TWI870195B - Image processing system and method thereof for early detection of ischemic core of ischemic stroke patient - Google Patents

Abstract

An image processing method thereof for early detection of ischemic core of ischemic stroke patient includes a model training phase and a model inference phase. In the model training phase, it first generates a first pseudo diffusion-weighted MR imaging (DWI) and a generator loss according to a non-contrast enhanced CT (NCCT) image for training; then, it generates a discriminator loss and a discriminator result according to a DWI for training and the first pseudo DWI; then, it judges whether to end the model training phase according to the generator loss and the discriminator loss, and if so, ends the model training phase and obtains a set of parameters for inference phase, and if not, updates the parameters used in the model training phase. In the model inference phase, it generates a pseudo DWI according to a NCCT image and the set of parameters for inference phase.

Description

Image processing system and method for early detection of cerebral infarction core in patients with ischemic cerebral stroke

The present invention relates to a medical image processing system and method, and more particularly to an image processing system and method for early detection of cerebral infarction core of ischemic stroke patients.

In the treatment of acute ischemic stroke, early detection of the infarct core is very important. The sooner the diagnosis and treatment, the better the patient's recovery.

Non-Contrast Enhanced CT (NCCT) is the first-line examination for acute stroke and is widely used to rule out cerebral hemorrhage. Although it can also be used to evaluate the infarction core of acute ischemic stroke, it has problems such as high subjectivity of physicians' interpretation, low accuracy, and large variability in interpretation.

There are two types of imaging tools used clinically to detect the infarct core, namely diffusion-weighted MR imaging (DWI) and CT perfusion (CTP) imaging. However, each of these imaging tools has its own shortcomings. Magnetic resonance imaging diffusion-weighted imaging is highly accurate. Although it is the gold standard for diagnosing the infarct core, its equipment is expensive, its accessibility is low, and its scanning time is long. Only very few hospitals have the ability to use it for patients with acute cerebral stroke. The disadvantages of tomographic cerebral blood perfusion imaging are the instability of quantification of cerebral infarction core, the need to inject contrast agent, high radiation dose, and the need to use expensive commercial software (e.g., RAPID.ai) for post-processing operations. In addition, its accuracy in diagnosing cerebral infarction core is not as good as that of magnetic resonance imaging diffusion-weighted imaging.

Known technologies, such as Patent No. I542328 and Patent No. I725813 of the Republic of China, are all based on the processing and judgment of diffusion-weighted magnetic resonance imaging, which has a high cost and a long scanning time. Therefore, there is an urgent need for an image processing technology with a low cost, which can be widely used by various hospitals and can accelerate the diagnosis of acute ischemic stroke.

Therefore, the purpose of the present invention is to provide an image processing system for early detection of cerebral infarction core of ischemic stroke patients, which includes a memory storing at least one instruction, a database, and a processor communicatively coupled to the memory and the database.

The database includes at least one training data set, wherein the training data set includes a training non-revelation brain tomography image and a training MRI diffusion-weighted image corresponding to the training non-revelation brain tomography image.

The processor is used to access and execute the instruction to operate in a model training phase and a model inference phase.

The model training phase includes a generative model operation, an identification model operation, a judgment step, and a parameter updating step. The generative model operation is used to perform a generative model learning operation using complex parameters based on the training non-imaging brain sectional scan image to generate a first virtual MRI diffusion weighted image and a generation error; the identification model operation is used to perform an identification model learning operation using complex parameters based on the training MRI diffusion weighted image and the first virtual MRI diffusion weighted image to generate an identification error and an identification result; The judgment step is used to judge whether to terminate the model training phase according to the generation error and the identification error. If the result of the judgment step is yes, the model training phase is terminated, and the parameters currently used in the generation model operation are used as an inference phase parameter set; the parameter update step is used to update the parameters used in the generation model operation and the identification model operation, wherein the parameter update step is performed only if the result of the judgment step is no.

