KR102635484B1 - Brain Neural Network Structure Image Processing System, Brain Neural Network Structure Image Processing Method, and a computer-readable storage medium - Google Patents

Abstract

The present invention relates to an image processing method for a brain neural network structure, and to a system, method, and computer-readable medium for image processing an input brain neural network image image to distinguish the brain neural network structure using deep learning. The brain neural network structure image processing system according to the embodiment includes an input unit for inputting a brain nerve network image image in conjunction with a brain nerve network image capture device; A ground truth label map storage unit and a deep neural network are provided in which ground truth label maps, which are ground truth images of the brain neural network to be used as a control group for learning of the brain neural network, are stored by structure, and the brain neural network video image input to the input unit is divided by brain neural network structure. It will be comprised of an image processing unit that generates a predicted label map and processes the image while learning the predicted label map for each structure by applying a loss function to reduce the deviation from the ground truth label map for each structure previously stored in the ground truth label map storage. You can.

Description

Brain Neural Network Structure Image Processing System, Brain Neural Network Structure Image Processing Method, and a computer-readable medium on which a computer program for providing the same is recorded. storage medium}

The present invention relates to an image processing system, method, and computer-readable medium with a brain neural network structure through deep learning, and in particular, a loss function used for learning to reduce the deviation between predicted images and actual images in the process of performing deep learning. It relates to an image processing system, method, and computer-readable medium with a brain neural network structure that enables image processing closer to the actual image based on an adaptive loss function.

Neurons, which transmit stimulation and excitement as a unit of the nervous system, generally exhibit a highly polarized structure with several dendrites and a single axon from the cell body, and the form of these neurons is It is used as important data for understanding various neural properties, including synaptic integration and neuronal excitability. For example, several neurodegenerative diseases, such as Alzheimer's disease or Parkinson's disease, have been observed through abnormalities in neuron morphology.

Meanwhile, imaging technology through machine learning has developed greatly, and large-scale image data sets based on automation and accurate analysis tools have emerged, and these data sets are also being widely used in the biomedical imaging field.

At this time, most algorithms used rely on supervised learning that uses clean labels for learning, that is, accurate correct answer labels. Neurons are narrow, thin, and sparsely distributed, so many parts of the image have an empty background. Due to their intertwined and intersecting structures, it is difficult to create accurate pixel-level ground truth labels from neuron images using deep learning algorithms that have been used in the biomedical image field.

For this reason, recently, Mean Absolute Error (MAE), Symmetric cross entropy Learning (SL), Reverse Cross Entropy (RCE), and Normalized Cross Entropy (NCE) have been developed. ), Active and Passive Loss (APL), and Noised Robust Dice (NR-Dice) were used, but all of these loss functions are strong against noisy labels. It is not possible to represent complex shapes with results that are close to actual measurements.

The present invention is proposed to solve the above problems. In image processing of labels in a brain neural network that are complex and have a lot of space per pixel, labels are classified according to the structure of the brain neural network and learned for each label, but the average Mean Squared Error (MSE) and Adaptive Mean Squared Error (ADMSE), an adaptive loss function that uses spatially varying weights to prevent incorrect label learning, are used together to provide better results in real-world images. It allows image processing to be closer to the actual image of the brain neural network by applying the Structure Prior (STPR) loss function, which is a loss function that reduces the loss in merging after completing learning between each label. The purpose is to provide systems, methods, and computer-readable media.

An image processing system with a brain neural network structure according to an embodiment of the present invention to solve the above problem includes an input unit that is linked with a brain neural network image capture device and inputs a brain neural network image image; A ground truth label map storage unit and a deep neural network are provided in which ground truth label maps, which are ground truth images of the brain neural network to be used as a control group for learning of the brain neural network, are stored by structure, and the brain neural network video image input to the input unit is divided by brain neural network structure. It will be configured to include an image processing unit that generates a prediction label map and processes the image while learning the prediction label map for each structure by applying a loss function to reduce the deviation of the prediction label map for each structure from the ground truth label map for each structure previously stored in the ground truth label map storage. You can.

Here, the structure of the brain nerve network may include the background of the brain nerve network image image, the neuron cell body, dendrites, and axons.

