Perfusion image processing method, system, electronic device and storage medium

Abstract

The present invention provides a perfusion image processing method, system, electronic device and storage medium, the method comprising: acquiring perfusion image data, the perfusion image data comprising an image sequence scanned at different times; determining a target region of interest in the perfusion image data; acquiring an attention coefficient based on the signal intensity of the target region of interest; processing the perfusion image data and the attention coefficient based on a preset data processing model to obtain perfusion parameters and / or target region segmentation results of the perfusion image data. The present invention can improve the accuracy of perfusion parameter calculation and / or region segmentation of perfusion image data.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a perfusion image processing method, system, electronic equipment and storage medium. Background Art

[0002] Perfusion imaging is a method of continuous scanning of the region of interest when contrast agent is rapidly injected intravenously, thereby obtaining the time-density change relationship of the region of interest. It can more effectively and quantitatively reflect the changes in local tissue blood perfusion, which is of great significance in clarifying the blood supply to the lesion. It is widely used in the early detection of cerebral infarction, stroke examination and cerebral ischemia.

[0003] At present, the main processing process of perfusion imaging in clinical practice is to first use different mathematical models to calculate various perfusion parameter values, and then use the perfusion parameter values ​​to further combine domestic and international diagnostic standards to perform a series of threshold and imaging processing to obtain regional segmentation results. At present, the main methods for calculating perfusion imaging parameters are non-deconvolution and deconvolution. Among them, the non-deconvolution method assumes that the injection rate of the contrast agent is instantaneous, which is inconsistent with the actual situation; the deconvolution method has low computational efficiency, and the calculation process is heavily dependent on the accuracy of the intermediate results, which is easily interfered by artifacts and noise. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, an object of the present invention is to provide a perfusion image processing method, system, electronic device and storage medium to improve the accuracy of perfusion parameter calculation and / or region segmentation of perfusion image data.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] In a first aspect, the present invention provides a perfusion image processing method, comprising:

[0007] Acquiring perfusion image data, wherein the perfusion image data includes an image sequence scanned at different times;

[0008] determining a target region of interest in the perfusion image data;

[0009] Based on the signal strength of the target region of interest, obtaining an attention coefficient;

[0010] The perfusion image data and the attention coefficient are processed based on a preset data processing model to obtain perfusion parameters and / or target area segmentation results of the perfusion image data.

[0011] Preferably, the processing of the perfusion image data and the attention coefficient based on a preset data processing model to obtain the perfusion parameters and / or target area segmentation result of the perfusion image data includes:

[0012] Based on the perfusion image data and the attention coefficient, acquiring input data;

[0013] The input data is input into the data processing model for processing to obtain the perfusion parameters and / or target area segmentation results of the perfusion image data.

[0014] Preferably, before determining the target region of interest in the perfusion image data, the method further comprises:

[0015] The perfusion image data is registered.

[0016] Preferably, at least one of the following four methods is used to determine the target region of interest in the perfusion image data:

[0017] Determine the target region of interest based on a first preset algorithm;

[0018] Determine the target region of interest based on a first user input instruction;

[0019] Acquire a plurality of candidate regions of interest based on a second preset algorithm, and determine the target region of interest from the plurality of candidate regions of interest based on a second user input instruction;

[0020] An initial region of interest is determined based on a third user input instruction, and the initial region of interest is modified based on a third preset algorithm to obtain the target region of interest.

[0021] Preferably, the step of acquiring a plurality of candidate regions of interest based on a second preset algorithm includes:

[0022] Acquire a reference value of each voxel point in the perfusion image data according to a relationship between a signal intensity and time of each voxel point in the perfusion image data;

[0023] The plurality of candidate regions of interest are acquired according to the reference value of each voxel point in the perfusion image data.

[0024] Preferably, obtaining the attention coefficient based on the signal strength of the target region of interest includes:

[0025] Based on the signal strength of the target region of interest in the image sequence at each moment and the maximum image data of the target region of interest in all image sequences, the attention coefficient corresponding to the image sequence at each moment is obtained.

[0026] Preferably, the acquiring input data based on the perfusion image data and the attention coefficient comprises:

[0027] Based on the attention coefficients corresponding to all the image sequences at the previous t moments and all the image sequences at the previous t moments, the input data corresponding to the t-th moment is obtained.

[0028] In a second aspect, the present invention provides a perfusion image processing system, comprising:

[0029] An image data acquisition module, used to acquire perfusion image data, wherein the perfusion image data includes image sequences scanned at different times;

[0030] A region of interest acquisition module, used to determine a target region of interest in the perfusion image data;

[0031] An attention coefficient acquisition module, used to acquire an attention coefficient based on the signal strength of the target region of interest;

[0032] The model processing module is used to process the perfusion image data and the attention coefficient based on a preset data processing model to obtain the perfusion parameters and / or target area segmentation results of the perfusion image data.