In the model inference phase, the processor generates a virtual MRI diffusion-weighted image based on a non-contrast brain CT scan image and using the inference phase parameter set.

Another object of the present invention is to provide an image processing method for early detection of cerebral infarction core in ischemic stroke patients, which includes a model training stage and a model inference stage.

The model training phase includes a generative model operation, an identification model operation, a judgment step, and a parameter updating step. The generative model operation is used to perform a generative model learning operation based on a training non-imaging brain sectional scan image using complex parameters to generate a first virtual MRI diffusion-weighted image and a generation error; the identification model operation is used to perform an identification model learning operation based on a training MRI diffusion-weighted image and the first virtual MRI diffusion-weighted image using complex parameters to generate an identification error and an identification result. ; The judging step is used to judge whether to terminate the model training phase according to the generation error and the identification error. If the result of the judging step is yes, the model training phase is terminated, and the parameters currently used in the generation model operation are used as an inference phase parameter set; The parameter updating step is used to update the parameters used in the generation model operation and the identification model operation. If the result of the judging step is no, the parameter updating step is performed.

In the model inference stage, a virtual MRI diffusion-weighted image is generated based on a non-contrast brain CT image and using the inference stage parameter set.

The utility of the present invention is that after the non-development brain sectional scan image passes through the core image processing system and method for early detection of cerebral infarction in ischemic stroke patients, the virtual MRI diffusion-weighted image can be generated. The patient does not need to undergo expensive and time-consuming MRI diffusion-weighted imaging, nor does he need to inject contrast agent and bear the risk of high radiation dose to undergo sectional cerebral blood flow perfusion imaging. The virtual MRI diffusion-weighted image can be quickly obtained only after the first-line examination (obtaining the non-development brain sectional scan image).

The above-mentioned other technical contents, features and effects of the present invention will be clearly presented in the following detailed description of two preferred embodiments with reference to the drawings.

Please refer to FIG. 1 , a first preferred embodiment of the image processing system 1 for early detection of cerebral infarction core of ischemic stroke patients of the present invention comprises a memory 11 storing at least one instruction, a database 12, and a processor 13 communicatively coupled to the memory 11 and the database 12. The image processing system 1 is operated in conjunction with an input device 2 and an output device 3. The input device 2 can be a non-development brain tomography scanning device to provide a non-development brain tomography scanning image to the image processing system 1 for processing, and the output device 3 can be a screen (but not limited thereto) to provide the result of the image processing system 1 to a user.

The database 12 includes at least one pre-established training data set, wherein the training data set includes a training non-revelation brain tomography image and a training MRI diffusion-weighted image corresponding to the training non-revelation brain tomography image. Furthermore, the training non-imaging brain tomography image and the training MRI diffusion-weighted image in each training data set are paired images obtained from actual examinations of the same patient within two hours. In the first preferred embodiment, the patient first undergoes a conventional non-imaging brain tomography to obtain the training non-imaging brain tomography image, and then immediately undergoes diffusion-weighted imaging to obtain the training MRI diffusion-weighted image. The database 12 is established using this paired image, and the training MRI diffusion-weighted image can be regarded as a ground truth.

Among them, the processor 13 is used to access and execute the instructions in the memory 11 to perform the image processing method of the present invention for early detection of the core of cerebral infarction in patients with ischemic cerebral stroke. Furthermore, the image processing method for early detection of the core of cerebral infarction in patients with ischemic cerebral stroke includes a model training stage and a model inference stage.

Please refer to Figures 1 and 2. The following will further explain the operations and steps performed during the model training phase.