In addition, the application of the loss function applies a mean squared error (MSE) loss function to the background of the brain neural network image and the predicted label map for the neuron cell body, and the mean squared error (MSE) loss function is applied to the prediction label map for the neuron cell body and the background of the brain neural network image. The prediction label map may be formed to apply an adaptive mean squared error (ADMSE) loss function that reflects the weight generated by determining the prediction value on a pixel-by-pixel basis.

In addition, the image processing unit further applies a structure prior (STPR) loss function that reduces the loss between prediction label maps divided by the structure of the brain neural network, and the STPR loss function uses prior information of the brain neural network structure. As a basis, it can be formed to reduce the loss between the predicted label maps for each structure of the brain neural network by using the MSE loss function.

Meanwhile, a brain neural network structure image processing method according to an embodiment of the present invention includes the steps of inputting a brain neural network image to an input unit; transmitting the input brain neural network image to an image processing unit; A brain neural network image transmitted to the image processing unit passes through a deep neural network to generate a predicted label map for each structure of the brain neural network, and the image processing unit stores the generated predicted label map for each brain neural network structure in a ground truth label map storage. It may be configured to include the step of comparing a pre-stored ground truth label map for each brain neural network structure, estimating a loss function for each structure, and learning through the estimated loss function.

Here, the application of the loss function applies the Mean Squared Error (MSE) loss function to the prediction label map for the brain neural network image background and neuron cell body, and the prediction for the dendrites and axons. The label map may be configured to apply an Adaptive Mean Squared Error (ADMSE) loss function that reflects the weight generated by determining the prediction value on a pixel-by-pixel basis.

In addition, the application of the loss function may be configured to further apply a Structure Prior (STPR) loss function that reduces the loss between prediction label maps divided by structure of the brain neural network.

Referring to FIGS. 5 to 7, the brain neural network structure image processing system, method, and computer-readable medium according to embodiments of the present invention, as a result, reduce state-of-the-art noise in various segmentation quality metrics by using ADMSE and STPR loss functions. It can be confirmed that it surpasses the robust loss function, and in particular, the ClDice score is the best, making it more effective in brain network segmentation problems.

1 is a configuration diagram of a brain neural network structure image processing system according to an embodiment of the present invention.
Figure 2 is a schematic diagram showing the process of a brain neural network structure image processing method according to an embodiment of the present invention.
Figure 3 is a flowchart of a brain neural network structure image processing method according to an embodiment of the present invention.
Figure 4 is a diagram showing in detail changes in image processing during the process of the brain neural network structure image processing method according to an embodiment of the present invention.
Here, (a) is the input image, (b) is the partial label map, (c) is the label map to which the ADMSE loss function is applied, (d) is the label map to which the STPR loss function is applied, and (e) is the final result label map. .
5A to 5B are diagrams showing comparison results of training images for a brain neural network structural image processing method according to an embodiment of the present invention using ADMSE and STPR loss functions and a brain neural network structural image processing method using other loss functions. .
Here, (a) is an input image, (b) is a partial label map, (c) is an example of applying the MSE loss function, (d) is an example of applying the NCE loss function, (e) is an example of applying the RCE loss function, (f) ) is an example of application of the NR-Dice loss function, and (g) is an example of application of the ADMSE and STPR loss functions according to an embodiment of the present invention.
6A to 6B are diagrams showing comparison results of test images for a brain neural network structural image processing method according to an embodiment of the present invention using ADMSE and STPR loss functions and a brain neural network structural image processing method using other loss functions. .
Here, (a) is the input image, (b) is the entire label map, (c) is an example of applying the MSE loss function, (d) is an example of applying the NCE loss function, (e) is an example of applying the RCE loss function, (f) is an example of application of the NR-Dice loss function, and (g) is an example of application of the ADMSE and STPR loss functions according to an embodiment of the present invention.
Figure 7 shows IoU, precision, recall, F1 score, and ClDice metric results for the brain neural network structure image processing method according to an embodiment of the present invention using ADMSE and STPR loss functions and the brain neural network structure image processing method using other loss functions. This is a comparison table.

