Ophthalmologic image processing program and ophthalmologic image processing device
The ophthalmologic image processing program uses a machine learning algorithm to reduce phase noise in OCT images from wavelength sweeps, enhancing image quality across all regions by employing a convolutional neural network to address the challenge of phase noise suppression in OCT images.
Patent Information
- Application Number
- PCT/JP2024/016228
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-08
- Filing Date
- 2024-04-25
- Publication Date
- 2024-11-14
AI Technical Summary
OCT images using wavelength swept light sources suffer from phase noise, which is difficult to suppress with conventional methods like k-space resampling and averaging processing, especially in regions farther from zero delay, leading to noise in tomographic images.
An ophthalmologic image processing program and apparatus that employs a mathematical model trained by a machine learning algorithm to reduce phase noise from wavelength sweeps in OCT images, using a convolutional neural network to improve image quality across all regions.
Effectively suppresses phase noise components in OCT images, resulting in higher quality tomographic images regardless of the depth position, even in regions farther from zero delay.
Smart Images

Figure JP2024016228_14112024_PF_FP_ABST
Abstract
Description
Ophthalmic image processing program and ophthalmic image processing device
[0001] The present disclosure relates to an ophthalmic image processing program and an ophthalmic image processing device that process OCT data of tissue of a subject's eye.
[0002] In the field of ophthalmology, optical coherence tomography (OCT) is a device that captures tomographic images of tissues in a subject's eye. One type of OCT is swept source-OCT (SS-OCT), which uses a wavelength-swept light source that changes the emission wavelength at high speed over time as a light source (see Patent Document 1). SS-OCT makes it easier to achieve a wider imaging range in the depth direction than conventional spectral domain-OCT (SD-OCT).
[0003] JP 2015-157182 A
[0004] However, when a wavelength-swept light source is used, the wavelength is changed very quickly to emit light, making control complex. The phase of the wavelength sweep at each sampling timing of the interference signal may oscillate irregularly (fluctuations), resulting in failure of resampling of the interference signal into wavenumber space, resulting in noise appearing in the tomographic image. This noise is more likely to be a problem in areas farther away from zero delay. This noise is difficult to suppress using computational processing to generate OCT data from spectral interference signals (e.g., k-space resampling methods such as the mapping arm method and the k-clock method).
[0005] Furthermore, in OCT, averaging of multiple images is widely used as a method for reducing noise, but phase noise in wavelength sweeps originating from the light source is difficult to remove using averaging.
[0006] The present invention has been made in view of the above circumstances, and has as its technical object to provide an ophthalmic image processing program and an ophthalmic image processing device that can acquire tomographic images with good image quality in each region.
[0007] An ophthalmic image processing program according to a first aspect of the present disclosure is an ophthalmic image processing program executed by an ophthalmic image processing device that processes ophthalmic images, which are images of the tissues of a test eye. When the ophthalmic image processing program is executed by a control unit of the ophthalmic image processing device, the ophthalmic image processing device is caused to execute a base image acquisition step of acquiring, as a base image, a tomographic image taken by an SS-OCT device having a wavelength swept light source as an OCT light source, and a target image acquisition step of acquiring a target image by inputting the base image into a mathematical model trained by a machine learning algorithm so as to output an image in which the phase noise of the wavelength sweep originating from the OCT light source in the input image is at least reduced.
[0008] An ophthalmological image processing apparatus according to a second aspect of the present disclosure executes the ophthalmological image processing program according to the first aspect.
[0009] According to the present disclosure, a tomographic image with good image quality is acquired in each region.
[0010] Fig. 1 is a diagram showing a schematic configuration of an OCT system according to an embodiment; Fig. 2 is a diagram showing an OCT optical system according to an embodiment; Fig. 3 is a flowchart showing an example of a calculation process for generating OCT data from a spectral interference signal; Fig. 4 is a flowchart showing a process for reducing phase noise in wavelength sweeping; Fig. 5 is a diagram for explaining a mathematical model; Fig. 6 is a diagram for explaining this reduction process; Fig. 7 is a diagram for explaining a modified example; Fig. 8 is a diagram for explaining a second modified example.