In a generative model operation 41, the processor 13 performs a generative model learning operation based on the training non-imaged brain sectional scan image pre-established in the database 12 to generate a first pseudo-DWI diffusion-weighted image and a generative error (G-Loss), wherein the generative error is the error between the first pseudo-DWI diffusion-weighted image and the training DWI diffusion-weighted image. The generative model learning operation may adopt an existing autoencoder (AE) architecture or an existing variational autoencoder (VAE) architecture. In this first preferred embodiment, a self-encoder architecture is used, which includes at least one encoder layer (Encoder Layer) operation, at least one bottleneck layer (Bottleneck Layer, or hidden layer) operation, and at least one decoder layer (Decoder Layer) operation, each layer corresponds to a plurality of nodes, and each node corresponds to a parameter; in other words, the processor 13 is based on the non-developmental brain tomography scan image used for training, and uses the parameters to perform the generative model operation.

In a discriminative model operation 42, the processor 13 performs a discriminative model learning operation based on the first virtual MRI diffusion weighted image and the training MRI diffusion weighted image pre-established in the database 12 to generate a discriminative error (D-Loss) and a discriminative result; in the first preferred embodiment, the discriminative error is the error in determining the first virtual MRI diffusion weighted image. The identification model operation 42 can use an existing Patch Generative Adversarial Network (Patch Generative Adversarial Network) to determine whether the first virtual MRI diffusion weighted image is a virtual image and whether the training MRI diffusion weighted image is a real image. The identification result is used to indicate whether the first virtual MRI diffusion weighted image is similar to the training MRI diffusion weighted image. The identification model architecture of the Patch-GAN includes at least one convolution layer, each layer corresponds to a plurality of nodes, and each node corresponds to a parameter; further, the processor 13 divides the first virtual MRI diffusion-weighted image and the training MRI diffusion-weighted image into m× n blocks, and the parameters are used to perform calculations based on the blocks, and the identification result is an m×n matrix whose content is 1 (True) or 0 (False). When a block in the first virtual MRI diffusion-weighted image is similar to a corresponding block in the training MRI diffusion-weighted image, the corresponding evaluation value in the matrix is 1.

In a judgment step 43, the processor 13 judges whether to terminate the model training phase according to the generation error and the identification error; if so, the parameters currently used by the generation model operation 41 are used as an inference phase parameter set and the phase is terminated; otherwise, a parameter update step 44 is performed to update the parameters used by the generation model operation 41 and the identification model operation 42, and return to the generation model operation 41 for the next round of training. In the first preferred embodiment, the judgment of the processor 13 is based on the termination of the model training phase when the generation error is the minimum and the identification error is the maximum, that is, when the difference between the generation error and the identification error is the minimum.

It is worth mentioning that the feature of the present invention is to train the generative model operation 41 and the identification model operation 42 at the same time, and the parameters used by the two are continuously updated during the training process. The two are constantly in conflict with each other, thereby making the first virtual MRI diffusion-weighted image generated by the generative model operation 41 close to its true standard (i.e., the MRI diffusion-weighted image used for training), while improving the discrimination power of the identification model operation 42.

Please refer to Figures 1 and 3. In the model inference stage, in a receiving step 51, the image processing system 1 receives a non-imaging brain tomography scan image of a patient from the input device 2; the processor 13 performs a generative model operation 52 based on the non-imaging brain tomography scan image of the patient and uses the inference stage parameter set to generate a virtual MRI diffusion-weighted image related to the patient; in an output step 53, the output device 3 provides the virtual MRI diffusion-weighted image related to the patient to the user (e.g., a doctor).

Please refer to Figures 1, 2, and 4, which are a second preferred embodiment of the image processing system 1 of the present invention for early detection of cerebral infarction core of ischemic stroke patients. The hardware and image processing method used are similar to those of the first preferred embodiment. In the following description, the same or similar components as those of the first preferred embodiment are represented by the same reference numerals.

The processor 13 of the image processing system 1 is used to perform a second preferred embodiment of the image processing method of the present invention for early detection of cerebral infarction core of ischemic stroke patients. The image processing method includes a model training phase and a model inference phase. The implementation details similar to the first preferred embodiment will not be elaborated below, and only the differences will be described.