Hereinafter, the description of the present invention with reference to the drawings is not limited to specific embodiments, and various changes may be made and various embodiments may be possible. In addition, the content described below should be understood to include all conversions, equivalents, and substitutes included in the spirit and technical scope of the present invention.

In the following description, the terms first, second, etc. are terms used to describe various components, and their meaning is not limited, and is used only for the purpose of distinguishing one component from other components.

Like reference numerals used throughout this specification refer to like elements.

As used herein, singular expressions include plural expressions, unless the context clearly dictates otherwise. In addition, terms such as “comprise,” “provide,” or “have” used below are intended to designate the presence of features, numbers, steps, operations, components, parts, or a combination thereof described in the specification. It should be construed and understood as not excluding in advance the possibility of the presence or addition of one or more other features, numbers, steps, operations, components, parts or combinations thereof.

Hereinafter, a brain neural network structure image processing system, method, and computer-readable medium according to an embodiment of the present invention will be described in detail with reference to the attached drawings.

1 is a configuration diagram of a brain neural network structure image processing system according to an embodiment of the present invention.

Referring to FIG. 1, a brain neural network structure image processing system according to an embodiment of the present invention may be configured to include an input unit 1, a ground truth label map storage unit 2, and an image processing unit 3.

Specifically, the input unit 1 is a place where a brain neural network image is input, and can be linked with the brain neural network imaging device 5. Here, the brain nerve network imaging device 5 is not limited to any specific type, and can be configured as any of various imaging devices capable of imaging the brain nerve network, such as CT or MRI.

The ground truth label map storage unit 2 may store a ground truth label map to be used for learning in the image processing unit 3, which will be described later. Here, the ground truth label map is a ground truth image of the brain neural network, and as the image processing unit 3 generates a prediction label map for each structure, it can be stored for each structure of the brain neural network to correspond to each prediction label map. The ground truth label maps stored for each structure of the brain neural network are used as a control for learning the predicted label map.

The image processing unit 3 divides the brain neural network image input into the input unit 1 according to the structure of the brain neural network and generates a prediction label map. For this purpose, the image processing unit 3 includes a deep neural network 20. It can be.

More specifically, the brain neural network image input to the input unit 1 may be transmitted to the deep neural network 20 provided in the image processing unit 3 to perform deep learning, and the deep neural network 20 may perform deep learning through deep learning. By dividing the brain neural network into structures, a prediction label map can be generated for each structure.

At this time, the structure of the brain neural network is a structure that includes the background of the video image of the brain neural network, the neuron cell body, dendrites, and axons. The brain neural network structure is divided into the four structures described above to generate a predicted label map for each. can do.

In other words, it is possible to generate a predicted label map for the brain neural network image background, a predicted label map for the neuron cell body, a predicted label map for dendrites, and a predicted label map for axons.

However, it is not necessarily limited to this, and the brain neural network structure may be divided into other structures in addition to the four parts above, and may be divided into a more simplified or more detailed structure than the above structure.

In addition, the image processing unit 3 is configured to process the image while learning by applying a loss function to reduce the deviation of the predicted label map generated for each brain neural network structure from the ground truth label map for each structure previously stored in the ground truth label map storage unit 2. It can be.

More specifically, the image processing unit 3 first generates a predicted label map for each brain neural network structure and compares it with the ground truth label map stored in the ground truth label map storage unit 2. Through this process, the loss function is estimated. can do. At this time, when the loss function is estimated, this loss function is back-propagated and applied, so that the predicted label map generated again becomes closer to the actual label map.

If the above process is repeated, the predicted label map generated by the image processing unit 3 is generated as close as possible to the desired ground truth label map, and through this learning, a more robust brain neural network classification can be achieved.

Here, the loss function used may be a mixture of the Mean Squared Error (MSE) loss function and the Adaptive Mean Squared Error (ADMSE) loss function, which may be used differently depending on the brain neural network structure. You can.

Specifically, the MSE loss function can be applied to the predicted label map for the background and neuron cell body of a brain neural network image that is relatively clearly distinguished and has few gaps between structures.

In addition, the ADMSE loss function is a loss function that reflects the weight generated by judging the prediction value on a pixel-by-pixel basis in the MSE loss function, and can be applied to the prediction label map for relatively undistinguished and complex dendrites and axons.