[0011] An embodiment of the present disclosure will be described. Items grouped in < > below can be used independently or in conjunction with each other. An ophthalmic image processing program according to the embodiment is executed by an ophthalmic image processing device that processes ophthalmic images, which are images of tissues of a subject's eye.
[0012] The ophthalmologic image processing program is executed by a control unit (also referred to as a "processor") of the ophthalmologic image processing device, thereby executing at least a base image obtaining step and a target image obtaining step.
[0013] In the base image acquisition step, a tomographic image captured by an SS-OCT apparatus having a wavelength swept light source as an OCT light source is acquired as a base image.
[0014] In the target image acquisition step, the target image is acquired by inputting the base image into a mathematical model trained by a machine learning algorithm so as to output an image in which at least the phase noise of the wavelength sweep originating from the OCT light source in the input image has been reduced. Phase noise is noise that occurs when the phase of the wavelength sweep at each sampling timing of the interference signal irregularly oscillates (fluctuations), resulting in failure of resampling the interference signal into wavenumber space. Phase noise is difficult to suppress using the mapping arm method, the k-clock method, and other methods, and is also difficult to suppress using conventional image processing such as averaging processing using multiple images. In contrast, in this embodiment, by utilizing a mathematical model trained by a machine learning algorithm so as to output an image in which at least the phase noise of the wavelength sweep originating from the OCT light source in the input image has been reduced, it is possible to appropriately suppress the phase noise component. In other words, by inputting the base image into the mathematical model, a target image in which the phase noise component in the base image has been appropriately reduced can be acquired.
[0015] The mathematical model refers to, for example, a data structure for predicting the relationship between input data and output data. The mathematical model is constructed by training using a training data set. In this embodiment, the training data set may be a set of input training data and output training data. For example, correlation data (e.g., weights) between each input and output is updated through training. In this embodiment, a program and data for realizing the constructed mathematical model are incorporated into the fundus image processing device.
[0016] Commonly known machine learning algorithms include, for example, neural networks, random forests, boosting, and support vector machines (SVMs).
[0017] In this embodiment, a multi-layer neural network is used as the machine learning algorithm. The neural network includes an input layer for inputting data, an output layer for generating data of the analysis result to be predicted, and one or more hidden layers between the input layer and the output layer. A plurality of nodes (also referred to as units) are arranged in each layer. In particular, in this embodiment, a convolutional neural network (CNN), which is a type of multi-layer neural network, is used. However, other machine learning algorithms may also be used. For example, a generative adversarial network (GAN) that uses two competing neural networks may be adopted as the machine learning algorithm.
[0018] The mathematical model used in the target image acquisition step of this embodiment may be trained using data of a tomographic image captured by placing a specified part of the test eye in a first region of the depth-wise imaging range in the tomographic image as output training data, and data of a noise-added image in which a phase noise component corresponding to a second region that is farther from zero delay than the first region has been added to the tomographic image captured by placing a specified part of the test eye in the first region as input training data.
[0019] This allows the phase noise component in the second region to be appropriately reduced for a tomographic image captured by placing a predetermined portion in the second region farther from the zero delay than the first region, thereby making it possible to obtain tomographic images of good image quality in each region.
[0020] The above-described mathematical model is trained using a training data set consisting of input training images having a phase noise component and output training images in which the phase noise component has been reduced, but this is not necessarily limited to this. For example, the mathematical model of this embodiment may be trained based on a technique called Noise2Noise to remove noise. In this case, the mathematical model that outputs images in which the phase noise of the wavelength sweep from the OCT light source has been reduced may be trained using images with a phase noise component without using images in which the phase noise component has been reduced as output training images. That is, in this embodiment, the mathematical model trained to reduce the phase noise component is trained based on images with a phase noise component.
