WO2019189972A1 - 치매를 진단을 하기 위해 홍채 영상을 인공지능으로 분석하는 방법 - Google Patents
치매를 진단을 하기 위해 홍채 영상을 인공지능으로 분석하는 방법 Download PDFInfo
- Publication number
- WO2019189972A1 WO2019189972A1 PCT/KR2018/003813 KR2018003813W WO2019189972A1 WO 2019189972 A1 WO2019189972 A1 WO 2019189972A1 KR 2018003813 W KR2018003813 W KR 2018003813W WO 2019189972 A1 WO2019189972 A1 WO 2019189972A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- dementia
- region
- interest
- image
- neural network
- Prior art date
Links
- 206010012289 Dementia Diseases 0.000 title claims abstract description 99
- 238000000034 method Methods 0.000 title claims abstract description 51
- 238000013473 artificial intelligence Methods 0.000 title claims abstract description 15
- 230000003902 lesion Effects 0.000 claims abstract description 51
- 238000013528 artificial neural network Methods 0.000 claims abstract description 49
- 238000001514 detection method Methods 0.000 claims abstract description 22
- 230000011218 segmentation Effects 0.000 claims abstract description 15
- 238000003745 diagnosis Methods 0.000 claims description 20
- 230000004913 activation Effects 0.000 claims description 11
- 238000013527 convolutional neural network Methods 0.000 claims description 9
- 238000012360 testing method Methods 0.000 claims description 8
- 208000024827 Alzheimer disease Diseases 0.000 claims description 7
- 210000001652 frontal lobe Anatomy 0.000 claims description 7
- 201000004810 Vascular dementia Diseases 0.000 claims description 6
- 208000009829 Lewy Body Disease Diseases 0.000 claims description 4
- 201000002832 Lewy body dementia Diseases 0.000 claims description 4
- 238000009533 lab test Methods 0.000 claims description 4
- 239000002131 composite material Substances 0.000 claims 1
- 238000005457 optimization Methods 0.000 claims 1
- 238000003384 imaging method Methods 0.000 abstract 1
- 230000007935 neutral effect Effects 0.000 abstract 1
- 230000006870 function Effects 0.000 description 19
- 238000010586 diagram Methods 0.000 description 12
- 238000012545 processing Methods 0.000 description 11
- 210000002569 neuron Anatomy 0.000 description 9
- 239000000284 extract Substances 0.000 description 7
- 238000011176 pooling Methods 0.000 description 5
- 210000004556 brain Anatomy 0.000 description 4
- 210000001519 tissue Anatomy 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 3
- 210000005036 nerve Anatomy 0.000 description 3
- 102000004169 proteins and genes Human genes 0.000 description 3
- 108090000623 proteins and genes Proteins 0.000 description 3
- 210000003478 temporal lobe Anatomy 0.000 description 3
- 230000002159 abnormal effect Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000007796 conventional method Methods 0.000 description 2
- 201000010099 disease Diseases 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000036541 health Effects 0.000 description 2
- 210000000056 organ Anatomy 0.000 description 2
- 230000002360 prefrontal effect Effects 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 208000000044 Amnesia Diseases 0.000 description 1
- 208000037259 Amyloid Plaque Diseases 0.000 description 1
- 208000018737 Parkinson disease Diseases 0.000 description 1
- 206010034719 Personality change Diseases 0.000 description 1
- 208000006011 Stroke Diseases 0.000 description 1
- 210000000467 autonomic pathway Anatomy 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 208000013404 behavioral symptom Diseases 0.000 description 1
- 238000009534 blood test Methods 0.000 description 1
- 210000000133 brain stem Anatomy 0.000 description 1
- 210000004027 cell Anatomy 0.000 description 1
- 210000001175 cerebrospinal fluid Anatomy 0.000 description 1
- 210000004720 cerebrum Anatomy 0.000 description 1
- 230000003930 cognitive ability Effects 0.000 description 1
- 208000010877 cognitive disease Diseases 0.000 description 1
- 230000006378 damage Effects 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 208000011977 language disease Diseases 0.000 description 1
- 210000004558 lewy body Anatomy 0.000 description 1
- 231100000863 loss of memory Toxicity 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000001087 myotubule Anatomy 0.000 description 1
- 210000001640 nerve ending Anatomy 0.000 description 1
- 210000000653 nervous system Anatomy 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 210000002682 neurofibrillary tangle Anatomy 0.000 description 1
- 238000002610 neuroimaging Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 230000001953 sensory effect Effects 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 208000024891 symptom Diseases 0.000 description 1
- 230000009885 systemic effect Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 238000012549 training Methods 0.000 description 1
- 230000003966 vascular damage Effects 0.000 description 1
Images
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
Definitions
- the present invention relates to a method of artificially analyzing an iris image for diagnosing dementia, and more specifically, to predict the type and prognostic diagnosis of dementia, and a lightweight neural network that can be diagnosed by a low-performance mobile device.