This ADMSE loss function is formed to suppress learning at an unlabeled pixel in the prediction label map for dendrites or axons by considering it as labeled if the prediction value is higher. It can be.

The weight applied here is It may be, and the ADMSE loss function will be explained in more detail in the brain neural network structure image processing method described later.

here, Is The predicted label map of the th class is of It means the second pixel value, is a user-defined parameter that controls the degree of weight applied to the loss function, Is It is a constant value added to the loss function to ensure stable convergence in the early stages of learning before convergence to correct .

Meanwhile, the image processing unit 3 may use the STPR loss function in addition to the MSE loss function and the ADMSE loss function described above.

The STPR loss function is a loss function used to reduce the loss between prediction label maps that occurs as the brain neural network structure imaging system of the present invention generates prediction label maps for each brain network structure.

Here, the STPR loss function can be formed to reduce the loss between the predicted label maps for each structure to the MSE loss function based on prior information of the brain neural network structure, and through this, when expressing the integrated brain network structure in the future, the overlap between brain neural network structures can be avoided to represent a more accurate brain neural network structure.

The STPR loss function will be explained in more detail in the brain neural network structure image processing method described later.

The brain neural network structure image processing system according to the embodiment of the present invention described above divides each brain neural network structure, performs prediction, learns, and applies the MSE loss function, ADMSE loss function, and STPR loss function to generate each brain neural network structure. By learning to reduce the deviation of the predicted label map from the actual label and, on the other hand, reducing the loss between the predicted label maps for each segmented brain neural network structure, accurate image processing is possible compared to a brain neural network structure image processing system that applies a different loss function. You can.

Figure 2 is a schematic diagram showing the process of a brain neural network structure image processing method according to an embodiment of the present invention, Figure 3 is a flowchart of a brain neural network structure image processing method according to an embodiment of the present invention, and Figure 4 is an embodiment of the present invention. This is a diagram showing in detail the changes in image processing during the process of the brain neural network structure image processing method. Here, (a) is the input image, (b) is the partial label map, (c) is the label map to which the ADMSE loss function is applied, (d) is the label map to which the STPR loss function is applied, and (e) is the final result label map. .

Referring to Figures 2 to 4, an image processing method using deep learning, a brain neural network structure image processing method according to an embodiment of the present invention, distribution Using paired random variables (X, Y) from A deep neural network parameterized with , sample image (here, is the distribution of the observed images) as input, the response map Y for all classes representing the brain network structure. (here, is about an image processing method to predict the distribution of brain neural network structure labels.

In this process, the present invention uses the Mean Squared Error (MSE) loss function and the Adaptive Mean Squared Error (ADMSE) loss function to predict the response map closest to the ground truth image while learning. Image processing can be performed to do so.

Here, the Adaptive Mean Squared Error (ADMSE) loss function is a loss function that applies a weight generated by determining the predicted value on a pixel basis to the Mean Squared Error (MSE) loss function. A more detailed explanation will be provided later. Do this.

Looking at the present invention in detail, the brain neural network structure image processing method according to an embodiment of the present invention includes the step of inputting a brain neural network image image to the input unit 1 (S10), and the input brain neural network image image is converted into an image. A step of transmitting to the processing unit 3 (S20), a step of generating a predicted label map for each structure of the brain neural network by passing the brain neural network image transmitted to the image processing unit 3 through the deep neural network 20 (S30), and The image processing unit 3 compares the generated predicted label map for each brain neural network structure with the ground truth label map for each brain neural network structure previously stored in the ground truth label map storage unit 2 to estimate a loss function for each structure, and estimate the loss function for each structure. It may be configured to include a step (S40) of learning through a loss function.

That is, the present invention can be configured to generate a prediction label map for each structure of the brain neural network from a brain neural network image input to a deep neural network during image processing through deep learning.

Here, the structure of the brain nerve network includes the background of the brain nerve network image image, nerve cell body, dendrites, and axon, so predictions are made for each of the four structures above. It is configured to create a label map.

At this time, the MSE loss function or the ADMSE loss function can be applied to each prediction label map.