In the second preferred embodiment, each training data set of the database 12 includes not only the non-imaging brain CT scan image and the MRI diffusion-weighted image for training of the same patient, but also the clinical data for training of the patient. Specifically, the clinical data used for training includes a demographic feature and a time feature; the demographic feature includes at least one of age, gender, NIH Stroke Scale (NIHSS), and laterality; the time feature includes at least one of a first time difference and a second time difference, the first time difference refers to the interval from the onset of the patient's stroke to the performance of a brain tomography scan without imaging, and the second time difference refers to the interval from the performance of a brain tomography scan without imaging to the performance of magnetic resonance imaging diffusion-weighted imaging. The clinical data used in the aforementioned training, such as the stroke scale and cerebral laterality, are all commonly known terms in the medical field to which this case belongs, and therefore will not be described in detail here.

In this second preferred embodiment, in the model training phase, the generative model operation 41 adopts a self-encoder architecture, which includes the operation of at least one coding layer 61, the operation of at least one bottleneck layer 62, and the operation of at least one decoding layer 63, wherein the operations of the coding layer 61 and the decoding layer 63 are similar to those of the first preferred embodiment, and the main difference lies in the operation of the bottleneck layer 62. In addition to receiving the output of the coding layer 61 and performing the bottleneck layer operation as in the first preferred embodiment, the bottleneck layer 62 is also used to receive the clinical data for training and perform a channel-wise attention (Channel-wise Attention) 621 operation and a spatial attention (Spatial Attention) 622 operation using the clinical data for training.

In the second preferred embodiment, the clinical data includes 6 values. During the calculation process, each value is first converted into a p×p matrix, which is filled with the value. Then, the matrix filled with the value is used to perform the calculation of the channel focus 621 and the spatial focus 622. For example, the age of a patient is converted into a p×p matrix, and each element in the matrix is filled with the age of the patient. If the gender is represented by 1 for male and 2 for female, then if the patient is male, each element in the p×p matrix is filled with 1, and so on. Then, the channel focus 621 and the spatial focus 622 are calculated using the 6 matrices filled with values. The operations of the channel focus 621 and the space focus 622 are operations that can be clearly understood by those with ordinary knowledge in the technical field to which this case belongs, so they are not described in detail here.

It is worth mentioning that, although in the second preferred embodiment, the generative model operation 41 adopts the autoencoder architecture, the generative model operation 41 may also adopt the variational autoencoder architecture, but is not limited thereto. When the variational autoencoder architecture is adopted, the clinical data used for training in some training data sets of the database 12 may be replaced with random numbers to perform the channel attention 621 and the spatial attention 622 operations.

Please refer to Figures 1 and 5. In a model inference stage of the second preferred embodiment, in a receiving step 71, the image processing system 1 receives a non-imaging brain sectional scan image and clinical data of a patient from the input device 2; the processor 13 uses an inference stage parameter set obtained in the model training stage to perform a generative model operation 72 based on the non-imaging brain sectional scan image and clinical data of the patient to generate a virtual MRI diffusion-weighted image related to the patient; in an output step 73, the output device 3 provides the virtual MRI diffusion-weighted image related to the patient to the user (e.g., a doctor). It should be noted that, since the present invention aims to generate the virtual MRI diffusion-weighted image using the generative model operation 72 after the patient has undergone a preliminary non-imaging brain tomography scan, that is, MRI diffusion-weighted imaging will not be performed on the patient, therefore, in the model inference stage, the second time difference of the time characteristic of the clinical data (i.e., the interval from the patient undergoing a non-imaging brain tomography scan to the patient undergoing MRI diffusion-weighted imaging) can be an arbitrarily given value.

Please refer to Figures 1 and 6. After a patient's non-imaged brain sectional scan image 81 passes through the image processing system 1 of the present invention, a virtual MRI diffusion-weighted image 82 as shown in Figure 6 can be obtained, which is very similar to the patient's actual MRI diffusion-weighted image 83; wherein, the bright areas in the virtual MRI diffusion-weighted image 82 and the MRI diffusion-weighted image 83 are the areas where the lesions are located, that is, the virtual MRI diffusion-weighted image 82 can be used for early detection of the cerebral infarction core of ischemic stroke patients.

