to introduce the process of the method for detecting MRI images shown in

[0052] S210, Obtain the MRI image to be detected.

[0053] Among them, the MRI image to be detected includes feature maps of MRI images of multiple sequences at different scales.

[0054] The MRI images of multiple sequences refer to MRI images obtained based on scans of multiple sequences. For example, it may be that P MRI images are obtained by scanning P sequences, where P is any integer greater than or equal to two. For example, P = 3, that is, three MRI images obtained by scanning three sequences are obtained.

[0055] In the field of computer vision, a feature map is relevant information required to complete a specific task. For example, when diagnosing CVST, the MRI image obtained by scanning the brain can be a feature map. Another example is that in face detection, an image formed by features related to a face, such as eyes, nose, mouth, etc., is a feature map. The feature map takes key point features as the main recognition points. For example, the key points can be special points that highlight features, such as the lesion area of CVST in the above-mentioned MRI image, or the eyes, nose, mouth, etc. in face recognition.

[0056] Scale, as an example, can be understood through the scale of a map. If the observation is carried out in units of 100 meters (referred to as scale A), a grassland can be observed. If the observation is carried out in units of 5 meters (referred to as scale B), it can be observed that there is a sheep near the grassland, and this sheep is invisible in the case of scale A. That is to say, in space, details can be seen at a low scale, and details will be smoothed out at a high scale, leaving only "macroscopic" features.

[0057] It is understandable that the manifestations of features at different scales are different. For example, a feature at a certain scale may not be the same at a larger scale. Therefore, it is necessary to extract features at the corresponding scale. As an example, since the size of the lesion area of CVST is not fixed and it will have different manifestation features at different scales, thus, by obtaining the feature maps of MRI images at different scales, more representative feature signals of CVST can be obtained, which is beneficial to better identify the lesion area.

[0058] As an embodiment, referring to Figure 3 the MRI detection model shown, multiple sequences can be, for example, three sequences, such as the commonly used T1 sequence, T2 sequence, and Flair sequence. Of course, other sequences can also be included. This application does not make specific restrictions on the number and type of sequences.

[0059] To obtain the feature maps of MRI images at multiple scales, for example, the process of obtaining the feature maps can be divided into multiple stages, and the scales of the feature maps obtained in each stage are different. For example, it can be divided into Q stages, where Q is any integer greater than or equal to 2. If Q = 4, then four different scales of feature maps can be obtained. That is to say, after each sequence generates a feature map in the i-th (1 ≤ i ≤ Q) stage, extract this feature map, and at the same time input this feature map into the (i + 1)-th stage to continue the feature map extraction, so as to obtain feature maps of different scales.

[0060] Referring to Figure 3 , the MRI detection model extracts the feature maps of the MRI images generated by three sequences at four different scales. As an example, the inputs of this model are respectively the MRI images generated by scanning the T1 sequence, T2 sequence, and Flair sequence. Then, the extraction of feature maps is performed on each sequence in four stages. Optionally, in the feature extraction of the first stage, only the first feature map generated by the Flair sequence can be extracted, such as Figure 3 the feature in Figure 1 , because the MRI images generated by the Flair sequence have better effects, so the MRI images in its first stage can be retained. Generally, the image scale corresponding to the first stage is larger and the calculation time-consuming is also relatively large, and it is usually discarded. It should be understood that in the first stage, the feature maps generated by all sequences can also be extracted. Of course, the feature maps of the first stage can also be discarded. The example given in this embodiment is only one example.

[0061] In the first stage, the first feature map generated by each sequence can be used as the input of the second stage to extract the second feature map, and so on, to extract the third and fourth feature maps corresponding to the three sequences respectively ( Figure 3 the first to fourth feature maps are not marked in

[0062] At S220, the MRI images of multiple sequences are respectively fused on the feature maps at each scale to obtain a first fused feature map.

[0063] Specifically, as described in step S210, after each sequence generates the i-th feature map at the i-th stage, in addition to extracting this feature map, this feature map can also be input into the (i + 1)-th stage to continue feature extraction, thereby realizing the extraction of feature maps at different scales. The feature maps (i.e., the i-th feature maps) extracted from multiple sequences at each stage are subjected to feature fusion to obtain a first fused feature map.