- the present invention relates to a method for designing and showing a location of a corresponding lesion area in order to increase the reliability of a dementia patient.
- Dementia is a condition in which humans have difficulties in daily or social life due to disorders in temporal and frontal lobe executive functions including memory, attention, language, and spatiotemporal ability. Dementia is classified into Alzheimer's dementia, vascular dementia, Lewy body dementia, and prefrontal dementia according to the cause of dementia. Alzheimer's dementia accounts for 70% of people with dementia today.
- Alzheimer's dementia gradually decreases cognitive abilities and often begins with loss of memory.
- the cause is Alzheimer's by killing cells by interlocking abnormal protein clumps called amyloid plaques and clumps of proteins called neurofibrillary tangles.
- Vascular dementia is a cognitive disorder caused by vascular damage in the brain. So stroke patients are likely to get dementia. In addition, it can appear similar to Alzheimer's disease, and it is very common to occur in parallel with vascular dementia.
- Lewy body dementia involves an abnormal mass of protein called alphacituclein in certain parts of the brain. This results in changes in movement, thinking and behavior. In addition, a lot of fluctuations occur in attention and thinking, and if the movement symptoms appear first, it is often diagnosed with Parkinson's disease.
- Frontal lobe dementia also known as prefrontal dementia, occurs when there is gradual damage to the frontal and temporal lobes of the brain. Behavioral symptoms and personality changes in the frontal lobe, and language disorders in the temporal lobe.
- the iris is an extension of the brain and has hundreds of thousands of nerve endings (autonomic nerves, middle nerves, sensory nerves), capillaries, and muscle fibers. Therefore, the iris is connected to all organs and tissues through the brain and nervous system, and thus can serve as a direct diagnostic indicator of systemic health.
- nerve endings autonomous nerves, middle nerves, sensory nerves
- capillaries capillaries
- muscle fibers muscle fibers. Therefore, the iris is connected to all organs and tissues through the brain and nervous system, and thus can serve as a direct diagnostic indicator of systemic health.
- iris science we immediately see the benefits and from all the changes on the iris diagnose the disease state of the tissues involved and figure out what the body needs.
- the iris is the most complex fibrous structure in our body and is connected to the cerebrum and parts of the body through nerves, information about chemical and physical changes occurring in each tissue and organ in the body is transmitted to the vibration. It changes the shape of the fibrous tissue. It can be used as a basis for diagnosing dementia through iris because it can be used to read and diagnose the status of personal health, response to treatment, the human skeleton, recovery and progression of disease.
- the conventional methods for diagnosing dementia are divided into a paper test and a laboratory test, which include an MRI or CT brain imaging test, a blood test, and a cerebrospinal fluid test.
- the neural network can be divided into a convolutional layer and a fully-connected layer (hereinafter referred to as an FC layer).
- FC layer loses spatial information that was originally a three-dimensional image by converting features extracted from the convolution layer into a one-dimensional vector, and most computations of the deep neural network are heavily focused on the FC layer.
- the processing speed is very slow.
- the object of the present invention is that the present invention compensates for the very slow processing speed, which is a problem of the conventional dementia diagnosis apparatus, and ultimately reduces the weight to focus on the processing speed that enables real-time diagnosis even in a mobile device having lower hardware performance than the desktop. Not only does it provide a neural network, but it is also able to accurately diagnose with high accuracy.
- Another object of the present invention is to provide a confidence in a patient by showing a statistical result for the diagnosis of dementia and visualizing the location of the lesion.
- Still another object of the present invention is to diagnose the likelihood of dementia and the degree of dementia progression according to the detection and analysis result, based on the big data indicating the dementia potential and the degree of dementia progression according to the location and shape of the lesion area, and adding as necessary. It is to notify in real time that inspection is necessary.
- a method of analyzing an iris image by artificial intelligence receives an input image photographing the user's eyes from the user terminal Doing; Extracting a region of interest to extract an iris from the input image; Resizing the extracted region of interest into a square shape and scaling the size of the region of interest; Applying a deep neural network to the resized and scaled region of interest; Detecting a lesion area by applying detection and segmentation to an image obtained by applying the deep neural network; Diagnosing dementia by determining the location of the lesion area through the detection and determining the shape of the lesion area through the segmentation, and extracting the region of interest from the input image.
- the method may further include extracting the region of interest, which is a minimum region necessary for extracting the iris by excluding an region not used for, and applying the deep neural network to the extracted image of the region of interest in a square form. Resizing and normalizing and converting the pixel information values into values between 0 and 1, compressing and optimizing the data into a single data, and diagnosing the dementia may include: And diagnosing the type of dementia based on the shape.
- the extracting the region of interest may include: using the virtual axis preset in advance when the input image is inclined with respect to a vertical direction.
- the method may further include extracting the region of interest after aligning by an inclination angle with respect to the vertical direction.