[0021] In this embodiment, the mathematical model may be trained using data on tomographic images captured by multiple OCT devices and noise-added images corresponding to the tomographic images. Since the phase noise characteristics of the wavelength sweep originating from the OCT light source differ for each device (each light source), training using data on tomographic images captured by multiple OCT devices and noise-added images corresponding to the tomographic images makes it easier to effectively reduce noise without necessarily training the mathematical model based on images from the device actually used.
[0022] The mathematical model may be trained using a plurality of noise-added images, each having a phase noise component corresponding to a different depth position, as input training data, which makes it easier to obtain a target image in which the phase noise of the wavelength sweep originating from the light source is well suppressed, regardless of the depth position of a predetermined portion in the input image.
[0023] EXAMPLES An OCT system (optical coherence tomography system) shown in FIGS. 1 and 2 will be described below as an example.
[0024] 1, the OCT system according to the embodiment includes at least an optical unit 10 and a control unit 50, which corresponds to the computer of the embodiment. In the embodiment, the optical unit 10 and the control unit 50 are integrated as an OCT device. The OCT system (OCT device) according to the embodiment has a basic configuration of swept-wavelength OCT (SS-OCT).
[0025] The optical unit 10 includes a light-guiding optical system 150. Furthermore, the optical unit 10 in this embodiment includes a fundus observation optical system 200 and an anterior eye observation optical system 300.
[0026] The optical unit 10 can be moved three-dimensionally by an XYZ moving unit 15. In this embodiment, the XYZ moving unit 15 is driven and controlled by an arithmetic controller 70. In this embodiment, the XYZ moving unit 15 moves the optical unit 10 three-dimensionally, thereby adjusting the three-dimensional position of the optical unit 10 with respect to the subject's eye E. This aligns the three-dimensional position of the optical unit 10 with respect to the subject's eye E. In addition, the subject's face is supported by a face support unit 17. The support position of the face by the face support unit 17 can be moved in the up and down directions.
[0027] The control unit 50 is an ophthalmologic image processing device in this embodiment, and includes at least an arithmetic controller (processor) 70 that controls the entire OCT system. The arithmetic controller 70 is configured with, for example, a CPU and a memory. As an example, in this embodiment, the arithmetic controller 70 also serves as an image processor in the OCT system.
[0028] The OCT system may also be provided with a storage unit (memory) 72, an input interface (operation unit) 75, a monitor 80, etc. Each unit is connected to the arithmetic and control unit 70.
[0029] Various programs, initial values, and the like for controlling the operation of the OCT apparatus may be stored in the memory 72. For example, a hard disk drive, a flash ROM, or a USB memory that is detachably attached to the OCT apparatus may be used as the memory 72. Furthermore, the memory 72 may store various information related to imaging in addition to OCT images generated from OCT data. The monitor 80 may display the OCT data (OCT images).
[0030] <OCT optical system> Next, the OCT optical system 100 in this embodiment will be described with reference to Fig. 2. The OCT optical system 100 guides measurement light to the subject's eye E using a light-guiding optical system 150. The OCT optical system 100 guides reference light to a reference optical system 110. The OCT optical system 100 causes a detector (light-receiving element) 120 to receive spectral interference signal light obtained by interference between the measurement light reflected by the subject's eye E and the reference light.
[0031] In this embodiment, the SS-OCT system is used for the OCT optical system 100. In this case, the OCT optical system 100 has a wavelength swept light source as the OCT light source 102. The OCT optical system 100 also has a point detector as the detector 120.
[0032] The wavelength of the swept light source is swept over time. The OCT light source 102 may be a VCSEL-type swept light source. The VCSEL-type swept light source includes a VCSEL that performs laser oscillation and a MEMS that realizes high-speed scanning.
[0033] In this embodiment, the detector 120 is a balanced detector that performs balanced detection using multiple (e.g., two) detectors. The arithmetic and control unit 70 samples the interference signal between the reference light and the returned light of the measurement light in response to changes in the wavelength emitted by the wavelength swept light source, and obtains OCT data of the subject's eye based on the interference signal at each wavelength obtained by sampling.
