TWI877875B - Brain tomography computer-aided detection system, its method and its computer program product thereof - Google Patents

Abstract

A brain tomography computer-aided detection system is provided. The system includes an image selection module, an image preprocessing module and a visual scale prediction module. The image selection module selects at least a suitable original image from a set of dopamine tomography (TRODAT) original images of a person being tested. The image preprocessing module converts the at least a suitable original image into at least a processed image. The visual scale prediction module generates a visual scale prediction stage of the person being tested according to the at least a processed image.

Description

Brain tomography computer-assisted detection system, method and computer program product

This invention belongs to the field of computer-assisted recognition technology, especially the field of computer-assisted recognition technology for brain tomography.

Parkinson's Disease (PD) is a progressive disease that affects movement and is usually associated with a lack of dopamine neurons. Dopamine tomography (TRODAT) is often used to detect whether the subject's brain lacks dopamine neurons. However, currently reviewing TRODAT images and making judgments still rely on manpower, but doctors are busy and lack of manpower often leads to a decline in judgment quality or a lot of time cost

Therefore, there is a need for a computer-assisted recognition system, method and computer program product for brain tomography to solve the above problems.

The present invention proposes a brain tomography computer-assisted recognition technology based on a deep neural network, and uses the TRODAT images of the subject (e.g., a patient) to train the training model of the deep neural network. After the training is completed, the deep neural network can assist in predicting the visual scale (VS) prediction level of the subject.

According to one viewpoint of the present invention, a brain tomography computer-aided detection system is proposed. The system includes an image selection module, an image preprocessing module and a visual grading prediction module. The image selection module executes an image selection procedure to select at least one suitable original image from a set of TRODAT original images of a subject. The image preprocessing module executes an image preprocessing procedure to convert at least one suitable original image into at least one processed image. The visual grading prediction module generates a Visual Scale prediction grade of the subject based on at least one processed image.

According to another aspect of the present invention, a brain tomography computer-assisted detection method is provided, wherein the method is performed by a brain tomography computer-assisted detection system, wherein the brain tomography computer-assisted detection system includes an image selection module, an image preprocessing module, and a visual grading prediction module. The method comprises the steps of: performing an image selection procedure by an image selection module to select at least one suitable original image from a set of TRODAT original images of a subject; performing an image preprocessing procedure by an image preprocessing module to convert at least one suitable original image into at least one processed image; and generating a Visual Scale prediction grade of the subject according to at least one processed image by a visual grading prediction module.

According to another aspect of the present invention, a computer program product is provided, which is stored in a non-transitory computer-readable medium and is used to operate a brain tomography computer-assisted detection system, wherein the brain tomography computer-assisted detection system includes an image selection module, an image pre-processing module and a visual grading prediction module. The computer program product includes: an instruction to enable an image selection module to execute an image selection procedure to select at least one suitable original image from a set of TRODAT original images of a subject; an instruction to enable an image preprocessing module to execute an image preprocessing procedure to convert at least one suitable original image into at least one processed image; and an instruction to enable a visual grading prediction module to generate a Visual Scale prediction grade of the subject based on at least one processed image.

1: Brain tomography computer-assisted detection system

10: Data acquisition interface

20: Image selection module

30: Image preprocessing module

40: Visual grading prediction module

41: First-class neural network

50: Microprocessor

51: Computer program products

60: Second prediction module

61: Second type of neural network

S21~S25: Steps

S41~S45a: Steps

S51~S57: Steps

S61~S62: Steps

S71~S73: Steps

S81~S83: Steps

FIG1 is a schematic diagram of a brain tomography computer-assisted detection system according to an embodiment of the present invention; FIG2 is a flow chart of the steps of a brain tomography computer-assisted detection method according to an embodiment of the present invention; FIG3 is a schematic diagram of a brain tomography computer-assisted detection system according to another embodiment of the present invention; FIG4 is a flow chart of a brain tomography computer-assisted detection method according to another embodiment of the present invention. FIG. 5 is a step flow chart of the image selection procedure of an embodiment of the present invention; FIG. 6 is a step flow chart of the image preprocessing procedure of an embodiment of the present invention; FIG. 7 is a flow chart of the training process of the first type of neural network of an embodiment of the present invention; FIG. 8 is a flow chart of the training process of the second type of neural network of an embodiment of the present invention.

