Method and system for detecting cerebral aneurysms using artificial intelligence using 3D angiographic Computed Tomography

Abstract

An artificial intelligence cerebral aneurysm detection system using 3D angiographic CT according to one embodiment of the present invention comprises: an input unit for acquiring a CT image of a patient; and an information processing unit including a patch model operating processor for operating a patch model that predicts the location of a cerebral infarction on a patch basis, and a patient model operating processor for operating a patient model that predicts the presence or absence of a cerebral aneurysm on a patient basis based on the results of the patch model. By doing so, the presence or absence of a cerebral aneurysm in a patient can be determined, thereby contributing to the improvement of the accuracy of cerebral aneurysm diagnosis.

Description

Technology Field

[0001] The present invention relates to a method and system for detecting cerebral aneurysms, and more specifically, to an artificial intelligence-based method and system for detecting cerebral aneurysms using three-dimensional CT angiography. Background Technology

[0003] A cerebral aneurysm refers to the abnormal expansion of a portion of the vital blood vessels supplying blood to the brain due to weakening. Because these aneurysms weaken like inflated balloons, they can rupture; since they are located within the brain, a rupture causes a cerebral hemorrhage.

[0004] When a cerebral aneurysm ruptures, one-third of patients die before arriving at the hospital, and more than half die within 30 days of the hemorrhage, or in some patients who survive, even with treatment, severe aftereffects remain.

[0005] Therefore, since it is important to detect cerebral aneurysms before they rupture using MRI, CT, etc., it is necessary to explore ways to improve the diagnostic accuracy of cerebral aneurysms.

[0006] delete Prior art literature

[65535] Korean Registered Patent No. 10-1974786 (Title of Invention: Method and System for Predicting Severity and Prognosis Using Characteristics of Cerebral Aneurysm Lesions) The problem to be solved

[0007] The present invention has been devised to solve the above-mentioned problems, and the objective of the present invention is to provide a method and system capable of determining the presence or absence of a cerebral aneurysm in a patient by using a deep learning artificial intelligence algorithm based on a convolutional neural network (CNN) and a recurrent neural network (RNN) on a brain computed tomography (CTA, CT Angiography) image in which a contrast agent is used. means of solving the problem

[0009] According to one embodiment of the present invention for achieving the above objective, an artificial intelligence cerebral aneurysm detection system using 3D angiography CT comprises: an input unit for acquiring a CT image of a patient; and an information processing unit including a patch model operating processor for operating a patch model that predicts the location of a cerebral infarction on a patch basis, and a patient model operating processor for operating a patient model that predicts the presence or absence of a cerebral aneurysm on a patient basis based on the results of the patch model.

[0010] And the patch model may include a patch extraction module that divides the patient's CT image into patches of a preset size, and includes an overlap portion at a certain ratio for each divided patch to minimize the failure of feature extraction for the boundary surface of each divided patch; and a 3D patch segmentation network implemented with a U-net structure to predict the cerebral aneurysm region present in each divided patch in pixel units.

[0011] Additionally, the 3D patch segmentation network may include: a contracting path that performs the role of extracting features of cerebral aneurysms present in a patch while reducing the size of the input patch by half (1 / 2); a skip path that connects the contracting path and the expansion path; a bottom layer that extracts additional feature information based on the features extracted through the contracting path; an expanding path that combines the first feature information, which has a feature size doubled after passing through the bottom layer from the contracting path, with the second feature information connected through the skip path, and then uses the combined feature information as input to predict the location of cerebral aneurysms present in a patch in pixel units; and a global positioning network that performs the role of learning the physical location information of the patch and, as a result of learning, distinguishes between patches containing blood vessels and patches that do not.

[0012] In addition, when operating a patient model, the patient model operation processor applies the results of the patch model to multiple patient models with different characteristics and fuses the cerebral aneurysm prediction results of each patient model to determine cerebral aneurysms at the patient level.

[0013] Additionally, the patient model may include: a first patient model comprising a patch overlap module that extracts patches by overlapping them when extracting patches from the patch model, and a noise reduction module that removes noise pixels by utilizing morphological techniques from the patches extracted through the patch overlap module; a second patient model comprising a patch slice module that extracts patches in the form of multiple slices corresponding to the number of CT Angiography (CTA) slices of the patient when extracting patches from the patch model, and a sequential network that performs patient-unit prediction by considering the characteristics of the patch slices extracted through the patch slice module and the order between the slices; and a fusion module that fuses the cerebral aneurysm prediction results of the first patient model and the cerebral aneurysm prediction results of the second patient model.