To summarize the above, by using the image processing system 1 of the present invention, the patient does not need to undergo expensive and time-consuming MRI diffusion-weighted imaging, nor does he need to inject contrast agents and bear the risk of high radiation doses to undergo CT-based cerebral blood flow perfusion imaging. Instead, the patient only needs to undergo a first-line examination (obtaining the non-developmental brain CT-based scan image 81) to quickly obtain the virtual MRI diffusion-weighted image 82. The accuracy of the virtual MRI diffusion-weighted image 82 in early detection of the core of cerebral infarction in patients with ischemic stroke is similar to that of the MRI diffusion-weighted image 83, so the purpose of the present invention can be achieved.

However, the above is only an example of the implementation of the present invention, and it should not be used to limit the scope of the implementation of the present invention. All simple equivalent changes and modifications made according to the scope of the patent application of the present invention and the content of the patent specification are still within the scope of the patent of the present invention.

1: Image processing system for early detection of cerebral infarction in ischemic stroke patients 11: Memory 12: Database 13: Processor 2: Input device 3: Output device 41: Generate model operation 42: Identify model operation 43: Determine step 44: Parameter update step 51: Receiving step 52: Generate model operation 53: Output step 61: Coding layer 62: Bottleneck layer 621: Channel attention 622: Spatial attention 63: Decoding layer 71: Receiving step 72: Generate model operation 73: Output step 81: Non-reflected brain cross-sectional scan image 82: Virtual MRI diffusion-weighted image 83: MRI diffusion-weighted image

Other features and functions of the present invention will be clearly presented in the embodiments with reference to the drawings, in which:

FIG. 1 is a block diagram illustrating a first preferred embodiment of an image processing system for early detection of cerebral infarction core of ischemic stroke patients according to the present invention;

FIG2 is a flow chart illustrating a model training phase of an image processing method corresponding to the first preferred embodiment;

FIG3 is a flow chart illustrating a model inference phase of the image processing method corresponding to the first preferred embodiment;

FIG4 is a schematic diagram illustrating a self-encoder architecture used in a generative model operation in a second preferred embodiment of an image processing system for early detection of cerebral infarction core of an ischemic stroke patient according to the present invention;

FIG5 is a flow chart illustrating a model inference stage of an image processing method corresponding to the second preferred embodiment; and

FIG6 is a schematic diagram illustrating a virtual MRI diffusion-weighted image obtained after a non-imaging brain sectional scan image of a patient is processed by the present invention, and an actual MRI diffusion-weighted image of the patient.

41: Generative model calculation

42: Identification model calculation

43: Judgment steps

44: Parameter update step

Claims (13)