- the resizing and scaling may include: resizing the ROI in a square shape, normalizing pixel information values to values between 0 and 1, and The method may further include optimizing data of the iris image by converting the pixel information value into a byte and compressing the ROI into one data.
- a method of artificially analyzing an iris image includes artificial convolutional neural networks (CNNs), thereby preventing loss of spatial information of the iris image. It features.
- CNNs convolutional neural networks
- a method of analyzing an iris image by artificial intelligence includes: a user terminal including a camera unit, and the camera unit includes a general mobile camera and an iris recognition dedicated camera, or the camera unit uses only iris recognition. A lens is attached.
- a method of artificially analyzing an iris image may include: applying the deep neural network, using separable and expanded convolution. It is characterized by.
- the method of artificially analyzing an iris image may further include generating a visualized image based on a diagnosis of dementia based on the location and shape of the lesion area. .
- the method of artificially analyzing an iris image according to an embodiment of the present invention may further include performing a prognostic diagnosis of dementia based on the location and shape of the lesion area.
- a method of analyzing an iris image by artificial intelligence includes: accumulating big data indicating a dementia potential and a degree of dementia progression according to the location and shape of the lesion area; Determining the likelihood of dementia and the degree of dementia progression according to the location and shape of the lesion area based on the big data; In accordance with the possibility of dementia and the degree of dementia, further comprising the step of real-time notification through the user terminal that further testing, including a paper test and laboratory tests, the type of dementia Alzheimer's dementia, vascular dementia, Lewy body It includes dementia and frontal lobe dementia, the likelihood of dementia is divided by a percentage, the degree of dementia progression is characterized by early, middle and late stages.
- an activation function and a focal loss method are used for the multiplicative neural network.
- the diagnosis according to the prognostic diagnosis and classification of dementia focuses not only on accuracy but also on processing speed, and shows the statistical result of the diagnosis and visualizes the location of the lesion to give the patient confidence.
- the lightweight neural network focuses on processing speed, ultimately enabling diagnosis on mobile devices with low hardware in real time.
- the present invention based on the big data indicating the dementia potential and the degree of dementia progression according to the location and shape of the lesion area, the possibility of dementia and the degree of dementia progression according to the detection and analysis results, and if necessary further examination Allows notification in real time, such as a push alarm.
- FIG. 1 is a conceptual diagram showing an entire system according to an embodiment of the present invention.
- FIG. 2 is a block diagram showing the configuration of a system according to an embodiment of the present invention.
- FIG. 3 is a flowchart illustrating a method of artificially analyzing an iris image for diagnosing dementia according to an embodiment of the present invention.
- FIG. 4 is a detailed block diagram of a deep neural network learning portion according to an embodiment of the present invention.
- FIG. 5 is a diagram illustrating general convolution and atrous convolution according to an embodiment of the present invention.
- FIG. 6 is a diagram illustrating normal and separable convolutions according to an embodiment of the present invention.
- FIG. 7 is an exemplary view showing a lesion location diagnosed by a neural network according to an embodiment of the present invention by a heat map method.
- FIG. 8 is a block diagram that detects features that have undergone convolution according to an embodiment of the present invention.
- FIG. 9 is a graph of an activation function of a deep neural network according to an embodiment of the present invention.
- ... unit described in the specification means a unit for processing at least one function or operation, which may be implemented in hardware or software or a combination of hardware and software.
- “a” or “an”, “one”, and the like shall be, in the context of describing the present invention, both singular and plural unless the context clearly dictates otherwise or is clearly contradicted by the context. It can be used as a meaning including.
- FIG. 1 is a conceptual diagram showing an entire system according to an embodiment of the present invention.
- the user terminal may be an electronic device including a smartphone, a tablet PC, a laptop, and the like.
- the user terminal may include a camera, and the camera may include an iris recognition dedicated camera or an iris recognition lens may be attached to the camera. Therefore, when the user's eyes are photographed using the user terminal, a screen or an image having a high resolution for a specific part including an iris can be secured.
- An image obtained by photographing an eye of a user through the user terminal may be input to a server and used for diagnosing dementia.
- FIG. 2 is a block diagram showing the configuration of a system according to an embodiment of the present invention.
- the system includes an extractor 110 that extracts a region of interest (RoI) to extract an iris from an input image, and polygons the extracted region of interest. That is, a resizing in a square shape having various sizes, a preprocessor 120 for scaling the size of the ROI, a learning unit 130 for applying a deep neural network to the resized and scaled ROI, and the deep neural network are applied.
- the detection unit 140 detects the lesion area by applying detection and segmentation to the obtained image and determines the position of the lesion area through the detection, and the shape of the lesion area through the segmentation. By determining this, it may include a diagnostic unit 150 for diagnosing dementia.
- the extraction unit 110 may acquire an image even from a low performance mobile device.
- the extraction unit 110 may remove an unnecessary image part in diagnosis for fast processing speed after obtaining images to be used to diagnose dementia.