[0034] The coupler (splitter) 104 is used as a first optical splitter and splits the light emitted from the light source 102 into a measurement optical path and a reference optical path. The coupler 104, for example, guides the light from the light source 102 to an optical fiber 152 on the measurement optical path side and to a reference optical system 110 on the reference optical path side.
[0035] <Light Guiding Optical System> The light guiding optical system 150 is provided to guide the measurement light to the eye E. The light guiding optical system 150 may be provided, for example, with an optical fiber 152, a collimator lens 153, a focusing lens 155, an optical scanner 156, and an objective lens system 158 (the objective optical system in this embodiment) in that order. In this case, the measurement light is emitted from the exit end of the optical fiber 152 and converted into a parallel beam by the collimator lens 153. The measurement light then travels toward the optical scanner 156 via the focusing lens 155. The focusing lens 155 is displaceable along the optical axis by a drive unit (not shown) and is used to adjust the focusing state. The light that has passed through the optical scanner 156 is irradiated onto the eye E via the objective lens system 158. In this embodiment, a pivot point P is formed at a position conjugate with the optical scanner 156 with respect to the objective lens system 158 (provided in the device main body).
[0036] The optical scanner 156 may scan the measurement light in the X and Y directions (transverse directions) on the tissue of the subject's eye E. In this embodiment, the optical scanner 156 is, for example, two galvanometer mirrors, the reflection angle of which is arbitrarily adjusted by a drive mechanism. The light beam emitted from the light source 102 has its reflection (propagation) direction changed, and is scanned in an arbitrary direction on the tissue of the subject's eye E. As the optical scanner 156, for example, a reflection mirror (galvanometer mirror, polygon mirror, resonant scanner), or an acousto-optical element (AOM) that changes the propagation (deflection) direction of light may be used.
[0037] The scattered light (reflected light) from the eye E due to the measurement light travels back along the path taken when the light was projected, enters the optical fiber 152, and reaches the coupler 104. The coupler 104 guides the light from the optical fiber 152 to an optical path leading to the detector 120.
[0038] <Reference Optical System> The reference optical system 110 generates reference light. The reference light is combined with the measurement light reflected from the subject's eye E. The reference light that has passed through the reference optical system 110 is combined with light from the measurement optical path by a coupler 148 and interferes with the light. The reference optical system 110 may be a Michelson type or a Mach-Zehnder type.
[0039] 2 is formed by a transmission optical system, for example. In this case, the reference optical system 110 transmits the light from the coupler 104 without returning it, thereby guiding it to the detector 120. However, the reference optical system 110 may be formed by a reflection optical system, for example, and may guide the light from the coupler 104 to the detector 120 by reflecting it using the reflection optical system. In this embodiment, an optical path length difference adjustment unit 145 and a polarization adjustment unit 147 are arranged on the optical path from the coupler 104 to the detector 120.
[0040] The optical path length difference adjustment unit 145 is used to adjust the optical path length difference between the measurement light and the reference light. When acquiring OCT data, it is necessary to adjust the optical path length difference between the measurement light and the reference light in advance, at least according to the depth position of the imaging target (a region of the test eye E). In this embodiment, a mirror 145a having two orthogonal surfaces is provided on the reference light path. The optical path length of the reference light path can be increased or decreased by moving this mirror 145a in the direction of the arrow by an actuator 145b. Of course, the configuration for adjusting the optical path length difference between the measurement light and the reference light is not limited to this. For example, in the light-guiding optical system 150, the collimator lens 153 and the coupler may be moved integrally to adjust the optical path length of the measurement light, thereby adjusting the optical path length difference between the measurement light and the reference light.
[0041] In this embodiment, the polarization adjuster 147 adjusts the polarization of the reference light. The polarization adjuster may be disposed on the measurement light path.