[0033] In a third aspect, the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the perfusion image processing method as described above when executing the computer program.

[0034] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the perfusion image processing method as described above.

[0035] The positive and progressive effect of the present invention is that: the present invention determines the target region of interest in the perfusion image data, and obtains the attention coefficient based on the signal strength of the target region of interest; finally, the perfusion image data and the attention coefficient are processed based on a preset data processing model to obtain the perfusion parameters and / or the target region segmentation result of the perfusion image data. Since the perfusion image data is processed by combining machine learning and attention mechanism, the implicit connection between the regions inside the perfusion image can be fully explored, which helps the data processing model to pay attention to the data that is more important for the diagnostic task. Compared with the traditional deconvolution scheme, the computational efficiency is high, the interference by noise and artifacts is small, and the regional segmentation result can be directly obtained without obtaining the perfusion parameters in advance, so the segmentation result will not be affected by information loss, thereby avoiding diagnostic errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flow chart of a perfusion image processing method according to embodiment 1 of the present invention;

[0037] Figure 2A A perfusion image comparison diagram of different regions selected as regions of interest in Example 1 of the present invention;

[0038] Figure 2B This is a comparison diagram of different perfusion parameter results generated by selecting different regions of interest for the same perfusion image in Example 1 of the present invention;

[0039] Figure 2C A comparison diagram of different time density curves generated by selecting different regions of interest for the same perfusion image in Example 1 of the present invention;

[0040] Figure 3 It is a schematic diagram of the structure of a recurrent neural network;

[0041] Figure 4 This is a schematic diagram of the structure of the data processing model in Example 1 of the present invention;

[0042] Figure 5 is a schematic block diagram of a perfusion image processing system according to embodiment 2 of the present invention;

[0043] Figure 6 This is a schematic diagram of the structure of an electronic device according to Embodiment 3 of the present invention. DETAILED DESCRIPTION

[0044] The present invention is further described below by way of examples, but the present invention is not limited to the scope of the examples.

[0045] At present, the main processing process of perfusion imaging in clinical practice is to first use different mathematical models to calculate various perfusion parameter values, such as blood flow (Blood Flow, BF), blood volume (Blood Volume, BV), mean transit time (Mean Transit Time, MTT), time-to-maximum (Tmax), time-to-peak (TTP), etc.; then use the perfusion parameter values ​​to further combine domestic and international diagnostic standards to perform a series of thresholds and imaging processing, and finally obtain regional segmentation data such as the core infarct area (Core of the Infar), low perfusion areas (Low Perfusion Zones), and mismatch ratio (Mismatch ratio).

[0046] At present, the main methods for calculating perfusion imaging parameters are non-deconvolution and deconvolution. The non-deconvolution method calculates parameters based on the relationship between the speed of contrast agent accumulation in tissues and organs = arterial inflow speed - venous outflow speed. The deconvolution method was proposed by Cenic et al. in 1999 based on the concept of non-deconvolution. Since the non-deconvolution method assumes that the injection rate of the contrast agent is instantaneous, which is inconsistent with the actual situation, in order to obtain quantitative results of blood flow and average transit time, the actual injection rate of the contrast agent must be taken into account, and the time course data of each pixel position must be converted into the corresponding impulse residual function (Impulse Residue Function, IRF), or impulse response function (Impulse Response Function, IRF), to reflect the amount of contrast agent in the tissue over time after intravenous injection of contrast agent. The deconvolution method uses the residual function to calculate the venous outflow of contrast agents, comprehensively considering the inflow artery and outflow vein of the perfusion. When calculating BF, BV and MTT, there is no need to make assumptions about the underlying vascular system, which is close to the actual hemodynamics. The calculated perfusion parameters and function graphs can better reflect the actual situation inside the lesion. At present, the clinical application of perfusion effects is mainly in the research of acute cerebral ischemia and oncology. When cerebral ischemia occurs, local blood flow decreases. The purpose of perfusion is to understand the degree of ischemia and quantify the hemodynamic changes within the unit tissue, so as to evaluate the ischemic tissue and provide guidance for timely restoration of normal blood supply. At present, the mainstream perfusion imaging auxiliary diagnosis is based on the above-mentioned deconvolution method. Its advantages are that it is based on the physical model of hemodynamics and has strong interpretability. Its disadvantages are low computational efficiency, the calculation process is heavily dependent on the accuracy of the intermediate results, and it is easily interfered by artifacts and noise.

[0047] In addition, some real information will be lost in the process of obtaining regional segmentation results based on the parameters calculated by the mathematical model, which may affect the diagnosis.