[0064] As an embodiment, referring to Figure 3 the MRI detection model, for example, the second feature maps respectively extracted from the T1 sequence, T2 sequence, and Flair sequence at the second stage are input into the multi-sequence feature fusion module 323 for feature fusion to obtain a feature fusion map at the second stage (i.e., Figure 3 the feature in Figure 2 ), the third feature maps respectively extracted from the three sequences at the third stage are input into the feature fusion module 323 to obtain a fused feature map at the third stage (i.e., Figure 3 the feature in Figure 3 ), and similarly, a fused feature map at the fourth stage is obtained (i.e., Figure 3 the feature in Figure 4 ).

[0065] Optionally, the fusion process of the feature maps of the multi-sequence MRI images at different scales can be implemented by Figure 4 way. It should be understood that the fusion method given in this embodiment is only one way of feature fusion, and other fusion methods can also be used for substitution. This embodiment does not limit the specific fusion method.

[0066] As an example, referring to Figure 4 , for example, it can be the fusion process of the i-th feature maps respectively extracted from the T1 sequence, T2 sequence, and Flair sequence at the i-th stage. The feature maps of the three sequences extracted at the i-th stage are respectively the T1 feature map 411, T2 feature map 412, and Flair feature map 413.

[0067] First, a merging operation is performed on the T1 feature map 411, T2 feature map 412, and Flair feature map 413. The merging operation can be, for example, a splicing operation or a stacking operation. This embodiment does not limit the specific way of merging. A fused feature map 420 is obtained through merging. Obviously, the fused feature map 420 contains the respective features of the T1 sequence, T2 sequence, and Flair sequence, realizing information complementarity among the three sequences.

[0068] Next, channel enhancement and spatial enhancement are respectively performed on the fused feature map 420 to obtain a channel-enhanced feature map 431 and a spatial-enhanced feature map 432. Channel enhancement can be, for example, strengthening the feature information in the fused feature map 420. Optionally, the channel enhancement process can perform a multiplication calculation on the fused feature map 420 and the feature map obtained after the fused feature map 420 undergoes operations such as spatial average pooling, a fully connected layer, and Sigmoid processing. Spatial enhancement can be, for example, strengthening the position information of the fused feature map 420. Optionally, the spatial enhancement process can perform a multiplication calculation on the fused feature map 420 and the feature map obtained after the fused feature map 420 undergoes operations such as feature max pooling, a convolutional layer, and Sigmoid processing.

[0069] Finally, the channel-enhanced feature map 431 and the spatial-enhanced feature map 432 are added to obtain a multi-sequence information fusion feature map 440 (also referred to as the original multi-sequence information fusion feature map). The addition calculation can be implemented, for example, by calling an addition function, such as performing an addition operation on the pixel values of an image. The obtained multi-sequence information fusion feature map 440 is the first fusion feature map. Obviously, the first fusion feature map realizes the strengthening of complementary features of multiple sequences and the strengthening of position features. Therefore, it is easier to identify the lesion area of CVST.

[0070] As an embodiment, in order to make the obtained first fusion feature map better highlight the difference between the features corresponding to the normal area and the features corresponding to the abnormal area in the MRI image, this embodiment introduces a prototype feature addressing stage, and the feature prototype addressing can be implemented by a prototype storage module and a prototype addressing module. For example, referring to Figure 3 , before the i-th feature map is input into the information fusion module 323, a prototype memory 321 and a feature addressing module 322 are added.

[0071] Feature addressing depends on prototype storage. In this embodiment, considering the lesion manifestation characteristics of CVST and the different manifestations at different degrees, a prototype feature library is constructed. The prototype feature library can be stored, for example, in the prototype storage module. The prototype feature library is used to classify and store the feature maps of MRI images of different sequences. For example, a prototype feature library can be constructed for each different sequence. The following combines Figure 3 to introduce the relevant content of the prototype addressing stage.

[0072] First, obtain MRI images of different sequences and construct prototype feature libraries corresponding to the MRI images of different sequences. The prototype feature libraries store and classify the MRI images. Assume the size of the prototype feature library is K, where K is an integer greater than or equal to 2. Optionally, the prototype feature library can be classified, for example, by the appearance, position, intensity, etc. of the features in the MRI images. The size of the prototype feature library can be, for example, the number of prototype feature maps (also called prototypes) in the prototype feature library. The feature prototype library can be constructed outside the MRI detection model. For example, it can be constructed based on the MRI images obtained from different patients and different scanned parts. The prototype feature library can also be constructed in the feature map extraction stage. For example, it can be constructed using the extracted feature maps. This application does not limit the construction method of the prototype feature library.

[0073] To make each prototype in the prototype feature library more representative, it is necessary to train the prototype feature library to obtain more expressive prototypes.