When read in conjunction with the attached drawings, the following embodiments are used to clearly demonstrate the above and other technical contents, features and/or effects of the present invention. Through the description of the specific implementation methods, people will further understand the technical means and effects adopted by the present invention to achieve the above-mentioned purpose. In addition, since the contents disclosed by the present invention should be easy to understand and can be implemented by technical personnel in this field, all equivalent substitutions or modifications that do not deviate from the concept of the present invention should be included in the claims.

It should be noted that in this article, unless otherwise specified, "a" element is not limited to a single element, but may also refer to one or more elements.

In addition, ordinals such as "first" or "second" in the specification and claims are only used to describe the claimed elements, and do not represent or indicate that the claimed elements have any order, and are not the order between the claimed elements and another claimed element or between the steps of the manufacturing method. These ordinals are used only to distinguish one claimed element with a specific name from another claimed element with the same name.

In addition, the descriptions of "when..." or "when..." in the present invention indicate "at the moment, before or after", etc., and are not limited to situations that occur at the same time, which is explained in advance. The descriptions of "set on..." and similar descriptions in the present invention indicate the corresponding position relationship between two elements, and do not limit whether the two elements are in contact, unless there is a special limitation, which is explained in advance. Furthermore, when the present invention records multiple effects, if the word "or" is used between the effects, it means that the effects can exist independently, but it does not exclude that multiple effects can exist at the same time.

In addition, the words "connected" or "coupled" in the specification and claims refer not only to direct connection with another element, but also to indirect connection or electrical connection with another element. In addition, electrical connection includes direct connection, indirect connection or communication between two elements using wireless signals.

In addition, in the specification and claims, the terms "about", "approximately", "substantially", and "roughly" usually indicate that the difference between a value and a given value is within 10%, 5%, 3%, 2%, 1%, or 0.5% of the given value. The quantity given here is an approximate quantity, that is, in the absence of specific description of "about", "approximately", "substantially", and "roughly", the meaning of "about", "approximately", "substantially", and "roughly" can still be implied. In addition, the terms "range is from a first value to a second value" and "range is between a first value and a second value" indicate that the range includes the first value, the second value, and other values between them.

In addition, in this article, the terms "system", "equipment", "device", "module", or "unit" refer to a digital circuit, an analog circuit, or other more general circuits that include an electronic component or are composed of multiple electronic components, and unless otherwise specified, they do not necessarily have a hierarchical or layered relationship. The above configurations are all based on actual applications.

In addition, as long as it is reasonable, the technical features of different embodiments disclosed in the present invention can be combined to form another embodiment.

FIG1 is a schematic diagram of a brain tomography computer-aided detection system 1 (hereinafter referred to as system 1) of an embodiment of the present invention. System 1 can be used to generate a Visual Scale predicted grade of a subject (e.g., a patient). As shown in FIG1 , system 1 can include a data acquisition interface 10, an image selection module 20, an image preprocessing module 30, and a visual grading prediction module 40.

In one embodiment, the data acquisition interface 10 can be used to obtain data from the outside, that is, the user (such as a doctor) can input image data into the system 1 through the data acquisition interface 10, wherein the image data can be, for example, a set of TRODAT original images of the subject.

The image selection module 20 can obtain the set of TRODAT original images of the subject from the data acquisition interface 10, and the image selection module 20 can execute an image selection procedure to select at least one suitable original image from the set of TRODAT original images of the subject.

The image preprocessing module 30 can obtain at least one suitable original image from the image selection module 20. The image preprocessing module 30 can execute an image preprocessing procedure to convert at least one suitable original image into at least one processed image.

The visual scale prediction module 40 may include a first type neural network 41 that has completed deep learning training. The visual scale prediction module 40 may obtain at least one processed image from the image pre-processing module 30, and perform feature analysis on at least one processed image, and then output the subject's Visual Scale prediction level. In one embodiment, the first type neural network 41 may include a training encoder and a linear classifier, wherein the training encoder may find various features in at least one processed image of the subject, and the linear classifier may analyze the subject's Visual Scale prediction level based on the features found by the training encoder, and is not limited thereto. In one embodiment, the visual grading prediction module 40 may output the subject's Visual Scale prediction grade. Here, "output" may be, for example, displaying the subject's Visual Scale prediction grade on a screen of a computer or other similar electronic device, or outputting in the form of a chart, graph, image, sound, text, document and/or file, and is not limited thereto.

Next, the details of the above components are described in more detail.