[0048] Example 1

[0049] This embodiment provides a perfusion image processing method for processing a perfusion image to obtain perfusion parameters and / or target region segmentation results of perfusion image data. A perfusion image is an image obtained by perfusion imaging (PWI), such as a CT (Computed Tomography) perfusion (CTP) image. Depending on the tissue and organ of the perfusion imaging, the CTP image may include a brain CTP image, a liver CTP image, a heart CTP image, etc.

[0050] like Figure 1 As shown, the perfusion image processing method of this embodiment specifically includes the following steps:

[0051] S1, acquiring perfusion image data, wherein the perfusion image data includes an image sequence scanned at different times.

[0052] In this embodiment, the perfusion image data may be CT perfusion image data X obtained by scanning with a CT device, where X∈(T×F×G×H), T represents the number of CT images acquired during the CT scanning process, and F, G, and H represent the number of layers, width, and height of the CT data acquired at each moment, respectively.

[0053] S2, determining a target region of interest in the perfusion image data.

[0054] In this embodiment, the target region of interest is preferably a predetermined neighborhood of an arterial point in each perfusion image data.

[0055] S3, obtaining an attention coefficient based on the signal strength of the target area of ​​interest.

[0056] In this embodiment, the attention coefficient is used to characterize the importance of different image sequences in the perfusion image. For different computing tasks, the importance of each sequence varies. For example, when calculating the Tmax parameter, the sequence when the contrast agent just arrives in the perfusion image and the sequence when the contrast agent accumulation reaches the peak value are more important than other sequences.

[0057] S4, processing the perfusion image data and the attention coefficient based on a preset data processing model to obtain perfusion parameters and / or target area segmentation results of the perfusion image data. The specific process is as follows:

[0058] S41, acquiring input data based on the perfusion image data and the attention coefficient;

[0059] S42, inputting the input data into the data processing model for processing to obtain the perfusion parameters of the perfusion image data and / or the target area segmentation result.

[0060] In this embodiment, the data processing model is a pre-trained model, which uses perfusion parameters and / or target area segmentation results as prediction targets. Perfusion parameters include but are not limited to blood flow (BF), blood volume (BV) and time to peak value (Tmax) of residual contrast agent. Target area segmentation results include but are not limited to low perfusion zones (LPZ) and core of infarct (CoI). The specific selection can be set according to the actual needs of clinical diagnosis, and this embodiment does not make any specific restrictions on this.

[0061] This embodiment determines the target region of interest in the perfusion image data, and obtains the attention coefficient based on the signal strength of the target region of interest; then obtains input data based on the perfusion image data and the attention coefficient; finally, the perfusion image data and the attention coefficient are processed based on a preset data processing model to obtain the perfusion parameters and / or target region segmentation results of the perfusion image data. Since the perfusion image data is processed by combining machine learning and attention mechanisms, the implicit connections between regions within the perfusion image can be fully explored, which helps the data processing model to pay attention to data that is more important for the diagnostic task. Compared with traditional deconvolution schemes, the computational efficiency is high, the interference from noise and artifacts is small, and the regional segmentation results can be directly obtained without pre-acquiring the perfusion parameters, so the segmentation results will not be affected by information loss, thereby avoiding diagnostic errors.

[0062] In an optional implementation, before executing step S2, the method of this embodiment further includes: registering the perfusion image data.

[0063] Specifically, the image sequences at one or more moments in the perfusion image data are used as reference data, and other sequences in the perfusion image data are registered with the reference data.

[0064] In this embodiment, the specific registration method can be implemented by referring to any existing suitable registration method, and the details are not described again. The perfusion image data after registration is recorded as X reg ,Registration can reduce the individual differences of image sequences caused by ,motion and improve the accuracy of subsequent data processing.

[0065] In an optional implementation, step S2 may determine the target region of interest in the perfusion image data in at least one of the following four ways:

[0066] The first one is to determine the target region of interest based on a first preset algorithm, that is, to process the perfusion image data using a first estimation method to automatically output the target region of interest in the perfusion image data. This method saves time and effort.

[0067] The second method is to determine the target region of interest based on the first user input instruction, that is, to support the user to select a suitable target region of interest in an interactive manner, which can incorporate the doctor's experience.

[0068] The third method is to obtain several candidate regions of interest based on the second preset algorithm, and determine the target region of interest from the candidate regions of interest based on the second user input instruction. This method combines the advantages of the first and second methods, and can ensure that the target region of interest is selected correctly by manually intervening and calibrating the automatic output results.

[0069] The fourth method is to determine the initial region of interest based on the third user input instruction, and to correct the initial region of interest based on the third preset algorithm to obtain the target region of interest. This method also combines the advantages of the first and second methods, and can improve the accuracy of the target region of interest selection by automatically calibrating the manual selection results.