[0074] Specifically, for example, a similarity training model can be used to separate the prototypes in the prototype library. For example, a prototype feature library can be constructed for the MRI images generated for each sequence. The training samples in the prototype feature library can include normal MRI images of normal people and abnormal MRI images of CVST patients. Assume the similarity degree between a prototype A in the prototype library and the remaining prototypes is R i , where i is an integer with 1 ≤ i ≤ K. The similarity degrees between prototype A and the remaining prototypes, sorted from largest to smallest, are R 1 > R 2 > …… > R K . In other words, R 1 is the similarity degree between prototype A and the most similar prototype B among the remaining prototypes, and R 2 is the similarity degree between prototype A and the second most similar prototype C among the remaining prototypes, and so on. Optionally, when the formula R 1 > R 2 + similarity distance D is satisfied, the training model of the prototype feature library will be penalized. In this way, after multiple trainings, the training model of the prototype feature library will learn to reduce the similarity between prototypes to avoid the penalty, thereby achieving the separation of prototypes. The training model of the prototype feature library will ultimately obtain a prototype feature library with an appropriate similarity distance between each prototype, that is, a prototype feature library with each prototype being representative.

[0075] Similarity degree R iSimilarity degree and similarity distance D are two main factors for measuring the similarity between individuals. The similarity distance measures the distance between individuals in space. The farther the distance, the greater the difference between individuals. The similarity degree is opposite to the distance degree. The smaller the value of the similarity degree, the smaller the similarity between individuals and the greater the difference. This embodiment does not limit the calculation methods of the similarity degree and the similarity distance. The similarity degree can be calculated, for example, by algorithms such as the Pearson correlation coefficient and the Cosine similarity. The similarity distance can be calculated, for example, by algorithms such as the Euclidean distance, the Manhattan distance, and the Chebyshev distance.

[0076] Suppose that the features of two individuals X and Y in space both contain M dimensions, X = (x1, x2, x3,..., xm), and Y = (y1, y2, y3,..., ym). First, calculate the similarity distance D(X, Y) between X and Y. For example, using the Euclidean distance algorithm, then

[0077] Calculate the similarity degree R(X, Y) between X and Y. If the calculation method of the Pearson correlation coefficient is used, the value of the Pearson correlation coefficient generally ranges from [-1, +1]. The larger the absolute value, the stronger the correlation. The similarity degree between X and Y is:

[0078] According to the calculation methods described above, calculate the similarity between each prototype in the prototype feature library corresponding to each sequence, and then through the formula R 1 >R 2 + similarity distance D, to train the most suitable prototype feature library. The prototype feature library is a learnable parameter. For example, it can be continuously updated through the iterative optimization of the network to achieve the optimal prototype feature library.

[0079] As an embodiment, the prototype addressing module calculates the similarity between the feature map of the MRI image extracted in step S210 and each prototype in the prototype feature library based on the prototype feature library, and selects N similar prototype feature maps therefrom, where N is an integer and N≥1.

[0080] Specifically, when the prototype addressing module receives feature maps of different scales of MRI images of different sequences, it will compare the feature map with each prototype in the prototype feature library and calculate the degree of similarity between them. For example, the Pearson correlation coefficient and other methods introduced above can be used for calculation. For example, the range of the degree of similarity is [-1, 1] or [0, 1]. For example, by setting a threshold, N prototypes that meet the threshold condition are selected from the K prototypes in the prototype feature library.

[0081] As an embodiment, the feature graphs of the selected N prototypes may be reorganized to obtain a first reorganized feature graph.

[0082] Refer to the following Figure 3 , the feature addressing stage is illustrated by an example. When the feature prototype addressing module 322 receives the feature graphs generated by the three sequences collected in the feature extraction stage 310 in each stage, such as the i-th feature graph of T1, the i-th feature graph of T2 and the i-th feature graph of Flair, it calls the prototype feature library of the corresponding sequence in the prototype memory 321. Taking the T1 sequence as an example, the feature prototype addressing module 322 compares the i-th feature graph of T1 with each prototype in the prototype feature library corresponding to T1, and calculates the similarity degree R i , if R i If the value is greater than the set threshold, the prototype is selected to obtain a similar prototype of the i-th feature map of T1. Similarly, similar prototypes of the i-th feature map of T2 and the i-th feature map of Flair can be obtained. The feature prototype addressing module 322 recombines the similar prototypes selected from each sequence to obtain the recombined feature map of T1, the recombined feature map of T2 and the recombined feature map of Flair respectively.