- the iris images may not exclude tilted images, the iris images may be extracted using the preset virtual axis and extracted only from the eye image.
- the region of interest refers to a minimum region for extracting the iris necessary for diagnosing dementia except for an unnecessary region.
- the reason to extract the region of interest is to reduce the amount of computation as much as possible by eliminating unnecessary regions for diagnosis of dementia in order to reduce the weight.
- the unnecessary region may be processed in gray and transferred to the preprocessor 120.
- the preprocessor 120 may resize the iris image obtained by the extractor 110 to a square size (NxN).
- NxN square size
- the learner 130 may include a convolutional neural network (CNN) capable of learning while maintaining spatial information of the iris image.
- CNN convolutional neural network
- the learner 130 and the detector 140 will be described in detail with reference to FIGS. 4 and 8, respectively.
- the diagnosis unit 150 finds the location and shape of the lesion area in detail in the detection unit 140 and extracts the segmentation and detection of the lesions from the learning unit 130.
- the type of dementia can be diagnosed by analyzing and classifying the diagnosis of dementia based on.
- FIG. 3 is a flowchart illustrating a method of artificially analyzing an iris image for diagnosing dementia according to an embodiment of the present invention.
- an input image capturing an eye of a user may be received from a user terminal (S11), and a region of interest may be extracted to extract an iris from the input image (S12). Since the user terminal includes a camera dedicated to iris recognition or uses a camera equipped with an iris recognition dedicated lens, the user terminal may easily extract an ROI including an iris at a higher resolution than a general camera.
- the extracted ROI may be resized in a polygonal shape, the size of the ROI may be scaled (S13), and a deep neural network may be applied to the ROG that has been resized and scaled (S14).
- the 'various sizes' may include enlarging and reducing the size of an image while maintaining a square shape.
- an area that is not extracted as the region of interest is an unnecessary region so that only the region of interest may be extracted by gray processing. Accordingly, since unnecessary computation amount of the neural network can be reduced, it can be the basis of a lightweight neural network focusing on processing speed.
- detection and segmentation are applied to an image obtained by applying the deep neural network (S15) to detect a lesion area, and the position of the lesion area is determined through the detection, and the segmentation is performed.
- the type of dementia can be determined, and the type of dementia includes, for example, Alzheimer's dementia, vascular dementia, Lewy body dementia, and frontal lobe dementia.
- the dementia potential is determined based on the position and shape of the lesion area. For example, as the color concentration of the lesion area increases, the likelihood of dementia increases as a percentage.
- the degree of dementia progression can be indicated, and can be expressed, for example, in the early, middle and late stages of dementia.
- Figure 4 shows a block diagram of a learning unit 130 for diagnosing dementia according to an embodiment of the present invention.
- the iris image which has undergone the preprocessing unit 120, may be modified to various sizes such as 500x500 and 400x400 as well as 300x300.
- the biggest feature of the deep neural network is first, feature maps are generated by applying various filters to the input image in the convolutional layer. That is, it serves as a kind of template for extracting features of the high-dimensional input image.
- down sampling refers to a neuron layer having reduced spatial resolution with respect to a generated feature map.
- ReLU rectified linear unit
- the factoring method is, for example, a 5x5 filter, which reduces the computation by about 2 to 3 times with a 25:10 ratio in terms of 1x5 + 5x1 filters. Furthermore, assuming that the input video is 7x7, using a 7x7 filter will output 1x1. However, if you use the 3x3 filter, the value of 5x5 is output, and if you use the 3x3 filter, the value of 3x3 is output, and if you use the 3x3 filter, the value of 1x1 is output.
- the present invention utilizes this concept to utilize a combination of a separable convolution method and a factorization method.
- the lesion area has smaller pixel information than the entire iris image during the operation of the learner 130, the more the layers are overlapped, the more the layers are overlapped. There may be limitations in extracting features of lesion areas that have features.
- the present invention can be changed to a specific loss function that is excellent in extracting features of a local area, that is, lesion area, by placing more weight on local features than global features when learning a deep neural network.
- the loss function refers to the deep neural network obtaining the error between the correct answer and the original correct answer with the features extracted from the last convolutional layer, and back-propagating to update the weight by the change amount of the error. It is said to learn.
- the loss function generally used for learning neural networks uses many cross-entropy (CE) loss methods, and the present invention uses a specific loss function as described above. I use it.
- the present invention uses a focal loss method for this purpose because the feature of the lesion area that is much smaller than the feature of the entire iris image should be well learned.
- CE Cross Entropy
- the focus loss method can be used to compensate.
- the size of the lesion to be searched in the iris image is very small compared to the entire image. Therefore, when performing classification on the extracted lesion area, the training is performed with a relatively smaller number of data than the entire image. Because the learning is not performed well, the accuracy is low.