[0014] And the fusion module may adopt the cerebral aneurysm prediction result of the second patient model as the final prediction result if the prediction result of the second patient model is smaller than a preset lower threshold or larger than a preset upper threshold, and may adopt the cerebral aneurysm prediction result of the first patient model as the final prediction result if the prediction result of the second patient model is greater than or equal to the preset lower threshold and less than or equal to the preset upper threshold.

[0015] In addition, the information processing unit may further include a post-processing processor that performs an additional task to reduce false positives on the prediction results of cerebral aneurysms calculated through the patch model and the patient model.

[0016] Meanwhile, according to another embodiment of the present invention, an artificial intelligence method for detecting cerebral aneurysms using 3D angiography CT comprises: an input unit acquiring a CT image of a patient; an information processing unit predicting the location of a cerebral infarction in patch units by utilizing a patch model; and an information processing unit applying the results of the patch model to a patient model to predict the presence or absence of a cerebral aneurysm in patient units. Effects of the invention

[0018] As described above, according to embodiments of the present invention, by using a deep learning artificial intelligence algorithm based on a convolutional neural network (CNN) and a recurrent neural network (RNN) on a CT angiography (CTA) image of the brain using a contrast agent, the presence or absence of a cerebral aneurysm in a patient can be determined, thereby contributing to the improvement of the accuracy of cerebral aneurysm diagnosis. Brief explanation of the drawing

[0020] FIG. 1 is a drawing provided in the description of an artificial intelligence cerebral aneurysm detection system using three-dimensional angiographic CT according to an embodiment of the present invention. FIG. 2 is a drawing provided for explaining the operation of an information processing unit according to an embodiment of the present invention. FIG. 3 is a drawing provided for the structural description of a patch model according to one embodiment of the present invention, FIG. 4 is a drawing provided for the structural description of a first patient model (patient model A) according to one embodiment of the present invention, FIG. 5 is a drawing provided for the structural description of a second patient model (patient model B) according to one embodiment of the present invention, FIG. 6 is a drawing provided for the description of a fusion module according to an embodiment of the present invention, FIG. 7 is a diagram illustrating an algorithm of a post-processing processor according to an embodiment of the present invention, and FIG. 8 is a flowchart provided in the description of an artificial intelligence cerebral aneurysm detection method using 3D angiography CT according to one embodiment of the present invention. Specific details for implementing the invention

[0021] The present invention will be described in more detail below with reference to the drawings. To clearly explain the invention, parts unrelated to the description have been omitted from the drawings, and in the drawings, the width, length, thickness, etc., of the components may be exaggerated for convenience.

[0022] FIG. 1 is a drawing provided for the description of an artificial intelligence cerebral aneurysm detection system using a three-dimensional angiographic CT according to one embodiment of the present invention, and FIG. 2 is a drawing provided for the description of the operation of an information processing unit according to one embodiment of the present invention.

[0023] The artificial intelligence brain aneurysm detection system using 3D angiography CT according to the present embodiment (hereinafter collectively referred to as the 'brain aneurysm detection system') is provided to determine whether a patient has a brain aneurysm by using a deep learning artificial intelligence algorithm based on a convolutional neural network (CNN) and a recurrent neural network (RNN) on brain computed tomography (CT Angiography) images using a contrast agent.

[0024] To this end, the cerebral aneurysm detection system may include an input unit (110), an information processing unit (120), and a storage unit (130).

[0025] The input unit (110) is equipped with a communication module connected to a network, and can receive a CT image of a patient (e.g., CT angiography image using contrast agent) necessary for the information processing unit (120) to operate, and can receive input for a specific command requested by a user and transmit it to the information processing unit (120).

[0026] The storage unit (130) is a storage medium that stores programs and data necessary for the information processing unit (120) to operate.

[0027] The information processing unit (120) can provide an analysis (determination) result regarding the presence or absence of a patient's cerebral aneurysm by utilizing a deep learning-based artificial intelligence algorithm in conjunction with a pre-approved medical image transmission device software (PACS).