In one embodiment, the system 1 may be a data processing device, which may be implemented by any electronic device having a microprocessor, such as a desktop computer, a laptop computer, a smart mobile device, a server or a cloud host, or the system 1 may also be implemented by a chip in an electronic device. In one embodiment, the system 1 may have a network communication function to transmit data through the network, wherein the network communication may be a wired network or a wireless network, so the system 1 may also obtain data through the network. In one embodiment, the system 1 can realize its functions by a microprocessor 50 executing a computer program product 51, wherein the computer program product 51 may have a plurality of instructions, which can enable the microprocessor 50 to perform special operations, thereby enabling the microprocessor to realize functions such as the image selection module 20, the image preprocessing module 30, and the visual grading prediction module 40, but not limited thereto. In one embodiment, the computer program product 51 can be stored in a non-transient computer-readable medium (such as a memory), but not limited thereto. In one embodiment, the computer program product 51 can also be pre-stored in a network server for user downloading. In one embodiment, the computer program product 51 can actually include multiple subroutines.

In addition, in one embodiment, the data acquisition interface 10 can be a physical connection port for the system 1 to obtain external data. For example, when the system 1 is a computer, the data acquisition interface 10 can be a USB interface on the computer, various transmission line connectors, etc., but it is not limited. In addition, the data acquisition interface 10 can also be integrated with a wireless communication chip, so that data can be received by wireless transmission. In one embodiment, the data acquisition interface 10 can be electrically connected to a register or a memory of the system 1 to store the acquired data.

In one embodiment, a set of TRODAT original images of the subject is, for example, a set of "Tc-99m brain TRODAT images" used in nuclear medicine, but not limited thereto. In one embodiment, the TRODAT original images are, for example, a set of images generated by a scanning machine performing a tomographic scan of the subject's brain a period of time (e.g., about four hours) after the subject is injected with the TRODAT tracer, and the set of images scans the head, but not limited thereto. In one embodiment, at least 2,000 sets of images are used to train the first type of neural network 41, and at least 150 sets of images are used to test the first type of neural network 41, but not limited thereto.

When the system 1 is actually used (for example, when the first type of neural network 41 has completed training), the system 1 can execute a brain tomography computer-assisted detection method. FIG2 is a step flow chart of the brain tomography computer-assisted detection method of an embodiment of the present invention, and please refer to FIG1 at the same time.

As shown in FIG2 , first, step S21 is executed, and the data acquisition interface 10 acquires a set of TRODAT original images of the subject. Then, step S22 is executed, and the image selection module 20 executes an image selection procedure to select at least one suitable original image from the set of TRODAT original images as an image input to the first type neural network 41 for analysis. Then, step S23 is executed, and the image pre-processing module 30 executes an image pre-processing procedure to convert the at least one suitable original image into at least one processed image. Then, step S24 is executed, and the first type neural network 41 of the visual grade prediction module 40 finds a plurality of features of at least one processed image. Then step S25 is executed, and the first type of neural network 41 analyzes the subject's Visual Scale predicted level based on the features. Afterwards, the visual grading prediction module 40 can output the subject's Visual Scale predicted level.

The above steps can be adjusted in order or increased or decreased according to needs, and are not limited thereto. In this way, the system 1 can analyze the subject's Visual Scale prediction level based on the subject's TRODAT image, which can assist doctors in making judgments, reduce doctors' fatigue, and improve accuracy.

In addition, the system 1 of the present invention may also have a variation. FIG3 is a schematic diagram of a brain tomography computer-assisted detection system 1 of another embodiment of the present invention. The embodiment of FIG3 is generally applicable to the description of the embodiment of FIG1, so the following mainly describes the differences.

As shown in FIG3 , the system 1 further includes a second prediction module 60. The second prediction module 60 may include a second type of neural network 61 that has completed deep learning training. The second prediction module 60 may obtain at least one processed image from the image pre-processing module 30, and perform feature analysis on at least one processed image, and then output the Hoehn-Yahr Scale prediction level of the subject. In one embodiment, the second prediction module 60 may include a training encoder and a linear classifier, wherein the training encoder may find various features in at least one processed image of the subject, and the linear classifier may analyze the Hoehn-Yahr Scale prediction level of the subject based on the features found by the training encoder, and is not limited thereto. In one embodiment, the second prediction module 60 can output the Hoehn-Yahr Scale prediction level of the subject. Here, "output" can be, for example, displaying the Hoehn-Yahr Scale prediction level of the subject on a screen of a computer or other similar electronic device, or outputting it in the form of a chart, a graph, an image, a sound, a text, a document and/or a file, but is not limited thereto. In one embodiment, at least 250 sets of images are used to train the second type of neural network 61, and at least 30 sets of images are used to test the second type of neural network 61, but is not limited thereto.