[0070] In this embodiment, the aforementioned first preset algorithm includes but is not limited to a threshold method, a key point detection method, and the like.

[0071] In an optional implementation, the specific process of obtaining several candidate regions of interest based on the second preset algorithm is as follows:

[0072] First, according to the relationship between the signal intensity and time of each voxel point in the perfusion image data, a reference value of each voxel point in the perfusion image data is obtained; then, according to the reference value of each voxel point in the perfusion image data, the candidate regions of interest are obtained. For example, regions with reference values ​​greater than a predetermined reference threshold are respectively regarded as arterial points, and the predetermined neighborhood of each arterial point is determined as a candidate region of interest. Specifically, according to the relationship between the signal intensity and time of each voxel point in the perfusion image data, the process of obtaining the reference value of each voxel point in the perfusion image data is as follows:

[0073] Obtaining the maximum signal intensity corresponding to each voxel point in the perfusion image data and the average value of the maximum signal intensity corresponding to all voxel points;

[0074] Acquire the time when each voxel point in the perfusion image data reaches the maximum signal intensity and the average value of the time when all voxel points reach the maximum signal intensity;

[0075] The reference value of each voxel point in the perfusion image data is obtained according to the maximum signal intensity corresponding to each voxel point in the perfusion image data and the average of the maximum signal intensity corresponding to all voxel points, as well as the time when each voxel point reaches the maximum signal intensity and the average of the time when all voxel points reach the maximum signal intensity.

[0076] Taking the above-mentioned CT perfusion image data X, X∈(T×F×G×H) as an example, when obtaining the reference value, firstly, the maximum signal intensity h(i,j,k) corresponding to each voxel point in the perfusion image data and the mean value of the maximum signal intensity corresponding to all voxel points are obtained. Where i, j, k represent the layer position, width position and height position of the corresponding voxel point in the perfusion image data, h = max (X, axis = 0), F, G, and H represent the number of layers, width, and height of the perfusion image data, respectively; then, the time a(i, j, k) at which each voxel point in the perfusion image data reaches the maximum signal intensity and the mean value of the time at which all voxel points reach the maximum signal intensity are obtained. in, Then, according to the maximum signal intensity h(i,j,k) corresponding to each voxel point in the perfusion image data and the mean maximum signal intensity corresponding to all voxel points And the time a(i,j,k) when each voxel reaches the maximum signal intensity and the mean of the time when all voxel points reach the maximum signal intensity The reference value c(i,j,k) of each voxel point in the perfusion image data is obtained, and the calculation formula is as follows:

[0077] In this embodiment, the maximum signal intensity h(i, j, k) corresponding to each voxel point in the perfusion image data and the average maximum signal intensity corresponding to all voxel points are calculated. And the time a(i,j,k) when each voxel reaches the maximum signal intensity and the mean of the time when all voxel points reach the maximum signal intensity The reference value c(i, j, k) of each voxel point obtained can reflect the brightness of the voxel point and the time when the signal intensity reaches the peak. The voxel points in the predetermined neighborhood of the arterial point are usually the brightest and reach the signal intensity peak the earliest. This embodiment can determine the candidate region of interest where the arterial point is located by integrating the two conditions, where k 1 ,k 2 In order to balance the importance and dimension weight of the two conditions, they can be set according to the needs, for example, k 1 =1.0, k 2 =3.5.

[0078] In an optional implementation, step S3 is specifically implemented by the following steps: based on the signal strength of the target region of interest in the image sequence at each moment and the maximum image data of the target region of interest in all image sequences, obtaining the attention coefficient corresponding to the image sequence at each moment.

[0079] Assume that the signal strength set S of the target region of interest (such as arterial point) in the image sequence at each moment is [s 1 ,s 2 ,…,s T ], where s i represents the signal strength of the target interest point in the i-th image sequence, T represents the number of the image sequences, then the attention coefficient f corresponding to the i-th image sequence iThe calculation formula is as follows: Among them, max(S) represents the maximum signal strength of each target point of interest in all image sequences, that is, s 1 ~s T For example, Figure 3 As shown, the AIF curve is a curve of the relationship between the arterial point signal strength set S and time, and the peak value of the curve is the maximum signal strength corresponding to the arterial point.

[0080] The above attention coefficient f i The following upper triangular matrix is ​​constructed to describe the entire perfusion process

[0081]

[0082] The change process of the arterial point affected by the contrast agent. The first row corresponds to the perfusion image acquired at the first moment, which can be regarded as the attention distribution of the sequence data at the first moment; the second row corresponds to the perfusion image acquired at the second moment, which can be regarded as the attention distribution of the sequence data at the second moment; and the distribution of the remaining moments is similar.