Specifically, the MSE loss function is used for the background of the brain neural network image image and the predicted label map for the neuron cell body, which has a relatively clear structure, and the ADMSE loss function is used for the dendrites and axons that have a relatively unclear structure and are complex. It can be used in the prediction label map for

That is, each predicted label map is sequentially marked as 0, 1, 2, and 3, If we define as a set of label indices, we get Defined as , the set of predicted label maps for the background and neuronal cell body of the brain neural network image image with a relatively clear structure is , a set of predicted label maps for dendrites and axons with unclear structures is , It can be defined as .

However, this is only a preferred example and is not necessarily limited, and the definition of the set for the prediction label map may vary depending on the criteria set. For example, the neuron cell body may be combined with dendrites or axons, and the background of the brain nerve network image may be combined with dendrites or axons.

Meanwhile, if a partial prediction label map is given for each brain neural network structure as described above, all unlabeled pixels cannot be treated as background images. For this reason, the background image is labeled by applying intensity-based thresholding to the input image. can be created.

We will look at the ADMSE loss function below, which is a loss function applied to prediction label maps such as complex dendrites or axons.

First, the MSE loss function (or mean square error), which is the basis of the ADMSE loss function, is widely used to minimize the distance between the predicted label map and the ground truth label map, and the formula is as follows [Equation 1].

[Equation 1]

here, Is It is a deep neural network parameterized with , where X is the input image and Y is the ground truth label map.

The MSE loss as above is In the sense of , it effectively penalizes large errors, but is not resilient to noise that makes up the complexity of the training data, as in the problem that the present invention aims to solve.

For this reason, in the present invention, the prediction label map for structurally complex dendrites and axons can be learned using the ADMSE loss function in which adaptive weights are applied to the MSE loss function to control pixel-level backpropagation. .

Specifically, cast is the label map of the th class, cast of As for the second pixel value, the adaptive weight function for each pixel is Using (·,·) Can be defined as [Equation 2] below.

[Equation 2]

At this time, the ADMSE loss function is used in the unlabeled pixels of the prediction label map for structurally complex dendrites and axons as shown in [Equation 3] below. If the prediction value appears high, the ADMSE loss function is used during backpropagation because it is an unlabeled pixel. By assigning weights, learning at that pixel can be suppressed.

[Equation 3]

here is a user-defined parameter that controls the degree of weight applied to the loss function, Is It is a constant value added to the loss function to ensure stable convergence in the early stages of learning before convergence to correct .

In other words, if there is no label in the pixel as above, it is trained using the ADMSE loss function applying a low weight according to the predicted value, and if there is a label, it is trained using the MSE loss function, so that the image can be accurately processed even with complex labels.

Meanwhile, the present invention further includes a step (S45) of applying a structure prior (STPR) loss function to reduce the loss between each prediction label map after the brain neural network image is generated as a prediction label map for each structure. It can be.

Here, the step (S45) of applying the dictionary structure loss function may preferably be performed simultaneously with step S40, but it may also be applied before step S40 or after step S40.

That is, the brain neural network structure image processing method according to an embodiment of the present invention includes an ADMSE loss function that reduces the loss between the predicted label map and the ground truth label map, as well as a structure prior (STPR) loss function that reduces the loss between the predicted label maps for each structure. You can also use functions to more accurately process images of the brain neural network structure.

To this end, the present invention can first store prior information on the neural structure, and use a predicted label map for each structure of the brain neural network, and the first class predicted label map that is the basis of the brain neural network structure, that is, the background of the brain neural network image. ) can be extracted by predicting from the predicted label map.

That is, the first class prediction label map ( ) from the second ( ) and the third or fourth ( ) is predicted and extracted.

At this time, as shown in the figure, the loss between maps can be reduced through the STPR loss function, which reduces the loss by utilizing the MSE loss function between the predicted label maps, which reduces the loss between maps based on prior information of the pre-stored neural structure. can be reduced.

This applies to the condition that only one correct class must be assigned to a given pixel location, for example, a pixel cannot belong to both axons and dendrites at the same time. Accurate classification by structure of the brain neural network can be achieved through the STPR loss function. there is.

In other words, the STPR loss function can constrain the pixel from being classified as a neuron cell body or an axon if the probability that the pixel is the background or a dendrite is predicted to be high, and through this process, the brain neural network structure can be determined from partial labels. Normalization between predicted values can be learned so that they do not overlap while generating an overall label for each variable.