Figure 4 is a flowchart of the steps of the brain tomography computer-assisted detection method of another embodiment of the present invention, and please refer to Figure 3 at the same time.

As shown in FIG4 , first, step S41 is executed, and the data acquisition interface 10 acquires a set of TRODAT original images of the subject. Then, step S42 is executed, and the image selection module 20 executes an image selection procedure to select at least one suitable original image from the set of TRODAT original images as an image input to the first neural network 41 and the second neural network 61 for analysis. Then, step S43 is executed, and the image pre-processing module 30 executes an image pre-processing procedure to convert the at least one suitable original image into at least one processed image. Then, step S44 is executed, and the first neural network 41 of the visual grade prediction module 40 finds a plurality of features of at least one processed image. In addition, step S44a can be executed, and the second type of neural network 61 of the second prediction module 60 finds multiple features of at least one processed image. Then step S45 is executed, and the first type of neural network 41 analyzes the subject's Visual Scale prediction level based on the features. In addition, step S45a can be executed, and the second type of neural network 61 analyzes the subject's Hoehn-Yahr Scale prediction level based on the features. Thereafter, the visual grading prediction module 40 can output the subject's Visual Scale prediction level, and the second prediction module 60 can output the subject's Hoehn-Yahr Scale prediction level.

The Hoehn-Yahr grade is an important evaluation indicator for assessing the severity of a patient's Parkinson's disease. Therefore, System 1 can simultaneously analyze the subject's Visual Scale predicted grade and Hoehn-Yahr Scale predicted grade through the subject's TRODAT images, which will assist doctors in assessing the patient's condition more accurately and improve efficiency.

It can be seen that the system 1 of the present invention can have a visual grading prediction module 40 and can optionally have a second prediction module 60. For the convenience of explanation, the subsequent paragraphs are all explained as if the system 1 has both the visual grading prediction module 40 and the second prediction module 60.

Furthermore, one of the features of the present invention includes that the TRODAT original image can be subjected to an image selection process and an image preprocessing process before being input into the visual grading prediction module 40 and the second prediction module 60, thereby improving the prediction effect of the visual grading prediction module 40 and the second prediction module 60.

First, the image selection process is explained. After the subject's brain is subjected to CT scan, the scanning machine can generate multiple TRODAT original images, each of which can, for example, show a cross-sectional image of a certain position of the brain. However, since the direction of the CT scan usually starts from the top of the subject's head, and the basal ganglia that can actually reflect the amount of brain is located in the middle of the brain, some of the original images may have more impurities, such as the image is occupied by the skull, etc., and these impurities will affect the accuracy of the analysis of the first type of neural network 41 and the second type of neural network 61. Therefore, the image selection module 20 needs to perform an image selection process to automatically select images suitable for the analysis of the first type of neural network 41 and the second type of neural network 61.

To achieve the above purpose, the image selection process may include multiple steps. Figure 5 is a step flow chart of the image selection process of an embodiment of the present invention, please refer to Figures 1 to 4 at the same time.

First, step S51 is executed, and the image selection module 20 determines whether the number of images in a set of TRODAT original images of the subject is greater than a first threshold value. When the number of images in the set of TRODAT original images is greater than the first threshold value, step S51a is executed, and the image selection module 20 starts to find an image that meets a preset condition from the first image in the set of TRODAT original images, and takes out the image that meets the preset condition and a specific number of images following the image as a set of preliminary selected images. When the number of images in the set of TRODAT original images is less than or equal to the first threshold value, the image selection module 20 directly sets the set of TRODAT original images as the set of preliminary selected images. After the group of preliminary selected images is selected, step S52 is executed to perform a reduction optimization process on the group of preliminary selected images to obtain at least a portion of the images from the group of preliminary selected images. Then step S53 is executed to perform center point alignment of at least a portion of the group of preliminary selected images obtained after the reduction optimization process with a mask, wherein the mask includes a base nucleus range. Then step S54 is executed to obtain a maximum pixel value of each image corresponding to the base nucleus range for each aligned image. Then step S55 is executed to compare the maximum pixel value of each aligned image corresponding to the base nucleus range. Then step S56 is executed, and the image with the maximum pixel value higher than other images is set as at least one suitable original image. Then step S57 can be executed, and the previous image and the next image of the image with the maximum pixel value are also set as the at least one suitable original image. In this way, the image selection process can be completed, and the image selection module 20 can automatically select at least one suitable original image suitable for the first type of neural network 41 and the second type of neural network 61 to analyze. It should be noted that the above steps are only examples, and as long as they are reasonably achievable, the above steps can be changed in order or increased or decreased according to needs.