[0083] This embodiment converts the signal intensity of the target region of interest in the image sequence at each moment into a corresponding probability as the attention coefficient of the image sequence at each moment, which can well reflect the importance of the corresponding sequence.

[0084] In an optional implementation, step S41 is specifically implemented by the following steps:

[0085] Based on the attention coefficients corresponding to all the image sequences at the previous t moments and all the image sequences at the previous t moments, the input data corresponding to the tth moment is obtained. The specific calculation formula is: Among them, x i represents the i-th image sequence, I t represents the input data at the tth moment, and T represents the number of image sequences.

[0086] This embodiment can accurately reflect the characteristics of all image data at time t and before time t by weighted summing up the attention coefficients corresponding to all image sequences before time t.

[0087] In an optional embodiment, the data processing model is constructed based on a neural network and trained based on sample perfusion image data and corresponding gold standard data, wherein the gold standard data includes perfusion parameter gold standard data and / or target area segmentation gold standard data. The sample perfusion image data can be obtained by collecting a large amount of CT perfusion scan clinical data. The perfusion parameter gold standard data and the target area segmentation gold standard data are specifically determined according to the prediction target.

[0088] For example, when the perfusion parameters to be predicted by the data processing model include BF, the gold standard data of the perfusion parameters used in training the model should include the gold standard data of blood flow, denoted as Y BF ; When the perfusion parameters to be predicted by the data processing model include BV, the gold standard data of perfusion parameters used in training the model should include the gold standard data of blood volume, denoted as Y BV When the perfusion parameters to be predicted by the data processing model include Tmax, the gold standard data of the perfusion parameters used in training the model should include the gold standard data of the time when the residual amount of contrast agent reaches the peak value, denoted as Y Tmax Similarly, when the target region segmentation result to be predicted by the data processing model includes LPZ, the target region segmentation gold standard data used in training the model should include the low perfusion region gold standard data, denoted as Y LPZ ; When the target region segmentation result to be predicted by the data processing model includes CoI, the target region segmentation gold standard data used in training the model should include the core infarct region gold standard data, denoted as Y CoI Of course, if the data processing model can also predict other perfusion parameters or other target areas, then the corresponding gold standard data should be used when training the model.

[0089] In this embodiment, the gold standard data can be calculated by the deconvolution method. In order to ensure that the calculation results reflect the actual perfusion parameters and target areas as much as possible, the entire calculation process is manually calibrated to ensure that the arterial points or venous points are selected correctly and the influence of artifact data is excluded. Using the gold standard data to train the model can ensure the accuracy of the model prediction. In the process of constructing the training set, the same sample perfusion image can generate different gold standard data according to different regions of interest. Figures 2A to 2C The experimental results shown in the figure show that when different regions of interest are selected (see Figure 2A When different time-density curves are obtained (as indicated by the arrow in the middle), different perfusion parameter result graphs will be generated.

[0090] In an optional embodiment, considering that the perfusion image data contains both time information (T) and spatial information (number of layers F, width G, height H), the neural network uses a recurrent neural network, preferably a recurrent neural network with a recursive structure that uses sequence data as input, including but not limited to RNN (Recurrent Neural Network), LSTM (Long Short-Term Memory), bidirectional RNN, bidirectional LSTM, etc. Among them, LSTM or bidirectional LSTM structure is preferred, and abnormal data caused by motion artifacts in CTP sequence are selectively excluded by gating. This method utilizes the memory characteristics of the recurrent neural network, combines the spatial information of the perfusion image data with the time information, encodes and maps the perfusion image data with time series characteristics, thereby quickly and efficiently generating perfusion parameters and target area segmentation results, simplifying the calculation process, improving the calculation speed, and reducing the loss of information contained in the original perfusion image data.

[0091] In this embodiment, the recurrent neural network is mainly composed of three convolutional neural networks U, V, and W. Figure 4 The cyclic structure of the convolutional neural network U, V, W is shown. The input data I at each time point is t After the convolutional neural network U is mapped to the hidden layer, it is recorded as I′ t , I′ t On the one hand, as a feature containing the current moment and the previous moment, it is input into the convolutional neural network W to obtain the encoded feature s t ,s t and the attention coefficient f at the corresponding moment t After multiplication, it is used as the input of the next cycle, I′ t On the other hand, compared with the feature s encoded at the previous moment t-1 The bitwise sum operation is used as the input of the convolutional neural network V to obtain the predicted target Y′. The following is an explanation of the loop process: From time point t 0 Starting from, the convolutional neural network U will 0 The corresponding input data I 0 Mapped to the hidden layer, we get I′ 0 , I′ 0 Input into convolutional neural networks W and V at the same time, convolutional neural network W for I′ 0 Encode and get feature s 0 Then at time t 1 , the convolutional neural network U will 1 Corresponding perfusion space data I 1 Mapped to the hidden layer, we get I′ 1 , I′ 1Input into convolutional neural networks W and V at the same time, convolutional neural network W for I′ 1 Encode and get feature s 1 ,s 1 With s 0 The bitwise sum operation is used as the input of the convolutional neural network V to obtain the predicted target Y′. And so on until the time point t N-1 The loop ends.