- the size of the image of the feature map extracted from the last convolution layer is up sampled to x2 and combined with the feature map of the layer of m_conv (see FIG. 4). Reinforce.
- a feature map with enhanced detail is received as an input to the detection unit 140 for detecting detection and segmentation, which will be described in detail with reference to FIG. 8.
- the location and shape of the lesion area are determined to generate a visualized image of whether the dementia is diagnosed.
- an atrous convolution is receptive that is grayed out without loss of resolution.
- the size of the field can be extended, the number of points in the circle that performs the convolution is the same, so that information on various scales can be extracted.
- the pooling layer is extracted by extracting a high resolution image similar to the original image. It can be replaced.
- the 3D image is composed of red green blue (RGB), and when a neural network is trained, a general convolution is calculated by overlapping values of R, G, and B in one kernel (hereinafter referred to as a filter). To extract the feature.
- RGB red green blue
- a filter one kernel
- the separable convolution is a bit more color because each of R, G, and B values is separated into one filter, one filter for one value, one for R, one for G, and one for B. It is possible to extract features of the partial part more precisely, and thus to extract more various features instead of extracting duplicated features.
- the general convolution in computational amount has a computational amount of F2NK2M because it calculates and extracts features at the same time, while Separable convolution calculates the values of R, G, and B and generates a filter that extracts the features. By separating these, the calculation amount becomes F2NK2 + F2MN, which can increase the processing speed by 8 to 9 times than the conventional one.
- the neural network extracts the feature of the neural network in the Class Activation Map (CAM), and shows where the region is viewed and diagnosed.
- CAM Class Activation Map
- the FC layer is replaced with a global average pooling (GAP).
- GAP global average pooling
- each feature map becomes a number of neurons through GAP, and the neurons are classified by giving appropriate weights to the neurons.
- the weight map generates a class activation map (CAM) by overlapping with the original iris image.
- CAM class activation map
- FIG. 8 is a diagram illustrating the detection unit 140 of FIG. 2 and a diagram illustrating detection and segmentation region detection of FIG. 4.
- the location and shape of the lesion area are detected by using a feature map of the layer of the last stage of FIG. 4.
- RPN Region Proposal Network
- ROI Region of Interest
- ROI pooling layers convert ROIs of different sizes into the same size.
- Segmentation is performed on a pixel-by-pixel basis of the size of the region of interest that detects the transformed object, but the conventional method has not considered alignment.
- the pixel unit is calculated through the FC layer.
- the FC layer loses the spatial information that was the original three-dimensional image by converting the features extracted from the convolutional layer into a one-dimensional vector, thereby reducing accuracy.
- the present invention keeps the decimal point intact and accurately aligns using bilinear interpolation. Therefore, the present invention can know which lesion location and shape was found and detected.
- the present invention changes to the 1x1 convolutional layer instead of the FC layer to compensate for the shortcomings of the FC layer, because the FC layer connects all neuronal layers to calculate correlations and is called Fully-connected (FC). Calling 1x1 convolution is also convolved through a 1x1 filter, which can improve correlation by computing correlations between pixels and maintaining spatial information.
- FC Fully-connected
- FIG. 9 it is a schematic of an activation function as mentioned in the description of the great features of the deep neural network above.
- an activation function is a function of receiving an input signal and satisfying a certain threshold to produce an output signal and delivering it to the next layer.
- a ReLU rectified linear unit
- ReLU rectified linear unit
- the neural network calculates the error due to the loss function and back propagates the learning. Done. If the backpropagation is calculated by multiplying the derivatives in each layer, the change amount becomes very small toward the first layer, and convergence occurs without transmitting the change amount of the error. If this part is applied to ReLU (rectified linear unit), all values below 0 are treated as 0, so that 0 is undecomposed to 0, which is not easy to learn due to the deep neural network learning by updating the weight.
- the present invention can set the values below 0 to learn as shown in FIG. 7 so that smooth learning can be achieved and thus accuracy can be increased.
- the present invention accumulates big data indicating the dementia potential and the degree of dementia progression according to the position and shape of the lesion area, and based on the big data, the dementia according to the position and shape of the lesion area.
- Learning and determining the likelihood and the degree of dementia progression and can be notified in real time that the additional test, including a questionnaire test and laboratory test is required in accordance with the likelihood of dementia and the degree of dementia progression.
- the progress of the dementia may be notified in real time through the user terminal.
- the real time notification through the user terminal may be configured as a pop-up or push alarm.
- the above-described method may be written as a program executable on a computer, and may be implemented in a general-purpose digital computer which operates the program using a computer readable medium.
- the structure of the data used in the above-described method can be recorded on the computer-readable medium through various means.
- Computer-readable media that store executable computer code for carrying out the various methods of the present invention include, but are not limited to, magnetic storage media (eg, ROM, floppy disks, hard disks, etc.), optical reading media (eg, CD-ROMs, DVDs). Storage media).