Regarding step S51. In one embodiment, the first threshold value may be, for example, between 60 and 70 images, such as 64 images, but not limited thereto. Specifically, since each image may correspond to a scanning position on a scanning path starting from the top of the subject's head, when the number of images exceeds the first threshold value, it may indicate that some parts that do not need to be observed (such as bones) may also be scanned multiple times, so the set of TRODAT original images may contain many images with more impurities that are not conducive to analysis. Therefore, in this case, step S51a needs to be performed to preliminarily select the set of TRODAT original images. On the contrary, when the number of images exceeds the first threshold value, it means that the number of images corresponding to impurities in the set of TRODAT original images is relatively small, so the set of TRODAT original images can be directly used for subsequent processing.

Step S51a is the details of preliminary selection. This step will take the first image (e.g., corresponding to the top of the head) of the set of TRODAT original images as the starting point, and select images that meet the preset conditions from the set of TRODAT original images as preliminary selected images. In one embodiment, the preset conditions may include at least: (1) the image has a plurality of pixels with pixel values greater than a second threshold value; and (2) the number of pixels with pixel values greater than the second threshold value is greater than a third threshold value. In one embodiment, the second threshold value may be, for example but not limited to, between 3 and 10, between 4 and 8, or between 4 and 6, and is not limited thereto. In one embodiment, the third threshold value may be between 300 and 700, between 400 and 600, or between 450 and 550, and is not limited thereto. For example, if the second threshold value is 5 and the third threshold value is 500, then in step S51a, the image selection module 20 will start to search for an image with more than 500 pixels with pixel values greater than 5 from the first image in the group of TRODAT original images, thereby, the image first found by the image selection module 20 and multiple images following the image (the number of subsequent images can be, for example, up to 64, and together with the image, 65, but not limited to this) can be set as the group of preliminary selected images.

Specifically, after the basal ganglia absorbs the TRODAT tracking agent, the basal ganglia secrete dopamine, making its pixel value higher than other parts. Therefore, the image with more than 500 pixels with a pixel value greater than 5 is more likely to correspond to the basal ganglia. Therefore, when an image that meets the preset conditions is found, it means that the scanned position is close to the basal ganglia. Therefore, this image and subsequent images can be selected as preliminary selected images. Thus, steps S51 and S51a can be understood.

Regarding step S52, this step can further reduce and optimize the preliminary selected images so that the reduced and optimized images can better match the location of the basal ganglia. In one embodiment, the image selection module 20 can take the median value of the total number of preliminary selected images, and obtain multiple images before and after the median value, such as 1/4, 1/5, 1/6, 1/7, or 1/8 of the total number of images, and is not limited to this. For example, if the total number is 60 images, the median number can be the 30th (or 31st) image, and 1/6 of the total number is 10, so the image selection module 20 will select the 20th to 40th images from the preliminary selected images as the reduced and optimized images. Thus, step S52 can be understood.

Regarding step S53, this step is to use a mask to obtain the range of the basal nucleus in the reduced and optimized image. For example, the mask can mark the range of the basal nucleus in the image, or can cover the range outside the basal nucleus in the image, but is not limited to this. In this way, the subsequent steps can process the relevant parts of the basal nucleus in each image.

Regarding steps S54 to S56, the image selection module 20 can obtain the maximum pixel value of each image from the pixel points within the basal ganglia range corresponding to each image, and then compare the maximum pixel value of each image to obtain an image with a maximum pixel value higher than other images. Specifically, the basal ganglia can have a higher pixel value when secreting dopamine, so the higher the pixel value usually means the closer to the region of interest (ROI) to be observed, so the image with a maximum pixel value higher than other images can be regarded as the most suitable image for analysis among the images. Therefore, the image selection module 20 can select at least one suitable original image suitable for analysis by the first type neural network 41 and the second type neural network 61. Thus, steps S54 to S56 can be understood.

Regarding step S57, specifically, the basal ganglia range of the previous image and the next image whose maximum pixel value is higher than other images usually also has a higher pixel value, and is also suitable for analysis by the first type neural network 41 and the second type neural network 61, and the amount of input data of the first type neural network 41 and the second type neural network 61 can be increased. Therefore, step S57 uses the previous image and the next image with the highest maximum pixel value as suitable original images, which can improve the analysis quality. Thus, step S57 can be understood.