[0092] based on Figure 4 The model structure of step S42 can be specifically implemented by the following steps: inputting the input data into the recurrent neural network in the order of scanning time, multiplying the output data of the hidden layer of the recurrent neural network at each moment with the attention coefficient corresponding to the corresponding moment together with the input data of the next time point as the input of the next cycle of the recurrent neural network, and finally outputting the perfusion parameters and the target area segmentation result.

[0093] In this embodiment, a recurrent neural network model suitable for perfusion image data processing is given in combination with the characteristics of perfusion image data containing temporal and spatial information. Different from the traditional RNN structure, the recurrent structure of the recurrent neural network in this method is completed in the hidden layer, mapping the input data to the feature space, which is conducive to removing noise and artifacts in the image and improving the robustness of the mapping network.

[0094] In an optional manner, the data processing model includes a plurality of single-task data processing models, each of which is used to predict and output one of the perfusion parameters and the target area segmentation results based on the input data. For example, a first data processing model for predicting only blood flow, a second data processing model for predicting only blood volume, a third data processing model for predicting only the time when the residual amount of contrast agent reaches the peak, a fourth data processing model for predicting only low perfusion areas, and a fifth data processing model for predicting only the core infarct area. The advantage of using a single-task data processing model is that training a single-task data processing model is faster and simpler. In addition, since each single-task data processing model predicts different targets, the sample data used can be different, and sample data that is more targeted to the predicted target can be selected for training.

[0095] In an optional manner, considering the strong correlation between the various parameter maps and the time efficiency of the model in the application stage, a multi-task approach is used in actual applications to implement the data processing model. The data processing model includes a multi-task data processing model, which is used to simultaneously predict and output perfusion parameters and target area segmentation results based on the input perfusion image data.

[0096] In one implementation, the multi-task data processing model may be a model trained based on sample perfusion image data of all perfusion parameters and target region segmentation results to be predicted and gold standard data of all perfusion parameters and gold standard data of target region segmentation results. Figure 4 The model shown is a multi-task data processing model, including the convolutional neural networks U, V, and W as described above, and also includes a feature mapping network for the prediction target, which is used to map the output results of the convolutional neural network V to the corresponding prediction results. The figure shows the feature mapping network H 1 , H 2 , H 3 , H 4 , H 5 , which correspond to blood flow, blood volume, peak time of residual contrast agent, low perfusion area, and core infarction area, respectively. The output of the convolutional neural network V is mapped to the blood flow prediction result Y′ BF , blood volume prediction result Y′ BV , the prediction result of the time when the residual amount of contrast agent reaches the peak value Y′ Tmax , prediction result of low perfusion area Y′ LPZ , prediction result of core infarction area Y′ CoI .

[0097] The loss calculated from the output at each time point and the gold standard data The expression is shown in formula (1):

[0098]

[0099] where w i ∈[0,1] is the weight of the loss between the output and the gold standard data at each time point, set to w i =(i+1) / N,i=0,1,…,N-1.

[0100] The loss function expression corresponding to blood flow is shown in formula (2):

[0101]

[0102] The loss function expression corresponding to blood volume is shown in formula (3):

[0103]

[0104] The loss function expression corresponding to the time when the residual amount of contrast agent reaches the peak value is shown in formula (4):

[0105]

[0106] The loss function expression corresponding to the low perfusion area is shown in formula (5):

[0107]

[0108] The loss function expression corresponding to the core infarct area is shown in formula (6):

[0109]

[0110] in, and They are the L1 loss and TV (Total Variation) loss of the constraint parameter graph mapping respectively.

[0111] The L1 loss expression is shown in formula (7):

[0112]

[0113] The TV loss expression is shown in formula (8):

[0114]

[0115] Among them, δ is a hyperparameter used to adjust the smoothness of the parameter map generated by the data; and is the loss function for constraining the segmentation of infarct and hypoperfusion areas;

[0116] The expression of is shown in formula (9):

[0117]

[0118] The expression of is shown in formula (10):

[0119]

[0120] In formula (10), S is the set of coordinate points of the predicted image in three-dimensional space, P′ is the predicted probability output by the regional segmentation network, Y is the gold standard data of the core infarct area or low perfusion area, and α and γ are hyperparameters.