[0083] The prototype addressing module performs targeted enhancement on the features of the acquired MRI images based on the prototype feature library, further highlighting the difference between the features corresponding to the normal area and the features corresponding to the abnormal area. The output of the prototype addressing module is a feature map of multiple sequences of similar prototypes that have been reorganized, which is more discriminative for normal and abnormal features. However, the recombination will cause the information of the abnormal position to be erased, which is very unfavorable for the later positioning of the abnormal area, that is, the identification of the lesion area. Therefore, as an embodiment, the prototype addressing module can merge the first recombined feature map and the first fused feature map to obtain a second fused feature map.

[0084] Reference Figure 4In the feature fusion process, the first fusion feature map can be the original multi-sequence information fusion feature map 440. The first recombined feature map can include the T1 recombined feature map 451, the T2 recombined feature map 452, and the Flair recombined feature map 453. The first fusion feature map and the first recombined feature map are merged to obtain the multi-sequence information fusion feature map 460 (i.e., the second fusion feature map). The multi-sequence information fusion feature map 460 is feature-enhanced based on the prototypes of the three sequences, and the distinction between normal features and abnormal features is strengthened through the recombined feature map, making it more conducive to identifying the lesion area of CVST.

[0085] S230. According to the first fusion feature map, identify the lesion area of the MRI image.

[0086] As an embodiment, the characteristic signals of the lesion area of the MRI image in the first fusion feature map can be enhanced and optimized before identifying the lesion area of the MRI image. Optionally, Figure 5 As an implementation manner of feature enhancement, the feature enhancement 500 can be implemented, for example, through a feature enhancement module, such as Figure 3 the multi-scale feature enhancement module 331 in Figure 5 The specific implementation manner of the feature enhancement 500 will be introduced below in combination with

[0087] As Figure 5 shown, the input of the feature enhancement module is the first fusion feature map. Referring to Figure 3 , the first fusion feature map can be the feature Figure 1 to the feature Figure 4 , the feature Figure 1 to the feature Figure 4 are the fusion feature maps of different scales of the three sequences. For example, they can be the feature fusion maps of the i-th stage. First, a sampling operation is performed on the i-th stage fusion feature map. Sampling can include, for example, upsampling and downsampling, such as upsampling and / or downsampling by a factor of two, four, or eight. Upsampling can be used to enlarge the image, and downsampling can be used to shrink the image. By sampling, the i-th stage fusion feature map is made to conform to the size of the display area, thereby obtaining a higher-quality image.

[0088] After the merged operation on the sampled image, convolution calculation is performed. Convolution is to restore the merged image to the scale size of the original input fusion feature map. That is to say, the feature enhancement module will enhance the input feature maps of different scales respectively without changing the size of the original feature map. The finally output is still the feature map of the original scale, but for the feature map of the original scale, the lesion area of CVST in the feature-enhanced MRI image will be more obvious.

[0089] As Figure 3As shown, the multi-scale feature enhancement module 331 outputs a feature map 331. This first feature enhancement map can go through the subsequent stages of Mask-RCNN, such as the object detection stage. The object detection stage can be subdivided into two subtasks, namely object detection and recognition. "Detection" is the first step in visual perception. It searches for every region of interest in the image as much as possible and marks them in the form of rectangular boxes. "Recognition" is similar to image classification and is used to determine the category of the target object in each region of interest. The object detection stage generates detection results for the MRI image. The object detection stage can include, for example, a Feature Pyramid Network 332, a Region Proposal Network 333, Region of Interest Alignment 334, Region of Interest Recognition 335, and Non-Maximum Suppression 336, and outputs the detection result 337 of the MRI image.

[0090] The Feature Pyramid Network (FPN) 332, whose input is, for example, the feature map 331. Without increasing the original computational cost, FPN can significantly improve the detection performance of the small-scale feature maps in the first feature enhancement map.

[0091] The Region Proposal Network (RPN) 333 can be responsible for generating candidate regions. For example, the input of RPN 333 can be an image of any shape, such as the feature map 332 output after the feature map 331 is processed by FPN 332. The output of RPN 333 can be a box-shaped object proposal region and object confidence, such as marking the features in the form of candidate boxes on the feature map 332. The candidate box can be outlined by a square, circle, ellipse, irregular polygon, etc., and the candidate box can be represented by data such as coordinates. That is to say, the feature map 333 output by RPN 333 can contain candidate boxes.