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Public Health (AREA)
- Physics & Mathematics (AREA)
- Medical Informatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Pathology (AREA)
- Animal Behavior & Ethology (AREA)
- Molecular Biology (AREA)
- Surgery (AREA)
- Heart & Thoracic Surgery (AREA)
- Biophysics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Veterinary Medicine (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Epidemiology (AREA)
- Primary Health Care (AREA)
- Image Analysis (AREA)
Abstract
Description
Claims (10)
- 스마트폰 상에서 실시간으로 치매를 진단을 하기 위해 홍채 영상을 인공지능으로 분석하는 방법으로서,사용자 단말로부터 사용자의 눈을 촬영한 입력 영상을 수신하는 단계;상기 입력 영상으로부터 홍채를 추출하기 위해 관심 영역(Region of Interest)을 추출하는 단계;추출된 상기 관심 영역을 정사각형 형태로 리사이징하고, 상기 관심 영역의 크기를 스케일링하는 단계;리사이징 및 스케일링된 상기 관심 영역에 대해 심층 신경망을 적용하는 단계;상기 심층 신경망을 적용함으로써 획득된 영상에 대해 디텍션(detection) 및 세그먼테이션(segmentation)을 적용하여 병변 영역을 검출하는 단계;상기 디텍션을 통해 상기 병변 영역의 위치를 판단하고, 상기 세그먼테이션을 통해 상기 병변 영역의 형상을 결정함으로써, 치매를 진단하는 단계를 포함하고,상기 관심 영역을 추출하는 단계는, 상기 입력 영상으로부터 치매 진단에 이용되지 않는 영역을 배제함으로써 홍채를 추출하기 위해 필요한 최소 영역인 상기 관심 영역을 추출하는 단계를 더 포함하고,상기 심층 신경망을 적용하는 단계는, 상기 입력 영상 중 상기 추출한 관심 영역을 정사각형 형태로 리사이징하고 픽셀 정보값을 0과 1사이의 값들로 정규화 및 바이트(byte) 형태로 변환하여 하나의 데이터로 압축 및 최적화하는 단계를 더 포함하고,상기 치매를 진단하는 단계는 상기 병변 영역의 위치 및 형상에 기초하여 치매 종류를 진단하는 단계를 더 포함하는,홍채 영상을 인공지능으로 분석하는 방법.
- 제1항에 있어서,상기 관심 영역을 추출하는 단계는, 상기 입력 영상이 수직 방향에 대해 기울어져 있는 경우, 미리 설정한 가상의 축을 이용하여 상기 수직 방향에 대해 기울어진 각도만큼 정렬시킨 후 상기 관심 영역을 추출하는 단계를 더 포함하는,홍채 영상을 인공지능으로 분석하는 방법.
- 제1항에 있어서,상기 리사이징 및 스케일링하는 단계는, 상기 관심 영역을 정사각형 형태로 리사이징하고, 픽셀 정보값을 0과 1 사이의 값들로 정규화하며, 상기 픽셀 정보값을 바이트(byte) 형태로 변환하고 상기 관심 영역을 하나의 데이터로 압축함으로써 상기 홍채 영상의 데이터를 최적화하는 단계를 더 포함하는,홍채 영상을 인공지능으로 분석하는 방법.
- 제1항에 있어서,상기 심층 신경망은 합성곱 신경망(Convolutional Neural Network: CNN)을 포함함으로써, 상기 홍채 영상의 공간 정보가 손실되는 것을 방지하는,홍채 영상을 인공지능으로 분석하는 방법.
- 제1항에 있어서,상기 사용자 단말은 카메라부를 포함하고,상기 카메라부는 일반 모바일 카메라 및 홍채 인식 전용 카메라를 포함하거나, 상기 카메라부에 홍채 인식 전용 렌즈가 부착되는,홍채 영상을 인공지능으로 분석하는 방법.
- 제4항에 있어서,상기 심층 신경망을 적용하는 단계는, 분리 가능한(Separable) 컨볼루션 및 팽창된(atrous) 컨볼루션을 이용하는 단계를 더 포함하는,홍채 영상을 인공지능으로 분석하는 방법.
- 제1항에 있어서,상기 병변 영역의 위치 및 형상에 기초하여, 치매 진단에 근거가 되는 가시화 이미지를 생성하는 단계를 더 포함하는,홍채 영상을 인공지능으로 분석하는 방법.
- 제1항에 있어서,상기 병변 영역의 위치 및 형상에 기초하여, 치매를 전조 진단하는 단계를 더 포함하는,홍채 영상을 인공지능으로 분석하는 방법.