By executing the image selection process, the image selection module 20 can automatically select images suitable for analysis, thereby improving the quality and effect of the analysis. Thus, the image selection process can be understood.

Next, the image preprocessing procedure is described. FIG6 is a flow chart of the steps of the image preprocessing procedure of an embodiment of the present invention, and please refer to FIG1 to FIG5 at the same time.

First, step S61 is executed to scale all pixel values of at least one suitable original image according to a maximum pixel value corresponding to at least one suitable original image and a preset maximum pixel value. Then step S62 is executed to normalize the scaled at least one suitable original image to generate the at least one processed image, wherein the normalization process includes binarization process.

Regarding step S61, specifically, since the scanning time of TRODAT tomography may be inconsistent in different testing centers, the pixel value range of the scanned image may also be inconsistent. For example, a shorter scanning time will result in a lower maximum pixel value in the image, while a longer scanning time will often result in a higher maximum pixel value in the image. Therefore, if the data from different testing centers are directly input into the first type of neural network 41 and the second type of neural network 61, the analysis will lose accuracy due to the inconsistent standards of the maximum pixel value. Therefore, step S61 is required to solve this problem. In one embodiment, the system 1 may be preset with a preset maximum pixel value, and the image preprocessing module 30 may scale the pixel values of all pixels in at least one suitable original image according to the ratio between the maximum pixel value in at least one suitable original image and the preset maximum pixel value, so that the pixel value ranges of all images input to the first type neural network 41 and the second type neural network 61 can be consistent. Thus, step S61 can be understood.

Regarding step S62, in one embodiment, the image pre-processing module 30 can perform binarization processing on at least one suitable original image after scaling adjustment, for example, pixel values greater than (or greater than or equal to) a specific value are all converted to 1, and pixel values less than or equal to (or less than) the specific value are all converted to 0, so that the image can better highlight the absorption effect of the basal ganglia. It should be noted that the above steps are only examples, and as long as they are reasonably achievable, the above steps can be changed in order or increased or decreased according to needs.

Thus, the image preprocessing process can be understood.

Next, the details of the visual rating prediction module 40 are described.

In one embodiment, the first neural network 41 of the visual rating prediction module 40 is an artificial intelligence model that uses a deep convolutional neural network (CNN) to analyze the features of an image. In one embodiment, the first neural network 41 is formed by training a first training model (e.g., a training deep convolutional neural network) through deep learning. In one embodiment, the first training model is trained using a large amount of training images in a supervised comparative learning manner. In one embodiment, when the first training model is trained, a feature path is generated. The feature path can be regarded as a neuron conduction path in the artificial intelligence model, where each neuron can represent an image feature detection point, and each image feature detection point may have a different weight value. After the first training model is trained, a first type of neural network 41 can be formed.

In one embodiment, supervised contrastive learning may use, for example, autonomous supervised contrastive learning technology, but is not limited thereto. The benefit of using supervised contrastive learning for the present invention is that supervised contrastive learning technology can enable the analysis of the artificial intelligence model to more effectively utilize the information of the label, pull clusters of the same category together in the vector (Embedding) space, and push away sample clusters from different categories, so that the specificity between features corresponding to different Visual Scale levels can be more clearly distinguished. In this way, the present invention can have good accuracy and robustness, and is robust to image corruption and hyperparameter changes. In one embodiment, the hardware device of the system 1 or the visual rating prediction module 40 may adopt, for example, an Nvidia V100 graphics processor, but is not limited thereto.

In one embodiment, the architecture of the first type of neural network 41 may include, for example, 1 to 3 input layers, 4 to 8 convolutional layers, 1 to 3 flat layers, 1 to 3 fully-connected layers, and 1 to 3 output layers, but is not limited thereto. In one embodiment, the convolutional layer may be used to find and integrate multiple image features from the training data. The flat layer may perform dimension conversion on the features found by the convolutional layer. The fully-connected layer may establish the correlation between the image features and the "labels of the training data (e.g., different visual scale levels)" (e.g., establish a feature path). In addition, the fully connected layer may include a loss function layer, wherein the loss function layer may be implemented, for example, using cross entropy loss, such as using Categorical Cross-entropy, but not limited thereto. In addition, the above architecture may also be regarded as the architecture of the first training model of the first type of neural network 41, but the characteristic paths and neurons of the first training model have not yet been trained to maturity.