[0092] During the image processing process, if a specific region of the image is of interest, this region is called the Region of Interest (ROI). For example, various operators and functions commonly used in machine vision software such as Halcon, OpenCV, and Matlab can be used to obtain the ROI. After setting the ROI, the next operations can be performed on this region, such as Region of Interest Alignment 334 and Region of Interest Recognition 335. The alignment and recognition of the ROI can be implemented through relevant programs, for example. The feature map 335 output by ROI Recognition 335 contains the region of interest. After setting the region of interest, the influence of some redundant features can be reduced, the image processing time can be reduced, and the image accuracy can be increased.

[0093] Non - maximum suppression (NMS) 336 can be used to suppress redundant regions of interest. The suppression process is an iterative - traversal - elimination process. For example, the scores of all regions of interest can be sorted, and the highest - scoring one and its corresponding ROI are selected. Then, the remaining ROIs are traversed. If the overlapping area with the current highest - scoring ROI is greater than a certain threshold, the box is deleted. Then, continue to select the highest - scoring one from the unprocessed boxes and repeat the above process. Finally, the highest - scoring ROI and its corresponding score are selected. As an example, NMS 336 can output the possible lesion regions and their corresponding probabilities in the MRI image feature map, that is, the output of NMS 336 is the detection result 337 of the MRI image. Obviously, such an MRI image that can show the lesion region and the corresponding probability is easier for doctors to make identification and diagnosis, thus improving the recognition efficiency and diagnostic accuracy of CVST.

[0094] As an embodiment, the present application also proposes a method for training a learning model. The deep - learning model trained by this method can, through learning a large amount of sample data, achieve accurate detection of MRI images, thereby helping doctors with less experience to diagnose CVST patients more accurately.

[0095] Figure 6 It is a schematic flowchart of the method for training a learning model provided by an embodiment of the present application. Figure 7 It is the implementation scenario for training the learning model provided by an embodiment of the present application. The following combines Figure 7 to Figure 6 introduce the method in detail.

[0096] S610, determine the training sample images, and the training sample images include multi - sequence MRI images.

[0097] The sample images for training can contain multi - sequence MRI images. For example, they can include MRI images of different patients, such as normal or abnormal MRI images. The training samples can contain as many MRI images generated by different sequence scans as possible to ensure the integrity of the training samples.

[0098] Taking Figure 7 the training model shown as an example, its training sample images can be, for example, the multi - sequence MRI images 710 for training. The present application does not specifically limit the source of the training sample images. For example, they can be MRI images generated by a magnetic resonance instrument, or images obtained by processing the MRI images generated by a magnetic resonance instrument, such as the fusion of multi - sequence MRI images.

[0099] S620, perform data annotation on the training sample images to obtain the first annotation data.

[0100] Specifically, data annotation is performed on the training sample images. The way of data annotation can be, for example, manual annotation. For example, the lesion area is marked on the sample image by means of a marking block. The marked lesion area block is a kind of first annotation data. The first annotation data can be, for example, Figure 7 the annotation data 720 in

[0101] Data annotation can be implemented by a computer, for example. In this embodiment, the annotation method and implementation method of the training sample images are not specifically limited, as long as the required feature information can be annotated.

[0102] S630, preprocess the training sample images to obtain the first training data.

[0103] Optionally, the preprocessing process can be implemented by Figure 7 the preprocessing 730 in

[0104] Since the angles, directions, etc. of the MRI images scanned by different sequences are inconsistent in terms of spatial position, this will affect the training of the learning model. Through the image registration process, the angles, directions, etc. of the MRI images formed by different sequences are adjusted to be as consistent as possible, that is, they are matched in position, which is beneficial to the subsequent location positioning of the lesion area.

[0105] For the scanning device, such as a magnetic resonance instrument, there will be deviations in the pixels of the scanned images. For example, for the MRI images obtained by scanning the same sequence of the same patient, the pixel coordinates of the lesion area will be different. Also, due to device factors, the pixel coordinates of the scanned MRI images may be abnormal. Therefore, through intensity bias correction, the values of the pixel points of the training sample images can be adjusted to the normal range.

[0106] Intensity normalization can scale the intensity values of all training sample images to a certain range. For example, the pixel intensities of multi-sequence MRI images can be adjusted as a whole to the range of [-1, 1] or [0, 1], so that the intensity ranges are consistent, which is more conducive to image recognition.

[0107] Central cropping can be to crop the training sample images. For example, to remove redundant information such as the background and retain the image information that can represent the characteristic signals.

[0108] After the above preprocessing, the obtained first training data can, for example, reflect the characteristic signals of the MRI images.

[0109] S640. Train the learning model based on the first labeled data and the first training data to generate a detection result for identifying a lesion area from MRI images of different sequences.