[0028] Specifically, the information processing unit (120) may include a patch model operating processor (121) that operates a patch model for predicting the location of a cerebral infarction on a patch basis, a patient model operating processor (122) that operates a patient model for predicting the presence or absence of a cerebral aneurysm on a patient basis based on the results of the patch model, and a post-processing processor (123) that performs additional post-processing tasks.

[0029] That is, the information processing unit (120) can divide the CTA slices into small-sized patches, perform patch-unit analysis, and then combine the patch analysis results on a patient-unit basis to finally predict the presence or absence of a cerebral aneurysm and the location of the cerebral aneurysm on a patient-unit basis.

[0030] FIG. 3 is a drawing provided to describe the structure of a patch model according to one embodiment of the present invention.

[0031] Because the size of the cerebral aneurysm is very small and the number of CTA slices is high, averaging about 170 per patient, the standard slice processing method cannot achieve the desired performance.

[0032] Therefore, the patch model can perform the role of extracting the characteristics of cerebral aneurysms on a patch-by-patch basis after dividing them into patches of a certain size.

[0033] To this end, the patch model may include a patch extraction module and a 3D patch segmentation network to predict the location of cerebral infarction on a patch-by-patch basis.

[0034] The total slice size of the patient can be expressed as (Z, Y, X).

[0035] Here, Z corresponds to the number of slices of the patient, Y to the vertical size of the slice, and X to the horizontal size of the slice. Among these, the Z value varies from patient to patient because the size of the face varies from patient to patient. As a result of the analysis, the Z value may be average (170, 512, 512).

[0036] In this case, to process in patch units, (170, 512, 512) is a patch (Z) of a fixed size. patch , Y patch , X patch Although it can be divided into ) but when dividing a patch, if there is a cerebral aneurysm at the boundary of the patch, the patch model may not be able to properly extract the features of the patch.

[0037] Therefore, when splitting patches, it is necessary to extract patches by overlapping them by a certain size.

[0038] Accordingly, the patch extraction module divides the patient's CT image into patches of a preset size, and in order to minimize the failure of feature extraction for the boundaries of each divided patch, each divided patch may include an overlap portion at a certain ratio.

[0039] A 3D patch segmentation network can be implemented as a U-net structure to predict the cerebral aneurysm region present within each segmented patch in pixel units.

[0040] Specifically, the 3D patch segmentation network may include a contracting path, a skip path, a bottom layer, an expanding path, and a global positioning network.

[0041] The contracting path can perform the role of extracting cerebral aneurysm features present within the patch while reducing the size of the input patch by half (1 / 2).

[0042] Specifically, the contracting path uses a 3D convolutional neural network as its basis and can reduce the patch size using max pooling.

[0043] A skip path can connect a contraction path and an expansion path.

[0044] In this case, a skip path is necessary for the following reasons.

[0045] 1) As it passes through the contraction path, location information of the cerebral aneurysm region is gradually lost, while conversely, overall characteristic information of the cerebral aneurysm becomes distinct.

[0046] 2) Features are gradually restored to their original size through an expansion path, and missing location information is required during this process, and location information can be obtained through a skip path.

[0047] The bottom layer can extract additional feature information from the features extracted through the shrinkage path. In this case, the most compressed feature information can actually be delivered to the bottom layer.

[0048] The expanding path combines the first feature information, which has a feature size doubled after passing through the bottom layer in the contraction path, and the second feature information connected through the skip path into one, and then uses the combined feature information as input to predict the location of the cerebral aneurysm within the patch in pixel units.

[0049] In this case, the patch model can predict the location of cerebral aneurysms within the patch in pixel units through an expansion path.

[0050] Since cerebral aneurysms occur only along the blood vessels of the brain, if the physical location of the patch is known, it becomes easy to determine that a patch that does not involve a blood vessel is not a cerebral aneurysm patch.

[0051] Therefore, the global positioning network performs the role of learning the physical location information of the patches, and through this learning, it can consequently distinguish between patches containing blood vessels and patches that do not.

[0052] FIG. 4 is a drawing provided for the structural description of a first patient model (patient model A) according to an embodiment of the present invention, and FIG. 5 is a drawing provided for the structural description of a second patient model (patient model B) according to an embodiment of the present invention.

[0053] The patient model can extract characteristic information of cerebral aneurysms on a patch-by-patch basis from the patch model, aggregate all patches at the patient level, and then predict the presence or absence of cerebral aneurysms at the patient level.