[0042] <OCT Calculation> Next, an example of calculation processing for generating OCT data from a spectral interference signal will be described with reference to Fig. 3. First, the calculation controller 70 processes (performs Fourier analysis on) the spectral interference signal detected by the detector 120 to obtain OCT data (OCT image) of the subject's eye (S101).
[0043] In this embodiment, the detector 120 detects the spectral interference signal as a beat signal whose intensity changes over time based on the wavelength sweep. For example, the calculation controller 70 performs a mapping process (S104) on the spectral interference signal for each A-scan after performing DC-subtraction (a process for removing direct current components) (S102) and zero padding (S103).
[0044] In the mapping process (S104), the arithmetic and control unit 70 converts the spectral interference signal into a function I(k) equally spaced with respect to the wavenumber k (=2π / λ). In other words, the spectral interference signal is resampled in wavenumber space. For a more specific method, see, for example, Japanese Patent Application Laid-Open No. 2015-68775 filed by the present applicant. Next, the arithmetic and control unit 70 performs a dispersion correction process (S105) on the resampled spectral interference signal I(k). The result is subjected to a Fourier transform (FFT) (S106), and mirror images are removed (S108), thereby obtaining A-scan OCT data. Multiple A-scan OCT data thus obtained are acquired in the transverse direction, and B-scan OCT data is constructed by consecutively arranging these data (S109).
[0045] <Processing for reducing wavelength sweep phase noise> Next, in this embodiment, processing for reducing wavelength sweep phase noise originating from the light source is performed on the generated B-scan OCT data. The phase noise σ of the wavelength sweep originating from the light source occurs due to irregular oscillation of the wavelength sweep phase, which causes failure of resampling of the interference signal into wavenumber space. In OCT, averaging processing using multiple images is widely used as a method for reducing noise, but it is difficult to remove the wavelength sweep phase noise σ originating from the light source using averaging processing.
[0046] Here is the equation for the interference signal after resampling: For ease of explanation, an ideal multi-layer SS-OCT A-scan signal can be expressed in integral form as follows:
[0047]
[0048] z s is the sample depth l s is the reflectivity from that depth, t is time, φ zs is an arbitrary phase. 0 t indicates the case where the wave number of the wavelength swept light source is linear with respect to time. However, since the wave number of the wavelength swept light source is usually not linear with respect to time, k is expressed as k = k(t) as a function of t. Furthermore, when noise σ in the phase of the wavelength sweep is taken into consideration,
[0049]
[0050] It can be seen that the effect of phase noise increases as the sample depth increases, that is, in areas farther away from zero delay. Therefore, the noise in the image due to phase noise becomes more noticeable as the imaging range in the depth direction becomes wider. In other words, it appears as image fuzziness and a decrease in resolution in the depth direction.
[0051] In contrast, the reduction process of this embodiment uses a mathematical model trained by a machine learning algorithm to obtain OCT data with reduced phase noise. That is, in this reduction process, as shown in Fig. 4, OCT data is input to the mathematical model as a base image (S201). As a result, OCT data with higher image quality than the input OCT data is obtained as a target image (S202).
[0052] <Machine Learning Algorithm and Mathematical Model> A mathematical model refers to, for example, a data structure for predicting the relationship between input data and output data. The mathematical model is constructed by training using a training data set. In this embodiment, the training data set is a set of input training data and output training data. For example, correlation data (e.g., weights) between each input and output is updated through training. In this embodiment, a program and data for realizing the constructed mathematical model are incorporated into the control unit 50 and the arithmetic controller 70.
[0053] Commonly known machine learning algorithms include, for example, neural networks, random forests, boosting, and support vector machines (SVMs).
[0054] In this embodiment, a multi-layer neural network is used as the machine learning algorithm. The neural network includes an input layer for inputting data, an output layer for generating data of the analysis result to be predicted, and one or more hidden layers between the input layer and the output layer. A plurality of nodes (also referred to as units) are arranged in each layer. In particular, in this embodiment, a convolutional neural network (CNN), which is a type of multi-layer neural network, is used. However, other machine learning algorithms may also be used. For example, a generative adversarial network (GAN) that uses two competing neural networks may be adopted as the machine learning algorithm.