[0110] Input the first labeled data obtained in step S620 and the first training data obtained in step S630 into the learning model for training. The learning model can be implemented, for example, Figure 7 by the deep learning model 740 in []. Input the labeled data 720 and the first training data output from the preprocessing 730 into the deep learning model 740 for training. The deep learning model 740 will perform one-to-one matching training on the first labeled data and the first training data until an accurate detection result is finally generated.

[0111] The learning model can complete the automatic detection of MRI images and output a detection result. For example, referring to Figure 7 when the input of the learning model is the test multi-sequence MRI image 750, it will output the detection result 760. Obviously, the larger the number of training sample images, the more accurate the diagnostic result obtained by the learning model. In other words, the learning model is a continuously updated process. As the first labeled data and the first training data input are continuously updated, the learning model will continuously improve the accuracy of its diagnosis. Obviously, the learning model can assist doctors in more quickly and accurately differentiating and diagnosing CVST patients.

[0112] Figure 8 is a schematic block diagram of a device for detecting MRI images provided in an embodiment of the present application. It should be understood that Figure 8 the shown device 800 for detecting MRI images is only an example, and the device 800 in the embodiment of the present application may further include other modules or units.

[0113] It should be understood that the device 800 can execute Figure 2 each step in the method of [], and for the sake of avoiding repetition, it will not be elaborated here.

[0114] In a possible implementation manner of the present application, the device 800 may include:

[0115] An extraction module 810, configured to obtain an MRI image to be detected, where the MRI image to be detected includes feature maps of multiple sequences of MRI images at different scales;

[0116] A fusion module 820, configured to fuse the feature maps of each scale of the multiple sequences of MRI images respectively to obtain a first fused feature map;

[0117] An identification module 830, configured to identify the lesion area of the MRI image according to the first fused feature map.

[0118] Optionally, the detection device 800 for MRI images further includes: a storage module, configured to obtain MRI images of different sequences, respectively construct prototype feature libraries corresponding to the MRI images of different sequences, classify and store the MRI images of different sequences in the prototype feature libraries, and the size of the prototype feature libraries is K, where K is an integer and K≥2;

[0119] An addressing module, configured to calculate the similarity between the feature maps of the MRI images of multiple sequences at different scales and each prototype feature map in the prototype feature libraries, and obtain N similar prototype feature maps, where N is an integer and N≥1;

[0120] A recombination module, configured to recombine the N prototype feature maps to obtain a first recombined feature map;

[0121] The fusion module 820 is further configured to merge the first fusion feature map and the first recombined feature map to obtain a second fusion feature map.

[0122] Optionally, the storage module further includes: performing similarity training on the prototype feature maps in the prototype feature libraries, and the similarity between each prototype feature map and other prototype feature maps is Ri, where R 1 >R 2 >……>R K , 1≤i≤K, when R 1 >R 2 + similarity distance D, the prototype feature libraries will reduce the similarity between the prototype feature maps.

[0123] Optionally, the MRI image detection device 800 further includes: a feature enhancement module, configured to enhance and optimize the feature signals of the lesion regions of the MRI images in the first fusion feature map, and identify the lesion regions of the MRI images.

[0124] It should be understood that the device 800 for detecting MRI images is embodied in the form of functional modules here. The term "module" here can be implemented in software and / or hardware forms, and no specific limitation is made thereto. For example, a "module" can be a software program, a hardware circuit, or a combination of the two that implements the above functions. The hardware circuit may include an application specific integrated circuit (ASIC), an electronic circuit, a processor (such as a shared processor, a proprietary processor, or a group of processors, etc.) for executing one or more software or firmware programs, a memory, a merging logic circuit, and / or other suitable components that support the described functions.

[0125] As an example, the device 800 for detecting MRI images provided in the embodiments of the present application may be a processor or a chip for performing the methods described in the embodiments of the present application.

[0126] Figure 9 is a schematic block diagram of a device 900 for training a learning model provided in an embodiment of the present application. It should be understood that the device 900 can execute Figure 6 each step in the method. To avoid repetition, details are not described here again. Figure 9 The device 900 shown may include:

[0127] A determination module 910, configured to determine training sample images, where the training sample images include multi-sequence MRI images;

[0128] An annotation module 920, configured to perform data annotation on the training sample images to obtain first annotation data;

[0129] A preprocessing module 930, configured to perform preprocessing on the training sample images to obtain first training data;

[0130] A training module 940, configured to train the learning model based on the first annotation data and the first training data to generate a detection result for identifying a lesion area from the MRI images of different sequences.