In one embodiment, the first type of neural network 41 may include more than one million neurons, but is not limited thereto. In one embodiment, the hyperparameters of the first type of neural network 41 are set as follows: the batch size is set to be between 200 and 600, and the training epoch is set to be between 300 and 500, but is not limited thereto.

Next, the training process of the first type of neural network 41 (first training model) is described. FIG7 is a flow chart of the training process of the first type of neural network 41 of an embodiment of the present invention. Please refer to FIG1 to FIG6 at the same time.

As shown in FIG7 , first, step S71 is executed, and a plurality of training images are input into the first training model, wherein the training images have labels of corresponding visual scale levels. In one embodiment, the training images may be TRODAT original images of different subjects, wherein the images are directly in the original image format of the digital imaging and communications in medicine (DICOM) protocol without format conversion, wherein the pixel values of the DICOM images may be read by the pydicom function of the software tool python 3.7.0, but not limited thereto. In addition, the training images input into the first training model may first be subjected to the image selection procedure and image preprocessing procedure described in the previous paragraph, so that the training images may be regarded as processed images. In addition, the training images for each subject may be, for example, a set of training images, such as a set of three training images, such as the three images selected by the image selection process in FIG. 5 .

Then step S72 is executed, and the first training model performs supervised comparative learning using the training images to find image features in the training images.

Then step S73 is executed, and the first training model establishes a feature path according to the found image features to complete the training of the first training model. In one embodiment, the first training model can be trained, for example, using a cross entropy loss method to establish an image feature path, but is not limited thereto.

In one embodiment, the first training model needs to undergo at least one "training phase" to train and establish a feature path, and needs to undergo at least one "testing phase" to test the accuracy of the feature path. Only when the accuracy meets the requirements can it be used as the first type of neural network 41 for subsequent actual use. In one embodiment, the first training model will undergo multiple trainings, and different feature paths will be generated after each training, and the feature path with the highest accuracy will be set as the actual feature path of the first type of neural network 41, and is not limited to this. In addition, the actual feature path of the first type of neural network 41 can also be adjusted at any time.

Thus, the establishment process of the first type of neural network 41 can be understood. Accordingly, the visual grading prediction module 40 of the present invention can predict the visual scale level of the subject, which can be used to assist doctors in making judgments and greatly improve medical efficiency.

Next, the detailed structure of the second prediction module 60 is described.

In one embodiment, the second neural network 61 of the second prediction module 60 is an artificial intelligence model that uses a deep convolutional neural network (CNN) to analyze the characteristics of an image. In one embodiment, the second neural network 61 is formed by training a second training model (e.g., a deep convolutional neural network for training) through deep learning. In one embodiment, the second neural network 61 of the second prediction module 60 can adopt the same architecture as the first neural network 41, so the details are not described in detail. In other words, the second training model can have the same architecture as the first training model, but different prediction capabilities are generated by different training data.

Next, the training process of the second type of neural network 61 (the second training model) is described. FIG8 is a flow chart of the training process of the second type of neural network 61 of an embodiment of the present invention. Please refer to FIG1 to FIG7 at the same time.

As shown in FIG8 , first, step S81 is executed, and a plurality of training images are input into the second training model, wherein the training images have corresponding Hoehn-Yahr Scale level labels. In one embodiment, the training images may be TRODAT original images of different subjects, and an image selection process and an image preprocessing process may be performed before the training images are input into the second training model, so the training images may be regarded as processed images.

Then step S82 is executed, and the second training model performs supervised comparative learning using the training images to find the image features in the training images.

Then step S83 is executed, and the second training model establishes a feature path according to the found image features to complete the training of the second training model. In one embodiment, the second training model can be trained, for example, using a cross entropy loss method to establish an image feature path, but is not limited thereto.

Similar to the training of the first training model, the second training model also needs to go through at least one "training phase" and at least one "testing phase". In addition, the actual feature path of the second type of neural network 61 after training can also be adjusted at any time.

Thus, the establishment process of the second type of neural network 61 can be understood. Accordingly, the second prediction module 60 of the present invention can predict the Hoehn-Yahr Scale level of the subject, which can enhance the assistance to doctors.

Furthermore, in one embodiment, after the visual grading prediction module 40 or the second prediction module 60 generates a prediction result, the system 1 may output a preset result. In one embodiment, the system 1 may be connected to a user interface (UI) of a specific application (APP), and the prediction result of the visual grading prediction module 40 or the second prediction module 60 may be displayed on the user interface, but not limited thereto. In another embodiment, the system 1 may automatically generate a relevant report based on the prediction result of the visual grading prediction module 40 or the second prediction module 60, but not limited thereto.