[0121] During the training process, the Adam (a random optimization algorithm) optimizer is used to train the entire neural network.

[0122] In another implementation, the multi-task data processing model can also be a model formed by fusing multiple single-task data processing models. The single-task data processing model trained in the previous text is first used, and then all single-task data processing models are fused through the Boosting method and the Bagging method.

[0123] Example 2

[0124] This embodiment provides a perfusion image processing system, such as Figure 5 As shown, including:

[0125] An image data acquisition module 11, used to acquire perfusion image data, wherein the perfusion image data includes image sequences scanned at different times;

[0126] A region of interest acquisition module 12, used to determine a target region of interest in the perfusion image data;

[0127] An attention coefficient acquisition module 13, used to acquire an attention coefficient based on the signal strength of the target region of interest;

[0128] The model processing module 14 is used to process the perfusion image data and the attention coefficient based on a preset data processing model to obtain the perfusion parameters of the perfusion image data and / or the target area segmentation result.

[0129] Preferably, the model processing module 14 includes:

[0130] An input data acquisition unit, configured to acquire input data based on the perfusion image data and the attention coefficient;

[0131] The model processing unit is used to input the input data into the data processing model for processing to obtain the perfusion parameters and / or target area segmentation results of the perfusion image data.

[0132] Preferably, the system further comprises: a registration module, configured to register the perfusion image data before the region of interest acquisition module 12 determines the target region of interest in the perfusion image data.

[0133] Preferably, the region of interest acquisition module 12 determines the target region of interest in the perfusion image data by at least one of the following four methods:

[0134] Determine the target region of interest based on a first preset algorithm;

[0135] Determine the target region of interest based on a first user input instruction;

[0136] Acquire a plurality of candidate regions of interest based on a second preset algorithm, and determine the target region of interest from the plurality of candidate regions of interest based on a second user input instruction;

[0137] An initial region of interest is determined based on a third user input instruction, and the initial region of interest is modified based on a third preset algorithm to obtain the target region of interest.

[0138] Preferably, the step of acquiring a plurality of candidate regions of interest based on a second preset algorithm includes:

[0139] Acquire a reference value of each voxel point in the perfusion image data according to a relationship between a signal intensity and time of each voxel point in the perfusion image data;

[0140] The plurality of candidate regions of interest are acquired according to the reference value of each voxel point in the perfusion image data.

[0141] Preferably, the attention coefficient acquisition module 13 is specifically used for:

[0142] Based on the signal strength of the target region of interest in the image sequence at each moment and the maximum signal strength of the target region of interest in all image sequences, the attention coefficient corresponding to the image sequence at each moment is obtained.

[0143] Preferably, the input data acquisition unit is specifically used for:

[0144] Based on the attention coefficients corresponding to all the image sequences at the previous t moments and all the image sequences at the previous t moments, the input data corresponding to the t-th moment is obtained.

[0145] Preferably, the data processing model is trained based on sample perfusion image data and corresponding gold standard data, and the gold standard data includes perfusion parameter gold standard data and / or target area segmentation gold standard data.

[0146] This embodiment determines the target region of interest in the perfusion image data, and obtains the attention coefficient based on the signal strength of the target region of interest; then processes the perfusion image data and the attention coefficient based on a preset data processing model to obtain the perfusion parameters and / or the target region segmentation result of the perfusion image data. Since this embodiment combines machine learning and attention mechanisms to process the perfusion image data, it is possible to fully explore the implicit connections between regions within the perfusion image, which helps the data processing model to focus on data that is more important for the diagnostic task. Compared with traditional deconvolution schemes, it has high computational efficiency and is less affected by noise and artifacts. It can also directly obtain regional segmentation results without obtaining perfusion parameters in advance, so that the segmentation results will not be affected by information loss, thereby avoiding diagnostic errors.

[0147] Example 3

[0148] Figure 6 The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the method of Embodiment 1 is implemented. Figure 6The electronic device 30 shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

[0149] The electronic device 30 may be in the form of a general-purpose computing device, for example, it may be a server device. The components of the electronic device 30 may include, but are not limited to: at least one processor 31, at least one memory 32, and a bus 33 connecting different system components (including the memory 32 and the processor 31).

[0150] The bus 33 includes a data bus, an address bus, and a control bus.

[0151] The memory 32 may include a volatile memory, such as a random access memory (RAM) 321 and / or a cache memory 322 , and may further include a read-only memory (ROM) 323 .

[0152] The memory 32 may also include a program / utility 325 having a set (at least one) of program modules 324, such program modules 324 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0153] The processor 31 executes various functional applications and data processing, such as the method of embodiment 1 of the present invention, by running the computer program stored in the memory 32 .