This STPR loss function can be defined by Equation 4 below, and the present invention utilizes the above-described ADMSE loss function and STPR loss function to perform high-accuracy prediction to perform image processing as close to the actual image as possible. there is.

[Equation 4]

here, is a sample image, am.

Equation 4 above is because one pixel cannot be a dendrite and an axon at the same time. … It is defined to have only one label, meaning that each pixel can have exactly one label.

In addition, the present invention can be implemented by storing computer-readable code in a computer-readable storage medium. The computer-readable storage medium includes all types of storage devices that store data that can be read by a computer system.

The computer-readable code is configured to perform steps for implementing the brain neural network structure image processing method according to the present invention when an image is processed by the image processor 3 from the input unit 1 from the computer-readable storage medium. do. The computer-readable code may be implemented in various programming languages. And functional programs, codes and code segments for implementing the present invention can be easily programmed by those skilled in the art to which the present invention pertains.

Examples of computer-readable storage media include ROM, RAM, CD-ROM, magnetic tape, floppy disk, and optical data storage, and also include those implemented in the form of carrier waves (e.g., transmitted via the Internet). Additionally, the computer-readable storage medium can be distributed across networked computer systems, so that computer-readable code can be stored and executed in a distributed manner.

Hereinafter, the following experimental examples are presented to explain in more detail the brain neural network structure image processing system and method of the present invention. However, the following experimental examples are only illustrative to aid understanding of the present invention, and the present invention is in the following experimental examples. It is not limited to the content.

[Experiment example]

[Experimental conditions]

The neuron image data used in this experiment consists of a total of 23 single cells and paired training data with partial annotations from the sum of three fully labeled test sets. Additionally, each neuron image has a size of 1024 did.

Here, because the number of pixel labels in dendrites was approximately 5 times greater than the number of pixel labels in axons, oversampling was applied to the labels in axons to prevent class imbalance problems.

The network proposed in this experiment is based on the existing U-Net using the ResNet-34 backbone encoder, the model is trained using Adam Optimizer with an initial learning rate of 3e-3 and a random batch size of 300, and cosine annealing ( The learning rate was gradually reduced to 3e-4 using the cosine annealing method.

Additionally, for the input image we set the threshold to 0.3 so that the background image was created, a user-defined parameter that controls the degree of weight applied to the loss function. applies 0.01 to structurally complex labels, is a constant value added to the loss function to ensure stable convergence in the early stages of learning before convergence to correct . was converged in initial learning by applying 0.03, and the experiment was performed through 10-fold cross-validation.

[Experimental Evaluation]

To prove the efficacy of partial labels, the brain neural network structure was compared with the MSE loss function, NCE loss function, APL loss function, RCE loss function, and NR-Dice loss function, with each image processing method set as a control group.

To quantitatively evaluate the performance of each method, we used three error metrics commonly used in image segmentation: Intersection over Union (IoU), F1 score, and ClDice.

[Results and Discussion]

The comparative results of image processing methods of the brain neural network structure according to the application of various loss functions are shown in Figures 5 and 6. Figure 5 shows the results of the training image, and Figure 6 shows the results of the test image.

As can be seen in Figures 5 and 6, the MSE loss function produced incomplete results, and other loss functions showed some improvement compared to the MSE loss function, but none produced accurate results like the image processing method of the present invention.

Additionally, other loss functions did not learn about axons during training due to partial labels, but the loss functions used in the present invention correctly reconstructed most axons. This is believed to be because other loss functions treat unlabeled pixels as negative labels.

Table 1 shown in Figure 7 is a table showing the segmentation accuracy measured using IoU, Precision, Recall, F1 score, and ClDice metrics compared to other control groups. The last three rows of this table show the research results of the loss function proposed in the present invention.

From these results, it can be seen that axons show extremely low recall values for MSE and noise robust loss functions compared to other classes, at 0.0954 and 0.1676.

However, it can be seen that the ADMSE and STPR loss functions are more resilient to label noise and achieve a recall rate of up to 0.5731.