Thus, through the present invention, as long as a set of TRODAT original images of the subject is input into the system 1, the system 1 can automatically select the images suitable for analysis, and predict the visual scale prediction level or Hoehn-Yahr Scale prediction level of the subject based on the images, thereby assisting doctors in making medical judgments. Through deep learning training, the system 1 of the present invention can accurately provide prediction levels, which can assist the subject in seeking the best medical care method.

Although the present invention has been described through the above embodiments, it is understandable that many modifications and variations are possible according to the spirit of the present invention and the scope of the patent application claimed by the present invention.

S41~S45a: Steps

Claims (10)

A brain tomography computer-aided detection system comprises: an image selection module (20) executing an image selection procedure to select at least one suitable original image from a set of dopamine tomography (TRODAT) original images of a subject; an image preprocessing module (30) executing an image preprocessing procedure to convert the at least one suitable original image into at least one processed image; and a visual scale prediction module (40) generating a visual scale (Visual Scale) prediction level of the subject according to the at least one processed image. The brain tomography computer-aided detection system as claimed in claim 1, wherein the image selection procedure comprises the steps of: the image selection module (20) determining whether the number of images in the set of TRODAT original images is greater than a first threshold value; when the number of images in the set of TRODAT original images is greater than the first threshold value, finding an image that meets a preset condition from the first image in the set of TRODAT original images; and selecting the image that meets the preset condition. The image and a specific number of images following the image are taken out as a set of preliminary selected images, wherein the preset conditions include: the image has a plurality of pixels with pixel values greater than a second threshold value, and the number of pixels with pixel values greater than the second threshold value is greater than a third threshold value; and when the number of images in the set of TRODAT original images is less than or equal to the first threshold value, the set of TRODAT original images is used as the set of preliminary selected images. A brain tomography computer-aided detection system as described in claim 2, wherein the image selection procedure comprises the steps of: aligning at least a portion of the image in the group of preliminary selected images with a mask, wherein the mask includes a basal ganglia range; obtaining a maximum pixel value corresponding to the basal ganglia range for each aligned image; comparing the maximum pixel value corresponding to the basal ganglia range for each aligned image; and setting the image with the maximum pixel value as the at least one suitable original image. A brain tomography computer-aided detection system as described in claim 3, wherein the at least one suitable original image also includes a previous image and a subsequent image of the image having the maximum pixel value. As described in claim 3, the brain tomography computer-aided detection system, wherein the image selection process includes the steps of: reducing and optimizing the group of preliminary selected images, and aligning the center point of the group of preliminary selected images after the reduction and optimization process with the mask. The brain tomography computer-aided detection system as described in claim 1, wherein the image preprocessing procedure includes the steps of: performing a normalization process on the at least one suitable original image to form the at least one processed image, wherein the normalization process includes a binarization process. A brain tomography computer-aided detection system as described in claim 6, wherein the normalization process comprises the steps of: scaling all pixel values of the at least one suitable original image according to a maximum pixel value corresponding to the at least one suitable original image and a preset maximum pixel value. The brain tomography computer-aided detection system as described in claim 1 further includes a second prediction module that generates a Hoehn-Yahr Scale prediction level of the subject based on the at least one processed image. A brain tomography computer-aided detection method is implemented by a brain tomography computer-aided detection system, wherein the brain tomography computer-aided detection system comprises an image selection module (20), an image preprocessing module (30) and a visual grading prediction module (40), wherein the method comprises the steps of: executing an image selection module (20) A selection process is used to select at least one suitable original image from a set of TRODAT original images of a subject; an image preprocessing process is executed by the image preprocessing module (30) to convert the at least one suitable original image into at least one processed image; and a visual scale prediction module (40) is used to generate a visual scale prediction level of the subject according to the at least one processed image. A computer program product is stored in a non-transitory computer-readable medium and is used to operate a brain tomography computer-assisted detection system, wherein the brain tomography computer-assisted detection system includes an image selection module (20), an image pre-processing module (30) and a visual grading prediction module (40), wherein the computer program product includes: an instruction to enable the image selection module (20) to execute An image selection procedure is performed to select at least one suitable original image from a set of TRODAT original images of a subject; an instruction is used to make the image pre-processing module (30) perform an image pre-processing procedure to convert the at least one suitable original image into at least one processed image; and an instruction is used to make the visual scale prediction module (40) generate a Visual Scale prediction level of the subject according to the at least one processed image.

Images