Adaptive near amblyopia therapy device

Abstract

The application discloses an adaptive near amblyopia therapeutic instrument, which comprises a pair of optical lens groups, a shielding assembly, a laser assembly, an optical perception assembly and a control module connected with the shielding assembly, the laser assembly and the optical perception assembly, which are symmetrically arranged in sequence, wherein the control module calculates the irradiation area according to the image collected by the optical perception assembly and outputs control instructions to the shielding assembly and the laser assembly to control the light intensity, frequency, irradiation angle and irradiation area. The application calculates the light intensity, light frequency and irradiation area of the required laser irradiation through the image recognition technology, so as to adaptively adjust the light intensity, frequency, irradiation angle and irradiation area while excluding the macula region.

Description

Technical Field

[0001] This invention relates to a technology in the field of laser ophthalmic equipment, specifically an adaptive myopia and amblyopia treatment device that can automatically adjust the light intensity, frequency, irradiation angle, and irradiation area. Background Technology

[0002] Current techniques for treating myopia and amblyopia use low-intensity lasers to irradiate the eye, improving blood circulation in the fundus, increasing retinal blood flow, and restoring choroidal thickness. However, existing technologies cannot adaptively and dynamically adjust the output power; too low an intensity will not produce any effect, while excessive irradiation can easily lead to retinal damage. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies that directly apply low-intensity laser irradiation to myopia and amblyopia by proposing an adaptive myopia and amblyopia treatment device. Through image recognition technology, it calculates the required laser irradiation intensity, frequency, and irradiation area, thereby adaptively adjusting the irradiation intensity, frequency, irradiation angle, and irradiation area while excluding the macular region.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to an adaptive myopia and amblyopia treatment device, comprising: a pair of optical lens groups, a blocking component, a laser component, an optical sensing component arranged symmetrically in sequence, and a control module connected to the blocking component, the laser component, and the optical sensing component respectively, wherein: the control module calculates the irradiation area based on the image acquired by the optical sensing component and outputs control commands to the blocking component and the laser component respectively to control the light intensity, frequency, irradiation angle, and irradiation area.

[0006] The images acquired by the optical sensing component include pupil images and retinal images.

[0007] The optical sensing components include: an optical camera for capturing images of the human eye and a fundus camera for capturing images of the retina, wherein: the optical camera outputs the captured images of the human eye to the control module for locating the pupil position, thereby adjusting the orientation of the laser generator; the fundus camera outputs the retinal images to the control module for identifying and segmenting the position of the macula, and combining the human eye images to confirm and optimize the light intensity and light frequency, as well as the area that the laser mask needs to block.

[0008] The laser assembly includes two laser generators, each used to stimulate and treat the eyes. The laser emitters are connected to a control module and generate stable single-frequency parallel lasers with adjustable intensity and frequency according to instructions.

[0009] The shielding component is a liquid crystal mask or a mechanical mask connected to the control module. Under the control of the control module, the area through which the laser passes is adjusted to achieve the effect of shielding the macula region.

[0010] The control module includes a pupil image segmentation unit, a retinal image segmentation unit, a mask control unit, and a laser control unit. Specifically: the pupil image segmentation unit obtains the pupil position in the human eye image using a pupil segmentation AI model; the retinal image segmentation unit obtains the positions of blood vessels, the optic disc, and the macula in the retinal image using a comprehensive retinal image segmentation AI model; the mask control unit controls the light-blocking mask of the occlusion component based on the macula position information obtained by the retinal image segmentation unit, ensuring that laser irradiation avoids the macula; the laser control unit calculates the corresponding irradiation angle based on the pupil position information obtained by the pupil image segmentation unit and the positions of blood vessels, the optic disc, and the macula in the retina obtained by the retinal image segmentation unit, generates control commands, and outputs them to the occlusion component; it also calculates the corresponding laser intensity and frequency based on the retinal information, generates control commands, and outputs them to the laser component, thereby adaptively adjusting the laser irradiation angle, intensity, and frequency. Technical effect

[0011] This invention uses an AI model to select areas for low-intensity laser irradiation treatment, thereby effectively avoiding irradiation of the macular region and preventing macular degeneration caused by laser irradiation, making the treatment safer; it can also intelligently select the light intensity and light frequency more suitable for the user, making the treatment more effective. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the structure of the present invention;

[0013] Figure 2 This is a schematic diagram of the device's workflow;

[0014] Figure 3 This is a schematic diagram of a laser generator;

[0015] Figure 4 This is a schematic diagram of the structure of the blocking component (liquid crystal);

[0016] Figure 5 This is a schematic diagram of the shielding component (mechanical) structure. Figure 6 and Figure 7 This is a schematic diagram of the segmentation results for an example. Detailed Implementation

[0017] like Figure 1As shown in this embodiment, an adaptive myopia and amblyopia treatment device includes: a housing and a laser component, an optical lens group, an optical sensing component, a blocking component, and a control module disposed therein. The housing has two cavities; the innermost cavity houses the sensing component for capturing images of the human eye and retina. In front of the sensing component is a retractable laser component for generating the laser. The blocking component for blocking the laser is stored in the middle of the cavity. The remaining area of ​​the housing houses the control module, which can be used to calculate and control the light intensity, frequency, irradiation angle, and irradiation area.