In addition, the image processing method according to the embodiment of the present invention has a ClDice metric that evaluates the linearity of the structure, which has much better performance than other methods, and through this, the ADMSE and STPR loss functions applied according to the embodiment of the present invention are different loss functions. It shows that it is better suited to brain neural network segmentation.

As a result, it can be confirmed that the brain neural network structure image processing system and method according to an embodiment of the present invention using ADMSE and STPR loss functions surpass the state-of-the-art noise robust loss function in various segmentation quality metrics, and in particular, the ClDice score It was confirmed to be the best and more effective in brain neural network segmentation problems.

Although embodiments of the present invention have been described above with reference to the attached drawings, those skilled in the art will understand that the present invention can be implemented in other specific forms without changing the technical idea or essential features of the present invention. You will be able to understand it. Therefore, the embodiments described above are illustrative in all respects and are not restrictive.

1: input part
2: Ground truth label map storage unit
3: Image processing unit
5: Video recording device
10: input image
20: Deep neural network
30: Predicted label map
30a: Background prediction label map
30b: Neuron cell body prediction label map
30c: Dendrite prediction label map
30d: Axon prediction label map
40: Ground truth label map
40a: Background ground truth label map
40b: Neuron cell body ground truth label map
40c: Dendrite ground truth label map
40d: Axon actual label map
A1: MSE loss function
A2: ADMSE loss function
A3: SPTR loss function
C: Neuron cell body
D: dendrites
A: axon

Claims (10)

An input unit for inputting a brain nerve network image in conjunction with a brain nerve network imaging device;
A ground truth label map storage unit in which a ground truth label map, which is a ground truth image of a brain neural network to be used as a control for learning a brain neural network, is stored for each structure;
A deep neural network is provided to generate a predicted label map by dividing the brain neural network image input into the input unit by brain neural network structure, and combines the predicted label map for each structure with the ground truth label map for each structure pre-stored in the ground truth label map storage unit. It includes an image processing unit that processes images while learning by applying a loss function to reduce deviation,
The image processing unit,
A brain neural network structure image processing system that applies a structure prior (STPR) loss function that reduces the loss between prediction label maps divided by the structure of the brain neural network.
According to claim 1,
The structure of the brain neural network is,
A brain nerve network structure image processing system including the background, nerve cell body, dendrites, and axons of the brain nerve network image image.
According to claim 2,
Application of the above loss function is,
A mean squared error (MSE) loss function is applied to the predicted label map for the background and neuron cell body of the brain neural network image image,
To the predicted label map for the dendrites and axons, an adaptive mean squared error (ADMSE) loss function is applied to the mean squared error loss function, in which weights generated by determining the prediction value on a pixel basis are reflected. A brain neural network structure image processing system characterized by:
delete According to claim 1,
The STPR loss function is,
A brain neural network structure image processing system that reduces the loss between predicted label maps for each structure of the brain neural network based on prior information of the brain neural network structure by using the MSE loss function.
Inputting a brain neural network image to an input unit;
transmitting the input brain neural network image to an image processing unit;
A step in which the brain neural network image transmitted to the image processing unit passes through a deep neural network to generate a predicted label map for each structure of the brain neural network;
The image processing unit compares the generated predicted label map for each brain neural network structure with the ground truth label map for each brain neural network structure previously stored in the ground truth label map storage, estimates a loss function for each structure, and applies the estimated loss function to learn. Including the steps of:
Application of the above loss function is,
A brain neural network structure image processing method that applies a structure prior (STPR) loss function that reduces the loss between prediction label maps divided by the structure of the brain neural network.
According to claim 6,
The structure of the brain neural network is,
A method of image processing of a brain nerve network structure including the background, nerve cell body, dendrites, and axons of the brain nerve network image image.
According to claim 7,
Application of the above loss function is,
A mean squared error (MSE) loss function is applied to the predicted label map for the background and neuron cell body of the brain neural network image image,
To the predicted label map for the dendrites and axons, an adaptive mean squared error (ADMSE) loss function is applied to the mean squared error loss function, in which weights generated by determining the prediction value on a pixel basis are reflected. A brain neural network structure image processing method characterized by:
delete A computer-readable recording medium on which a computer program is recorded, providing a method for processing images of a brain neural network structure according to claim 6.