- 제1항에 있어서,상기 병변 영역의 위치 및 형상에 따른 치매 가능성 및 치매 진행 정도를 나타내는 빅데이터를 축적하는 단계;상기 빅데이터에 기초하여, 상기 병변 영역의 위치 및 형상에 따른 상기 치매 가능성 및 치매 진행 정도를 결정하는 단계;상기 치매 가능성 및 상기 치매 진행 정도에 따라, 문진 검사 및 검사실 검사를 포함한 추가 검사가 필요함을 상기 사용자 단말을 통해 실시간으로 통지하는 단계를 더 포함하고,상기 치매 종류는 알츠하이머 치매, 혈관성 치매, 루이소체 치매 및 전두엽 치매를 포함하고, 상기 치매 가능성은 백분율로 구분되며, 상기 치매 진행 정도는 초기, 중기 및 말기로 구분되는,홍채 영상을 인공지능으로 분석하는 방법.
- 제4항에 있어서,상기 합성곱 신경망은 활성화 함수(activation function) 및 초점(Focal) 손실 방법이 이용되는,홍채 영상을 인공지능으로 분석하는 방법.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/KR2018/003813 WO2019189972A1 (ko) | 2018-03-30 | 2018-03-30 | 치매를 진단을 하기 위해 홍채 영상을 인공지능으로 분석하는 방법 |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/KR2018/003813 WO2019189972A1 (ko) | 2018-03-30 | 2018-03-30 | 치매를 진단을 하기 위해 홍채 영상을 인공지능으로 분석하는 방법 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2019189972A1 true WO2019189972A1 (ko) | 2019-10-03 |
Family
ID=68059519
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/KR2018/003813 WO2019189972A1 (ko) | 2018-03-30 | 2018-03-30 | 치매를 진단을 하기 위해 홍채 영상을 인공지능으로 분석하는 방법 |
Country Status (1)
Country | Link |
---|---|
WO (1) | WO2019189972A1 (ko) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111126279A (zh) * | 2019-12-24 | 2020-05-08 | 深圳市优必选科技股份有限公司 | 一种手势互动方法及手势互动装置 |
CN111369528A (zh) * | 2020-03-03 | 2020-07-03 | 重庆理工大学 | 基于深度卷积网络的冠状动脉血管造影图像狭窄区域标示方法 |
CN113435267A (zh) * | 2021-06-09 | 2021-09-24 | 江苏第二师范学院 | 一种基于改进的卷积神经网络的在线教育学生专注力判别方法 |
CN113505655A (zh) * | 2021-06-17 | 2021-10-15 | 电子科技大学 | 面向数字孪生***的轴承故障智能诊断方法 |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20030075614A (ko) * | 2002-03-20 | 2003-09-26 | 김종성 | 홍채 영상의 자동 분석을 통한 병소 진단 방법 |
KR20040017924A (ko) * | 2002-08-22 | 2004-03-02 | 조승호 | 동공 변환율을 이용한 건강체크 시스템 및 방법과 그방법에 대한 컴퓨터 프로그램 소스를 저장한 기록매체 및이를 이용한 홍채를 이용한 신원 확인 시스템 및 방법과그 방법에 대한 컴퓨터 프로그램 소스를 저장한 기록매체 |
JP3855025B2 (ja) * | 2004-08-30 | 2006-12-06 | 富山県 | 個人認証装置 |
KR20140141917A (ko) * | 2013-06-03 | 2014-12-11 | 남궁용주 | 홍채 병소 영상의 자동분석을 위한 효율적인 동공 및 홍채 열공 데이터 획득, 홍채 열공 특징추출, 데이터 분석(학습), 그리고 판단(분류) 방법 |
KR20160097786A (ko) * | 2015-02-10 | 2016-08-18 | 삼성전자주식회사 | 사용자 단말 및 이의 제공 방법 |
KR20160118508A (ko) * | 2015-04-02 | 2016-10-12 | 삼성에스디에스 주식회사 | 사용자 인증 장치 및 방법 |
-
2018
- 2018-03-30 WO PCT/KR2018/003813 patent/WO2019189972A1/ko active Application Filing
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20030075614A (ko) * | 2002-03-20 | 2003-09-26 | 김종성 | 홍채 영상의 자동 분석을 통한 병소 진단 방법 |
KR20040017924A (ko) * | 2002-08-22 | 2004-03-02 | 조승호 | 동공 변환율을 이용한 건강체크 시스템 및 방법과 그방법에 대한 컴퓨터 프로그램 소스를 저장한 기록매체 및이를 이용한 홍채를 이용한 신원 확인 시스템 및 방법과그 방법에 대한 컴퓨터 프로그램 소스를 저장한 기록매체 |
JP3855025B2 (ja) * | 2004-08-30 | 2006-12-06 | 富山県 | 個人認証装置 |
KR20140141917A (ko) * | 2013-06-03 | 2014-12-11 | 남궁용주 | 홍채 병소 영상의 자동분석을 위한 효율적인 동공 및 홍채 열공 데이터 획득, 홍채 열공 특징추출, 데이터 분석(학습), 그리고 판단(분류) 방법 |
KR20160097786A (ko) * | 2015-02-10 | 2016-08-18 | 삼성전자주식회사 | 사용자 단말 및 이의 제공 방법 |
KR20160118508A (ko) * | 2015-04-02 | 2016-10-12 | 삼성에스디에스 주식회사 | 사용자 인증 장치 및 방법 |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111126279A (zh) * | 2019-12-24 | 2020-05-08 | 深圳市优必选科技股份有限公司 | 一种手势互动方法及手势互动装置 |