[0055] In this embodiment, the mathematical model is pre-trained so that an image in which at least the phase noise of the wavelength sweep caused by the light source in the input image is reduced is output. Specifically, as shown in FIG. 5 , high-quality OCT data 402 with reduced phase noise is used as output training data, and noise-added data 403, which is obtained by adding phase noise to the high-quality OCT data, is used as input training data. For example, the high-quality OCT data 402 used as output training data may be generated by extracting a first region from OCT data 401 captured with a predetermined portion of the subject's eye positioned in the first region near zero delay within the imaging range in the depth direction. The noise-added data 403 may be generated by adding pseudo phase noise corresponding to the depth position of a second region, which is farther from zero delay than the first region, to the extracted OCT data 401.
[0056] By inputting an OCT image of a specified area taken in the second region into a mathematical model trained based on multiple pairs of such OCT data 402 and noise-added data for the OCT data 403, an OCT image can be generated in which the phase noise of the wavelength sweep originating from the light source is reduced.
[0057] <Method for Obtaining Pseudo Phase Noise> Here, the pseudo phase noise added to high-quality OCT data can be obtained, for example, as follows. For example, it can be derived based on the interference signal of a gold mirror. First, the interference signal of the gold mirror can be written as follows.
[0058]
[0059] This is subjected to Hilbert transformation [H], and the instantaneous phase is unwrapped [P] to obtain the phase slope φ IAu (t) is obtained.
[0060]
[0061] k(t) is the average (ensemble average) of the phase slopes of multiple interference signals〈φ IAU (t)〉. Also, 〈σ〉 = 0. Therefore, the noise is
[0062]
[0063] Using this, the phase noise function can be expressed as A·e Bσt By multiplying the phase noise function by the OCT analysis signal (Hilbert transformed exponential representation) of the first region depicting a predetermined part of the test eye, noise-added data that reproduces pseudo-resampling errors at any depth can be obtained. However, A and B appearing in the phase noise function are functions of the frequency of the acquired gold mirror signal (depth of the gold mirror) and must be adjusted appropriately.
[0064] In actual imaging, the depth position at which a predetermined region is depicted is not constant for each imaging. The depth position at which a predetermined region is depicted may vary for each imaging due to alignment errors, individual differences in the subject's eye, and the like. Therefore, in this embodiment, the pseudo phase noise added to each noise-added data of the training data set may correspond to different depth positions. A mathematical model trained with such a training data set can effectively suppress the phase noise of the wavelength sweep originating from the light source, regardless of the depth position at which the predetermined region is depicted.
[0065] Furthermore, the imaging range in the depth direction corresponding to the OCT data of the base image may be a wide range including the first region and the second region. Therefore, the reduction process of this embodiment may process a partial image, which is a partial depth region in which a predetermined part is depicted in the captured OCT data, as the base image.
[0066] In this embodiment, the phase noise of the wavelength sweep is reduced as described above, so that a tomographic image with good image quality can be obtained even when a specific tissue must be positioned away from the zero delay in the OCT data.
[0067] <Modifications> Although the present disclosure has been described above based on the embodiments, the present disclosure is not limited to the above embodiments and various modifications are possible.
[0068] For example, as described above, the depth position at which a predetermined part is depicted may differ for each imaging. In response to this, as shown in FIG. 7, a mathematical model may be prepared for each depth position at which a predetermined part is depicted. In FIG. 7, 1 , Z 2 , Z 3 Three mathematical models corresponding to the depth positions are prepared. A mathematical model corresponding to a certain depth position is trained using the above-described high-quality OCT data 402 as an output training image, and is also trained using noise-added data 403, in which phase noise at that depth position is added to the high-quality OCT data 402, as an input training image. OCT data with reduced phase noise may be obtained by inputting the base image to one of the multiple (three in FIG. 7 ) mathematical models that corresponds to the depth position at which a predetermined part is depicted in the base image.