[0054] In this case, the patient model can implement various methods for collecting patches at the patient level, and since the performance of the patient model may vary depending on the method of collecting patches at the patient level, the presence or absence of cerebral aneurysms can be predicted at the patient level by utilizing multiple patient models with different characteristics.

[0055] Specifically, when operating a patient model, the patient model operation processor applies the results of a patch model to multiple patient models with different characteristics and fuses the cerebral aneurysm prediction results of each patient model to determine a cerebral aneurysm at the patient level.

[0056] For example, the patient model may include a first patient model (patient model A), a second patient model (patient model B) in which the method of collecting patches is implemented differently from the first patient model, and a fusion module that fuses the cerebral aneurysm prediction results of the first patient model and the cerebral aneurysm prediction results of the second patient model.

[0057] Specifically, the first patient model (patient model A) may include a patch overlap module that extracts patches by overlapping them when extracting patches from a patch model as exemplified in FIG. 4, and a noise reduction module that removes noise pixels by utilizing morphological techniques from the patches extracted through the patch overlap module.

[0058] The patch overlap module can extract patches by overlapping them when extracting patches from a patch model, and by remembering the extraction location of each patch, it can reconstruct the patient's (Z, Y, X) from the patches through the reverse process of extraction.

[0059] In this case, the values ​​of the pixels in the patch correspond to the probability indicating whether or not the pixel is a cerebral aneurysm. For pixels in the overlapping area, these values ​​are all added together and then divided by the number of overlaps to obtain the result.

[0060] The patch model essentially takes the patch as input and performs the role of predicting whether or not every pixel within the patch is a cerebral aneurysm pixel.

[0061] As a result, some isolated pixels are identified as cerebral aneurysm pixels, or some of the cerebral aneurysm pixels clustered in an oval shape are predicted to be normal pixels.

[0062] Since all such pixels can be viewed as noise, the noise reduction module uses morphological techniques to remove abnormal pixels corresponding to this noise.

[0063] And the noise reduction module can subsequently remove brain aneurysm pixels consisting of pixels that are too small, and additionally remove brain aneurysm pixel regions that were incorrectly predicted to be located outside the skull.

[0064] The first patient model restored the patient's (Z, Y, X) by reversing the patches extracted from the patch model, but the second patient model collected the patches in parts (Z slice It can be made into a kind of multiple slice form having , Y, X).

[0065] The reason for doing this is to perform patient-level predictions through training using these slices. On the other hand, the first patient model does not perform separate training for patient-level predictions.

[0066] To this end, the second patient model (patient model B) may include a patch slice module that extracts the patch in the form of multiple slices equal to the number of CT Angiography (CTA) slices of the patient when extracting the patch from the patch model as exemplified in FIG. 5, and a sequential network that performs patient-unit prediction by considering the characteristics of the patch slices extracted through the patch slice module and the order between the slices.

[0067] A sequential network can make patient-level predictions by considering the features of patch slices and even the order between them, and can utilize stacked recurrent neurons consisting of a total of three layers to correct feature errors that may occur in the patch model during training.

[0068] FIG. 6 is a drawing provided to describe a fusion module according to one embodiment of the present invention.

[0069] The fusion module may adopt the cerebral aneurysm prediction result of the second patient model as the final prediction result if the result is smaller than a preset lower threshold or larger than a preset upper threshold.

[0070] Conversely, if the fusion module is above a preset lower threshold and below a preset upper threshold, the cerebral aneurysm prediction result of the first patient model can be adopted as the final prediction result.

[0071] FIG. 7 is a diagram illustrating an algorithm of a post-processing processor according to one embodiment of the present invention.

[0072] The cerebral aneurysm detection system according to the present embodiment can predict the presence or absence of a cerebral aneurysm on a patient-by-patient basis by passing from a patch model to a patient model.

[0073] Meanwhile, various experimental results generally show a tendency for false positives. Such false positives mean that patients without cerebral aneurysms are identified as having them.

[0074] Although most studies on cerebral aneurysms state that false positives are not a problem, there is a need to further reduce them when considering numerical performance.

[0075] Accordingly, the post-processing processor (123) can perform additional post-processing work to reduce false positives on the predicted results of cerebral aneurysms calculated through the patch model and the patient model. At this time, the algorithm of the post-processing processor is as exemplified in FIG. 7.

[0076] (1) A 3D mask image that can be expressed as (Z, Y, X) by collecting patch model results on a patient-by-patient basis ( mask patient You can obtain ). The value of each pixel in this mask is a probability value indicating whether or not the pixel corresponds to a cerebral aneurysm.