[0154] The electronic device 30 may also communicate with one or more external devices 34 (e.g., keyboards, pointing devices, etc.). Such communication may be performed via an input / output (I / O) interface 35. Furthermore, the model-generated device 30 may also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) via a network adapter 36. As shown, the network adapter 36 communicates with other modules of the model-generated device 30 via a bus 33. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the model-generated device 30, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (RAID) systems, tape drives, and data backup storage systems.

[0155] It should be noted that although several units / modules or sub-units / modules of the electronic device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to an embodiment of the present invention, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided into multiple units / modules to be embodied.

[0156] Example 4

[0157] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the method of Embodiment 1 are implemented.

[0158] The readable storage medium may include but is not limited to: a portable disk, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical storage device, a magnetic storage device or any suitable combination of the above.

[0159] In a possible implementation manner, the present invention may also be implemented in the form of a program product, which includes program code. When the program product is run on a terminal device, the program code is used to enable the terminal device to execute the steps of the method for implementing Embodiment 1.

[0160] Among them, the program code for executing the present invention can be written in any combination of one or more programming languages, and the program code can be executed completely on the user device, partially on the user device, as an independent software package, partially on the user device and partially on a remote device, or completely on the remote device.

[0161] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that this is only for illustration and the protection scope of the present invention is defined by the appended claims. Those skilled in the art may make various changes or modifications to these embodiments without departing from the principles and essence of the present invention, but these changes and modifications all fall within the protection scope of the present invention.

Claims

1. A perfusion image processing method, characterized in that: include: Acquiring perfusion image data, wherein the perfusion image data includes an image sequence scanned at different times; determining a target region of interest in the perfusion image data; Based on the signal strength of the target region of interest, obtaining an attention coefficient; Wherein, the attention coefficient is used to characterize the importance of different image sequences in the perfusion image; The perfusion image data and the attention coefficient are processed based on a preset data processing model to obtain perfusion parameters and / or target area segmentation results of the perfusion image data.

2. The perfusion image processing method according to claim 1, characterized in that: The processing of the perfusion image data and the attention coefficient based on a preset data processing model to obtain the perfusion parameters and / or the target area segmentation result of the perfusion image data includes: Based on the perfusion image data and the attention coefficient, acquiring input data; The input data is input into the data processing model for processing to obtain the perfusion parameters and / or target area segmentation results of the perfusion image data.

3. The perfusion image processing method according to claim 1, wherein: Before determining the target region of interest in the perfusion image data, the method further includes: The perfusion image data is registered.

4. The perfusion image processing method according to claim 1, wherein: The target region of interest in the perfusion image data is determined by at least one of the following four methods: Determine the target region of interest based on a first preset algorithm; Determine the target region of interest based on a first user input instruction; Acquire a plurality of candidate regions of interest based on a second preset algorithm, and determine the target region of interest from the plurality of candidate regions of interest based on a second user input instruction; An initial region of interest is determined based on a third user input instruction, and the initial region of interest is modified based on a third preset algorithm to obtain the target region of interest.

5. The perfusion image processing method according to claim 4, characterized in that: The step of acquiring a plurality of candidate regions of interest based on a second preset algorithm includes: Acquire a reference value of each voxel point in the perfusion image data according to a relationship between a signal intensity and time of each voxel point in the perfusion image data; The plurality of candidate regions of interest are acquired according to the reference value of each voxel point in the perfusion image data.

6. The perfusion image processing method according to claim 1, wherein: The acquiring the attention coefficient based on the signal strength of the target region of interest includes: Based on the signal strength of the target region of interest in the image sequence at each moment and the maximum signal strength of the target region of interest in all image sequences, the attention coefficient corresponding to the image sequence at each moment is obtained.

7. The perfusion image processing method according to claim 2, characterized in that: The step of acquiring input data based on the perfusion image data and the attention coefficient includes: Based on the attention coefficients corresponding to all the image sequences at the previous t moments and all the image sequences at the previous t moments, the input data corresponding to the t-th moment is obtained.

8. A perfusion image processing system, characterized in that: include: An image data acquisition module, used to acquire perfusion image data, wherein the perfusion image data includes image sequences scanned at different times; A region of interest acquisition module, used to determine a target region of interest in the perfusion image data; An attention coefficient acquisition module, used to acquire an attention coefficient based on the signal strength of the target region of interest; Wherein, the attention coefficient is used to characterize the importance of different image sequences in the perfusion image; The model processing module is used to process the perfusion image data and the attention coefficient based on a preset data processing model to obtain the perfusion parameters and / or target area segmentation results of the perfusion image data.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images