[0131] Optionally, the device 900 for training a learning model further includes: performing image registration, intensity bias correction, intensity normalization, central cropping, etc. on the training sample images.

[0132] Figure 10 is a schematic block diagram of a device 1000 for detecting MRI images provided in an embodiment of the present application. Figure 10 The device 1000 shown may include a memory 1010, a processor 1020, a communication interface 1030, and a bus 10404. Among them, the memory 1010, the processor 1020, and the communication interface 1030 are communicatively connected to each other through the bus 1040.

[0133] The memory 1010 may be a read only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 1010 may store a program. When the program stored in the memory 1010 is executed by the processor 1020, the processor 1020 is configured to execute each step of the method for detecting MRI images in the embodiments of the present application. For example, it may execute Figure 2 and Figure 6 each step of the embodiments shown.

[0134] The processor 1020 may adopt a general - purpose central processing unit (CPU), a microprocessor, an application - specific integrated circuit (ASIC), or one or more integrated circuits to execute relevant programs to implement the method for measuring bone density in the method embodiments of the present application.

[0135] The processor 1020 may also be an integrated circuit chip with the ability to process signals. In the implementation process, each step of the method for measuring bone density in the embodiments of the present application may be completed by the integrated logic circuit of the processor 1020 or instructions in the form of software.

[0136] The above - mentioned processor 1020 is a general - purpose processor, a digital signal processor (DSP), an application - specific integrated circuit (ASIC), a field - programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of the present application. The general - purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0137] The steps of the method disclosed in combination with the embodiments of the present application may be directly embodied as being executed and completed by a hardware decoding processor, or executed and completed by a combination of hardware and software modules in the decoding processor. The software module may be located in a mature storage medium in the art such as a random access memory, a flash memory, a read - only memory, a programmable read - only memory, or an electrically erasable programmable memory, a register, etc. This storage medium is located in the memory 1010. The processor 1020 reads the information in the memory 1010 and combines its hardware to complete the functions required to be executed by the units included in the device for measuring bone density in the embodiments of the present application, or executes the method for detecting MRI images in the method embodiments of the present application. For example, it can execute Figure 2 and Figure 3 each step / function of the embodiments shown.

[0138] The communication interface 1030 may use, but is not limited to, transceiver - type transceiver devices to implement the communication between the device 1000 and other devices or communication networks.

[0139] The bus 1040 may include a path for transmitting information between various components of the device 1000 (for example, the memory 1010, the processor 1020, the communication interface 1030).

[0140] It should be understood that the device 1000 shown in the embodiments of the present application may be a processor or a chip for executing the methods described in the embodiments of the present application.

[0141] It should be understood that the processor in the embodiments of the present application may be a central processing unit (CPU), and this processor may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or this processor may also be any conventional processor, etc.

[0142] It should be understood that the term "and / or" in this article is merely a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. These three situations, where A and B can be singular or plural. Additionally, the character " / " in this article generally represents an "or" relationship between the associated objects before and after, but it may also represent an "and / or" relationship. The specific meaning can be understood by referring to the context before and after.

[0143] In the present application, "at least one" means one or more, and "a plurality" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, or c can represent: a, b, c, a - b, a - c, b - c, or a - b - c, where a, b, and c can be single or multiple.

[0144] It should be understood that in various embodiments of the present application, the magnitudes of the sequence numbers of the above processes do not imply the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0145] The above are only the preferred embodiments of the present application and are not used to limit the present application. Any modifications, equivalent replacements, etc. made within the spirit and principles of the present application shall be included within the protection scope of the present application.

Claims

1. A method for detecting MRI images, which is executed by an MRI image detection model. The MRI image detection model is pre-trained based on a deep learning method. The MRI image detection method includes: Obtaining an MRI image to be detected, where the MRI image to be detected includes feature maps of MRI images of multiple different sequences at different scales; Fusing the feature maps of the MRI images of the multiple different sequences at each scale to obtain a first fused feature map; Identifying the lesion area of the MRI image according to the first fused feature map.

2. The method according to claim 1, wherein, The step of fusing the feature maps of the MRI images of the multiple different sequences at each scale to obtain a first fused feature map includes: Obtaining MRI images of different sequences, respectively constructing prototype feature libraries corresponding to the MRI images of different sequences. The prototype feature libraries classify and store the MRI images of different sequences. The size of the prototype feature library is K, where K is an integer and K≥2; Calculating the similarity between the feature maps of the MRI images of the multiple different sequences at different scales and each prototype feature map in the prototype feature library to obtain N similar prototype feature maps, where N is an integer and N≥1; Recombining the N prototype feature maps to obtain a first recombined feature map; The step of identifying the lesion area of the MRI image according to the first fused feature map includes: Identifying the lesion area of the MRI image according to a second fused feature map, where the second fused feature map is obtained by merging the first fused feature map and the first recombined feature map.