[0069] Furthermore, for example, as shown in FIG. 8 , OCT data with reduced phase noise can be obtained using a single mathematical model, regardless of the depth position at which a predetermined region is depicted in the base image. That is, the base image may be pre-processed to make the degree of phase noise uniform for each image, and then the base image may be input into a single mathematical model to obtain OCT data with reduced phase noise. In this case, the mathematical model is trained using, as input training images, OCT data to which a phase noise component corresponding to a predetermined depth position has been added to high-quality OCT data. For example, in the example of FIG. 8 , the mathematical model is trained using the above-mentioned high-quality OCT data 402 as output training images, and a depth position Z is added to the high-quality OCT data 402 as input training images. 1 , Z 2 , Z 3 The deepest Z 3 The training is performed using OCT data to which a phase noise component corresponding to the case where the depth position at which a predetermined part is depicted in the base image is Z 3 If the depth is shallower than Z (more precisely, if it is close to zero delay), a phase noise component is added according to the depth position by preprocessing. 1 and Z2 In each case, the depth position at which the predetermined part is depicted is Z 3 The phase noise component is added so that it is the same as in the case of A·e Bσt 8, A and B are both functions that depend on the depth Z.
[0070] In this disclosure, the term "processor" refers to one or more hardware processors configured to execute program code (i.e., one or more instructions of a program) contained in a program. In other words, a "processor" is a hardware device capable of executing one or more programmed processes. For example, a "processor" may be a general-purpose or special-purpose processor, such as, but not limited to, a CPU, a microprocessor, a GPU, and a data flow processor (DFP).
[0071] In this disclosure, the term "memory" refers to one or more hardware memories that are non-transitory tangible recording media configured to store computer program code and / or data accessible to a processor. The "memory" may be implemented using memory technologies such as SRAM, SDRAM, non-volatile / flash-type memory, or other types of memory. Computer program code constituting a program may be stored in the memory and executed by the processor to cause the ophthalmic image processing device to perform various functions.
Claims
1. An ophthalmic image processing program executed by an ophthalmic image processing device that processes ophthalmic images, which are images of tissues of a subject's eye, wherein the ophthalmic image processing program is executed by a control unit of the ophthalmic image processing device to cause the ophthalmic image processing device to execute the following steps: a base image acquisition step of acquiring, as a base image, a tomographic image taken by an SS-OCT device having a wavelength swept light source as an OCT light source; and a target image acquisition step of acquiring a target image by inputting the base image into a mathematical model trained by a machine learning algorithm so as to output an image in which the phase noise of the wavelength sweep originating from the OCT light source in the input image is at least reduced.
2. The ophthalmologic image processing program of claim 1, wherein the mathematical model is trained using data of a tomographic image captured by placing a specified part of the test eye in a first region of the depth-wise shooting range of the tomographic image as output training data, and data of a noise-added image in which the phase noise component corresponding to a second region farther from zero delay than the first region is added to a tomographic image captured by placing a specified part of the test eye in the first region as input training data.
3. The ophthalmologic image processing program according to claim 2, wherein the mathematical model is trained using data of tomographic images taken by a plurality of OCT devices and noise-added images corresponding to the tomographic images.
4. An ophthalmologic image processing program as described in claim 2 or 3, characterized in that the mathematical model is trained using a plurality of noise-added images to which phase noise components corresponding to different depth positions have been added as input training data.
5. An ophthalmological image processing device that executes the ophthalmological image processing program according to any one of claims 1 to 4.
Citation Information
Patent Citations
Oct apparatus and imaging control program
JP2022185838A
Ophthalmologic device and IOL power determination program
WO2018021561A1
Medical image processing device, medical image processing method, and program
WO2020183799A1
Information processing device and program
WO2022234828A1