CN111126279B (zh) * | 2019-12-24 | 2024-04-16 | 深圳市优必选科技股份有限公司 | 一种手势互动方法及手势互动装置 |
CN111369528A (zh) * | 2020-03-03 | 2020-07-03 | 重庆理工大学 | 基于深度卷积网络的冠状动脉血管造影图像狭窄区域标示方法 |
CN111369528B (zh) * | 2020-03-03 | 2022-09-09 | 重庆理工大学 | 基于深度卷积网络的冠状动脉血管造影图像狭窄区域标示方法 |
CN113435267A (zh) * | 2021-06-09 | 2021-09-24 | 江苏第二师范学院 | 一种基于改进的卷积神经网络的在线教育学生专注力判别方法 |
CN113435267B (zh) * | 2021-06-09 | 2023-06-23 | 江苏第二师范学院 | 一种基于改进的卷积神经网络的在线教育学生专注力判别方法 |
CN113505655A (zh) * | 2021-06-17 | 2021-10-15 | 电子科技大学 | 面向数字孪生***的轴承故障智能诊断方法 |
CN113505655B (zh) * | 2021-06-17 | 2023-10-13 | 电子科技大学 | 面向数字孪生***的轴承故障智能诊断方法 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
KR102058884B1 (ko) | 치매를 진단을 하기 위해 홍채 영상을 인공지능으로 분석하는 방법 | |
TWI823897B (zh) | 用於診斷腸胃腫瘤的系統和方法 | |
Pires et al. | A data-driven approach to referable diabetic retinopathy detection | |
Rathod et al. | Diagnosis of skin diseases using Convolutional Neural Networks | |
WO2019189972A1 (ko) | 치매를 진단을 하기 위해 홍채 영상을 인공지능으로 분석하는 방법 | |
Kou et al. | An enhanced residual U-Net for microaneurysms and exudates segmentation in fundus images | |
WO2019132590A1 (ko) | 영상 변환 방법 및 장치 | |
WO2021137454A1 (ko) | 인공지능 기반의 사용자 의료정보 분석 방법 및 시스템 | |
WO2019098415A1 (ko) | 자궁경부암에 대한 피검체의 발병 여부를 판정하는 방법 및 이를 이용한 장치 | |
WO2021212715A1 (zh) | 精神***症分类识别方法、运行控制装置及医疗设备 | |
WO2021045507A2 (ko) | Ct 영상 기반 부위별 대뇌 피질 수축율 예측 방법 및 장치 | |
WO2019132588A1 (ko) | 영상의 특징 및 맥락에 기초한 영상 분석 장치 및 방법 | |
Islam et al. | Understanding 3D CNN behavior for Alzheimer's disease diagnosis from brain PET scan | |
Vamsi et al. | Early Detection of Hemorrhagic Stroke Using a Lightweight Deep Learning Neural Network Model. | |
WO2020179950A1 (ko) | 딥러닝 기반 뇌 질환 진행 예측 방법 및 장치 | |
Viscaino et al. | Computer-aided ear diagnosis system based on CNN-LSTM hybrid learning framework for video otoscopy examination | |
WO2020246676A1 (ko) | 자궁경부암 자동 진단 시스템 | |
WO2023182702A1 (ko) | 디지털 병리이미지의 인공지능 진단 데이터 처리 장치 및 그 방법 | |
Valenzuela et al. | A spatio-temporal hypomimic deep descriptor to discriminate parkinsonian patients | |
Hussein et al. | Myositis Detection From Muscle Ultrasound Images Using a Proposed YOLO-CSE Model | |
Xu et al. | A Tuberculosis Detection Method Using Attention and Sparse R-CNN. | |
CN117115444B (zh) | 一种多任务图像分割方法、***、计算机设备和存储介质 | |
WO2021241830A1 (ko) | 신장 기능 이상을 검출하는 심층 신경망 기반 망막 영상 분석 방법 및 장치 | |
CN113706449B (zh) | 基于病理图像的细胞分析方法、装置、设备及存储介质 | |
WO2024072067A1 (ko) | 멀티모달 딥 러닝 모델을 이용한 녹내장 환자의 미래 시야 예측 방법 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 18912811 Country of ref document: EP Kind code of ref document: A1 |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 18912811 Country of ref document: EP Kind code of ref document: A1 |
|
32PN | Ep: public notification in the ep bulletin as address of the adressee cannot be established |
Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 21/04/2021) |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 18912811 Country of ref document: EP Kind code of ref document: A1 |