3. The method according to claim 2, wherein, The step of obtaining MRI images of different sequences and respectively constructing prototype feature libraries corresponding to the MRI images of different sequences includes: Perform similarity training on the prototype feature maps in the prototype feature library. The similarity between each prototype feature map and other prototype feature maps is Ri, where R 1 >R 2 >……>R K , 1 ≤ i ≤ K. When R 1 >R 2 + similarity distance D, the prototype feature library will reduce the similarity between the prototype feature maps.

4. The method according to claim 1, wherein, The step of identifying the lesion area of the MRI image according to the first fused feature map includes: Enhancing and optimizing the feature signals of the lesion area of the MRI image in the first fused feature map to identify the lesion area of the MRI image.

5. A method for training a learning model, wherein, includes: Determining a training sample image, where the training sample image includes a fused feature map, and the fused feature map is obtained by fusing feature maps of MRI images of multiple different sequences at different scales; Performing data annotation on the training sample image to obtain first annotation data; Performing preprocessing on the training sample image to obtain first training data; Training the learning model based on the first annotation data and the first training data to generate a detection result for identifying the lesion area of the MRI image of different sequences.

6. The method according to claim 5, wherein, The step of performing preprocessing on the training sample image includes: Performing image registration, intensity bias correction, intensity normalization, central cropping, etc. on the training sample image.

7. A device for detecting MRI images, wherein, including: an extraction module, configured to obtain an MRI image to be detected, where the MRI image to be detected includes feature maps of MRI images of multiple different sequences at different scales; a fusion module, configured to fuse the feature maps of the MRI images of the multiple different sequences at each scale to obtain a first fused feature map; an identification module, configured to identify a lesion area of the MRI image according to the first fused feature map.

8. The apparatus according to claim 7, wherein, it further includes: a storage module, configured to obtain MRI images of different sequences, respectively construct prototype feature libraries corresponding to the MRI images of different sequences, classify and store the MRI images of different sequences in the prototype feature libraries, and the size of the prototype feature library is K, where K is an integer and K≥2; an addressing module, configured to calculate the similarity between the feature maps of the MRI images of the multiple different sequences at different scales and each prototype feature map in the prototype feature library to obtain N similar prototype feature maps, where N is an integer and N≥1; a recombination module, configured to recombine the N prototype feature maps to obtain a first recombined feature map; the identification module, configured to identify a lesion area of the MRI image according to the first fused feature map, includes: identifying a lesion area of the MRI image according to a second fused feature map, where the second fused feature map is obtained by the fusion module by merging the first fused feature map and the first recombined feature map.

9. The apparatus according to claim 8, wherein, it includes: Perform similarity training on the prototype feature maps in the prototype feature library. The similarity between each prototype feature map and other prototype feature maps is Ri, where R 1 >R 2 >……>R K , 1 ≤ i ≤ K. When R 1 >R 2 + similarity distance D, the prototype feature library will reduce the similarity between the prototype feature maps.

10. The apparatus according to claim 7, wherein, it further includes: a feature enhancement module, configured to enhance and optimize the feature signals of the lesion area of the MRI image in the first fused feature map to identify the lesion area of the MRI image.

11. An apparatus for training a learning model, wherein, it includes: a determination module, configured to determine a training sample image, where the training sample image includes a fused feature map, and the fused feature map is obtained by fusing feature maps of MRI images of multiple different sequences at different scales; a labeling module, configured to perform data labeling on the training sample image to obtain first labeled data; a preprocessing module, configured to preprocess the training sample image to obtain first training data; a training module, configured to train the learning model based on the first labeled data and the first training data to generate a detection result for identifying a lesion area for the MRI images of different sequences.

12. The apparatus according to claim 11, wherein, it includes: performing image registration, intensity bias correction, intensity normalization, central cropping, etc. on the training sample image.

13. An apparatus for detecting an MRI image, wherein, it includes a processor and a memory, the memory is used to store program instructions, and the processor is used to call the program instructions to execute the method according to any one of claims 1 to 6.

14. A computer-readable storage medium, wherein, The computer-readable storage medium stores program instructions, and when the program instructions are run by a processor, the method described in any one of claims 1 to 6 is implemented.

Images