[0077] (2) 3D cerebral aneurysm volume area in this 3D mask ( Aneurysm_volume k Extracts )

[0078] (3) Calculate the length of the largest major axis for each in the volume area and use this as the length of the cerebral aneurysm ( Aneurysm_length k It is considered as ).

[0079] (4) If the length of the cerebral aneurysm is smaller than the threshold value, the corresponding cerebral aneurysm region is removed from the mask image. As a result, if no cerebral aneurysm region exists in the mask, the patient is considered to be without a cerebral aneurysm. That is, in the post-processing algorithm figure above Aneursym_set This applies when no value exists in.

[0080] FIG. 8 is a flowchart provided in the description of an artificial intelligence cerebral aneurysm detection method using 3D angiography CT according to one embodiment of the present invention.

[0081] The artificial intelligence method for detecting a cerebral aneurysm using a three-dimensional angiographic CT according to the present embodiment can be executed by the cerebral aneurysm detection system described above with reference to FIGS. 1 to 7.

[0082] Referring to FIG. 8, when a CT image of a patient is obtained through the input unit (110) (S810), the information processing unit (120) can predict the location of the cerebral infarction in patch units by utilizing a patch model (S820).

[0083] Afterwards, the information processing unit (120) can apply the results of the patch model to the patient model to predict whether there is a brain aneurysm on a patient-by-patient basis (S830).

[0084] And the information processing unit (120) can perform additional post-processing work to reduce false positives on the predicted results of cerebral aneurysms calculated through the patch model and the patient model (S840).

[0085] So far, an artificial intelligence method and system for detecting cerebral aneurysms using 3D angiography CT has been described in detail with preferred embodiments.

[0086] According to embodiments of the present invention, by using a deep learning artificial intelligence algorithm based on a convolutional neural network (CNN) and a recurrent neural network (RNN) on a CT angiography (CTA) image of a brain using a contrast agent, the presence or absence of a cerebral aneurysm in a patient can be determined, thereby contributing to the improvement of the accuracy of cerebral aneurysm diagnosis.

[0087] Meanwhile, it goes without saying that the technical concept of the present invention may also be applied to a computer-readable recording medium containing a computer program that enables the device and method according to the present embodiment to perform their functions. Furthermore, the technical concept according to various embodiments of the present invention may be implemented in the form of computer-readable code recorded on a computer-readable recording medium. A computer-readable recording medium may be any data storage device that can be read by a computer and store data. For example, a computer-readable recording medium may be a ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical disk, hard disk drive, etc. Additionally, computer-readable code or a program stored on a computer-readable recording medium may be transmitted through a network connected between computers.

[0088] That is, the artificial intelligence method for detecting cerebral aneurysms using 3D angiography CT according to the present embodiment can be implemented as a standalone software installed on customer-owned hardware at a customer-designated location, and can be used in conjunction with customer-owned medical image transmission device software. In addition, some upgrades and customization may be required to link with customer-owned medical image transmission device software.

[0089] Furthermore, although preferred embodiments of the present invention have been illustrated and described above, the present invention is not limited to the specific embodiments described above. Various modifications are possible by those skilled in the art without departing from the essence of the invention as claimed in the claims, and such modifications should not be understood individually from the technical spirit or perspective of the present invention. Explanation of the symbols

[0091] 110 : Input section 120 : Information processing unit 121 : Patch Model Operating Processor 122 : Patient Model Operation Processor 123 : Post-processor 130 : Storage unit