[0018] The images acquired by the optical sensing component include pupil images and retinal images.

[0019] The control module includes: a pupil image segmentation unit, a retinal image segmentation unit, a mask control unit, and a laser control unit. The pupil image segmentation unit has a built-in pupil segmentation AI model, the retinal image segmentation unit has a built-in retinal image comprehensive segmentation AI model, the mask control unit has a built-in control circuit connected to the occlusion component, and the laser control unit has a built-in regression model and a control circuit connected to the laser component.

[0020] The pupil and retinal images are preferably resized to 256×256 using the torchvision.transforms.Resize method in Python.

[0021] The pupil segmentation AI model includes an encoder, a feature processor, and a decoder, wherein: the encoder extracts five 16×16×1024 feature vectors from the input pupil image; the feature processor generates a soft attention vector based on the five feature vectors and fuses them to obtain a 16×16×1024 feature vector; and the decoder generates a pupil segmentation mask based on the feature vector.

[0022] The encoder contains five convolutional modules and a max pooling layer. Each convolutional module consists of a convolutional layer, a batch processing layer, a ReLU activation function, a convolutional layer, a batch processing layer, and a ReLU activation function. The max pooling layer is used to reduce the length and width of the feature vector, ultimately obtaining a 16×16×1024 feature vector.

[0023] The feature processor employs a Spatial Dilated Pooling Pyramid (ASPP) module with attention engineering. This module performs the following convolutions on the input vector: 1. Global average pooling + upsampling to the original size; 2. Convolution with a 1×1 kernel, stride of 1, no padding, and no dilation; 3. Convolution with a 3×3 kernel, stride of 1, padding of 6, and dilation of 6; 4. Convolution with a 3×3 kernel, stride of 1, padding of 12, and dilation of 12; 5. Convolution with a 3×3 kernel, stride of 1, padding of 18, and dilation of 18. The five vectors obtained above are then concatenated together at the channel level and passed through a convolutional layer with a 3×3 kernel, padding of 1, stride of 1, 5120 input channels, and 1024 output channels to obtain a 16×16×1024 vector as a soft attention vector. This vector is then multiplied by the original feature vector to obtain a 16×16×1024 feature vector.

[0024] The decoder includes four upconvolutional modules, four residual convolutional layers, and one convolutional layer with a 1×1 kernel. Each upconvolutional module consists of an upsampling layer, a convolutional layer, a batch processing layer, and a ReLU activation function. Each upsampling layer concatenates the vector of the corresponding layer in the encoder with the vector obtained by the upconvolutional layer at the channel number level. Then, the residual convolutional layer restores the channel number. After the above steps are repeated four times, a 256×256×64 vector is obtained. This vector is passed through the convolutional layer with a 1×1 kernel to obtain a 256×256×1 mask image. A threshold of 0.5 is set for this image. Values ​​greater than 0.5 become 1, and values ​​less than or equal to 0.5 become 0, finally obtaining the pupil segmentation mask image of the human eye image.

[0025] The pupil segmentation AI model described above is trained as follows: 3000 human eye images with pupil masks are used as a sample set, randomly divided into training and test sets at a 9:1 ratio. A dataset is then generated using the `Dataset` class in the PyTorch package of Python, and the `Dataloader` class is used to make it usable for model optimization training. The training set data is then processed in batches through the model to obtain results. Accuracy and cross-entropy are calculated using the results and the labels of the training set, with cross-entropy used as the loss function in the optimization process. The model is further optimized using the `optim.Adam` class of PyTorch's built-in Adam algorithm, ultimately obtaining a high-precision pupil segmentation model. The segmentation results are shown below. Figure 6 As shown, the first column is the labeled pupil mask, the second column is the predicted pupil mask, and the third column is the corresponding human eye image.

[0026] In the aforementioned sample set, the pupil outlines were labeled using the labelme tool, and then processed using the cv2 package in Python to obtain the pupil mask for each image.

[0027] The retinal image comprehensive segmentation AI model includes an encoder, a feature processor, and a decoder, wherein: the encoder extracts five 16×16×1024 feature vectors from the input retinal image; the feature processor generates a soft attention vector based on the five feature vectors and fuses them to obtain a 16×16×1024 feature vector; and the decoder generates a retinal segmentation mask map based on the feature vector.

[0028] The encoder contains five convolutional modules and a max pooling layer. Each convolutional module consists of a convolutional layer, a batch processing layer, a ReLU activation function, a convolutional layer, a batch processing layer, and a ReLU activation function. The encoder adds a max pooling layer after the first four convolutional modules to reduce the length and width of the feature vector, ultimately obtaining a 16×16×1024 feature vector.