An image processing system for early detection of cerebral infarction core of ischemic stroke patients, comprising: A memory storing at least one instruction; A database including at least one training data set, wherein the training data set includes a training non-imaging brain tomography scan image and a training magnetic resonance imaging diffusion-weighted image corresponding to the training non-imaging brain tomography scan image; and A processor, communicatively coupled to the memory and the database, the processor is used to access and execute the instruction to operate in a model training phase, which includes: A generative model operation is used to perform a generative model learning operation using complex parameters based on the training non-imaging brain sectional scan image to generate a first virtual MRI diffusion-weighted image and a generative error; An identification model operation is used to perform an identification model learning operation using complex parameters based on the training MRI diffusion-weighted image and the first virtual MRI diffusion-weighted image to generate an identification error and an identification result; A determination step for determining whether to terminate the model training phase according to the generation error and the identification error. If the determination step is yes, the model training phase is terminated, and the parameters currently used in the generation model operation are used as an inference phase parameter set; and A parameter updating step for updating the parameters used in the generation model operation and the identification model operation. If the determination step is no, the parameter updating step is performed. An image processing system for early detection of cerebral infarction core in patients with ischemic stroke as described in claim 1, wherein the processor is also used to access and execute the instruction to operate in a model inference stage, in which the processor generates a virtual magnetic resonance imaging diffusion-weighted image based on a non-imaged brain sectional scan image and using the inference stage parameter set. An image processing system for early detection of cerebral infarction core in patients with ischemic stroke as described in claim 1, wherein the training data set of the database also includes clinical data for training. In the model training stage, the processor performs the generative model operation based on the non-revealed brain sectional scan images for training and the clinical data for training. An image processing system for early detection of cerebral infarction core in patients with ischemic stroke as described in claim 3, wherein the processor is also used to access and execute the instruction to operate in a model inference stage, in which the processor generates a virtual MRI diffusion-weighted image based on a non-revealed brain sectional scan image and clinical data and using the inference stage parameter set. An image processing method for early detection of cerebral infarction core of ischemic stroke patients includes a model training stage, which includes the following operations and steps: A generative model operation, which is used to perform a generative model learning operation using complex parameters based on a training non-imaging brain sectional scan image to generate a first virtual MRI diffusion-weighted image and a generation error; An identification model operation, which is used to perform an identification model learning operation using complex parameters based on a training MRI diffusion-weighted image and the first virtual MRI diffusion-weighted image to generate an identification error and an identification result; A determination step for determining whether to terminate the model training phase according to the generation error and the identification error. If the determination step is yes, the model training phase is terminated, and the parameters currently used in the generation model operation are used as an inference phase parameter set; and A parameter updating step for updating the parameters used in the generation model operation and the identification model operation. If the determination step is no, the parameter updating step is performed. The image processing method for early detection of cerebral infarction core in ischemic stroke patients as described in claim 5 also includes a model inference stage, in which a virtual MRI diffusion-weighted image is generated based on a non-imaged brain sectional scan image and using the inference stage parameter set. As described in claim 5, the image processing method for early detection of cerebral infarction core in patients with ischemic stroke, wherein the training data set also includes clinical data for training. In the model training stage, the generative model operation is performed based on the non-revealed brain sectional scan images used for training and the clinical data used for training. The image processing method for early detection of cerebral infarction core in ischemic stroke patients as described in claim 7 also includes a model inference stage, in which a virtual MRI diffusion-weighted image is generated based on a non-reflective brain sectional scan image and clinical data and using the inference stage parameter set. An image processing method for early detection of cerebral infarction core in ischemic stroke patients as described in claim 7, wherein the clinical data used for training includes at least one numerical value. In the model training stage, the numerical value is converted into a matrix, the matrix is filled with the numerical value, and then the generative model operation is performed based on the non-revealed brain sectional scan image used for training and the matrix. An image processing method for early detection of cerebral infarction core in ischemic stroke patients as described in claim 9, wherein the generative model operation in the model training stage includes at least one encoding layer operation, at least one bottleneck layer operation, and at least one decoding layer operation, and the bottleneck layer operation is a channel attention operation based on the matrix. An image processing method for early detection of cerebral infarction core in ischemic stroke patients as described in claim 9, wherein the generative model operation in the model training stage includes at least one encoding layer operation, at least one bottleneck layer operation, and at least one decoding layer operation, and the bottleneck layer operation is a spatial attention operation based on the matrix. An image processing method for early detection of cerebral infarction core in ischemic stroke patients as described in claim 7, wherein the clinical data used for training includes a demographic characteristic, and the demographic characteristic includes at least one of age, gender, a stroke scale, and cerebral laterality. An image processing method for early detection of cerebral infarction core in patients with ischemic stroke as described in claim 7, wherein the clinical data used for training includes a time feature, and the time feature includes at least one of a first time difference and a second time difference, the first time difference refers to the interval between the onset of the patient's stroke and the performance of a brain tomography scan without imaging, and the second time difference refers to the interval between the patient's brain tomography scan without imaging and the performance of diffusion-weighted magnetic resonance imaging.