Claims

Claim 1 The system includes an input unit for acquiring a patient's CT image; and an information processing unit comprising a patch model operating processor for operating a patch model that predicts the location of a cerebral infarction in patch units, and a patient model operating processor for operating a patient model that predicts the presence or absence of a cerebral aneurysm in patient units based on the results of the patch model; wherein the patch model includes a patch extraction module that divides the patient's CT image into patches of a preset size, and ensures that each divided patch includes an overlapping portion at a certain ratio to minimize failure in feature extraction for the boundary surface of each divided patch. An artificial intelligence brain aneurysm detection system using 3D angiography CT, characterized by comprising: a 3D patch segmentation network implemented with a U-net structure to predict brain aneurysm regions present within each segmented patch in pixel units; wherein the 3D patch segmentation network comprises: a contracting path that performs the role of extracting brain aneurysm features present within a patch while reducing the size of the input patch by half (1 / 2); a skip path that connects the contracting path and the expansion path; a bottom layer that extracts additional feature information based on the features extracted through the contracting path; an expanding path that combines first feature information, which has a feature size doubled after passing through the bottom layer from the contracting path, with second feature information connected through the skip path, and then uses the combined feature information as input to predict the location of brain aneurysms present within the patch in pixel units; and a global positioning network that performs the role of learning physical location information of the patch and, as a result of learning, distinguishes between patches containing blood vessels and patches that are not. Claim 2 delete Claim 3 delete Claim 4 An artificial intelligence cerebral aneurysm detection system using 3D angiography CT according to claim 1, wherein the patient model operation processor, when operating a patient model, applies the results of a patch model to a plurality of patient models with different characteristics and fuses the cerebral aneurysm prediction results of each patient model to determine a patient-unit cerebral aneurysm. Claim 5 The system includes an input unit for acquiring a CT image of a patient; and an information processing unit comprising a patch model operating processor for operating a patch model that predicts the location of a cerebral infarction in patch units, and a patient model operating processor for operating a patient model that predicts the presence or absence of a cerebral aneurysm in patient units based on the results of the patch model; wherein the patient model operating processor, when operating the patient model, applies the results of the patch model to a plurality of patient models with different characteristics, fuses the cerebral aneurysm prediction results of each patient model, and determines a cerebral aneurysm in patient units; and the patient model includes a first patient model comprising a patch overlap module that extracts patches by overlapping them when extracting patches from the patch model, and a noise reduction module that removes noise pixels using morphological techniques from the patches extracted through the patch overlap module. An artificial intelligence cerebral aneurysm detection system using 3D angiography CT, characterized by comprising: a second patient model including a patch slice module that extracts patches in the form of multiple slices equal to the number of CT angiography (CTA) slices of the patient when extracting patches from a patch model, and a sequential network that performs patient-unit prediction by considering the characteristics of the patch slices extracted through the patch slice module and the order between the slices; and a fusion module that fuses the cerebral aneurysm prediction results of the first patient model and the cerebral aneurysm prediction results of the second patient model. Claim 6 An artificial intelligence brain aneurysm detection system using 3D angiography CT according to claim 5, wherein the fusion module adopts the brain aneurysm prediction result of the second patient model as the final prediction result if the brain aneurysm prediction result of the second patient model is smaller than a preset lower threshold or larger than a preset upper threshold, and adopts the brain aneurysm prediction result of the first patient model as the final prediction result if the brain aneurysm prediction result of the second patient model is greater than or equal to the preset lower threshold and less than or equal to the preset upper threshold. Claim 7 An artificial intelligence cerebral aneurysm detection system using 3D angiography CT according to claim 1, wherein the information processing unit further includes a post-processing processor that performs an additional task of reducing false positives on the prediction results of cerebral aneurysms calculated through a patch model and a patient model. Claim 8 The method comprises: an input unit acquiring a CT image of a patient; an information processing unit predicting the location of a cerebral infarction in patch units by utilizing a patch model; and an information processing unit applying the results of the patch model to a patient model to predict the presence or absence of a cerebral aneurysm in patient units; wherein the patch model includes a patch extraction module that divides the patient's CT image into patches of a preset size, and ensures that each divided patch includes an overlapping portion at a certain ratio to minimize failure in feature extraction for the boundary surfaces of each divided patch; An artificial intelligence method for detecting cerebral aneurysms using 3D angiography CT, characterized by including: a 3D patch segmentation network implemented with a U-net structure to predict cerebral aneurysm regions present within each segmented patch in pixel units; wherein the 3D patch segmentation network comprises: a contracting path that performs the role of extracting cerebral aneurysm features present within a patch while reducing the size of the input patch by half (1 / 2); a skip path that connects the contracting path and the expansion path; a bottom layer that extracts additional feature information based on the features extracted through the contracting path; an expanding path that combines first feature information, which has a feature size doubled after passing through the bottom layer from the contracting path, with second feature information connected through the skip path, and then uses the combined feature information as input to predict the location of the cerebral aneurysm present within the patch in pixel units; and a global positioning network that performs the role of learning physical location information of the patch and, as a result of learning, distinguishes between patches containing blood vessels and patches that are not.

Images