[0029] The feature processor employs a Spatial Dilated Pooling Pyramid (ASPP) module with attention engineering. This module performs the following operations on the input vector: 1. Global average pooling + upsampling to the original size; 2. Convolution with a 1×1 kernel, stride of 1, no padding and no dilation; 3. Convolution with a 3×3 kernel, stride of 1, padding of 6, and dilation of 6; 4. Convolution with a 3×3 kernel, stride of 1, padding of 12, and dilation of 12; 5. Convolution with a 3×3 kernel, stride of 1, padding of 18, and dilation of 18. The five vectors obtained above are then concatenated together at the channel level and passed through a convolutional layer with a 3×3 kernel, padding of 1, stride of 1, 5120 input channels, and 1024 output channels to obtain a 16×16×1024 vector as a soft attention vector. This vector is then multiplied by the original feature vector to obtain a 16×16×1024 feature vector.

[0030] The decoder includes four upconvolutional modules, four residual convolutional layers, and one convolutional layer with a 1×1 kernel. Each upconvolutional module consists of an upsampling layer, a convolutional layer, a batch processing layer, and a ReLU activation function. Each upsampling layer concatenates the vector of the corresponding layer in the encoder with the vector obtained by the upconvolutional layer at the channel number level. Then, the residual convolutional layer restores the channel number. After the above steps are repeated four times, a 256×256×64 vector is obtained. This vector is passed through the convolutional layer with a 1×1 kernel to obtain a 256×256×1 mask image. Three thresholds of 0.5, 1.5, and 2.5 are set for this image. Values ​​less than or equal to 0.5 become 0, values ​​greater than 0.5 and less than or equal to 1.5 become 1, values ​​greater than 1.5 and less than or equal to 2.5 become 2, and values ​​greater than 2.5 become 3. Here, 0 represents no segmentation result, 1 represents the pixel where the blood vessel is located, 2 represents the pixel where the optic disc is located, and 3 represents the pixel where the macula is located. Finally, a pupil segmentation mask image is obtained.

[0031] The aforementioned retinal image comprehensive segmentation AI model is trained as follows: 3000 human eye images with pupil masks are used as a sample set, randomly divided into training and test sets at a 9:1 ratio. A dataset is then generated using the `Dataset` class in the PyTorch package, and the `Dataloader` class is used to optimize the model. The training set data is then processed in batches through the model to obtain results. Accuracy and L2 loss are calculated using the results and the labels of the training set, with L2 loss used as the loss function in the optimization process. Finally, the `optim.Adam` class of the built-in Adam algorithm in PyTorch is used to optimize the model, resulting in a high-precision retinal comprehensive segmentation model. The segmentation results are shown below. Figure 7 As shown, the first column is the original retinal image, the second column is the predicted vascular mask, the third column is the optic disc mask, and the fourth column is the macular mask.

[0032] In the aforementioned sample set, the labelme tool was used to annotate the contours of blood vessels, the optic disc, and the macular region. Then, the cv2 package in Python was used to process these contours to obtain masks for the blood vessels, optic disc, and macular region in each image. This involved 3000 images.

[0033] The regression model includes three convolutional layers, two average pooling layers, and one fully connected layer. The convolutional layers extract a 256×256×64 feature vector from the input 256×256×1 retinal segmentation mask. The feature vector is then transformed into a 64×64×64 feature vector by the average pooling layer, then into a 64×64×256 feature vector by the convolutional layer, downsampled to a 1×256 feature vector by the global average pooling layer, and finally obtained two values, light intensity and light frequency, by the fully connected layer.

[0034] The regression model is trained as follows: 3000 images with light intensity and frequency are used as a sample set, randomly divided into training and test sets at a ratio of 9:1. A dataset is generated using the Dataset class in the PyTorch package of Python, and the Dataloader class is used to make it usable for model optimization training. The training set data is then processed in batches through the model to obtain results, and the L1 loss is calculated using the results and training set labels. The L1 loss is used as the loss function in the optimization process, and the optim.Adam class of the built-in Adam algorithm in PyTorch is used to optimize the model, finally obtaining a high-precision regression model.

[0035] The sample set includes retinal and ocular images of 3000 patients with myopia and amblyopia. An AI model was used to determine the approximate content range of various pigments (melanin, lipofuscin, lutein, etc.) in the corresponding patients' eyes. Based on the different absorption wavelength ranges of each pigment for laser light, a laser frequency that maximizes the total energy absorbed by all pigments was assigned. The AI ​​model also generated vascular segmentation masks in the retinal images of the corresponding patients and assigned different vascular density ratings based on the mask images. The ratings were divided into 10 levels, with corresponding laser intensities set to 1.1mW, 1.2mW, 1.3mW…2mW (increasing by 0.1mW each time). Lower vascular density resulted in lower laser intensity, and vice versa. The given laser frequency and intensity were used as labels, along with the corresponding 3000 retinal images, to form the sample set.

[0036] When the device is started, the sensing component will first collect information about the user's eyes. The ordinary camera in the sensing component will first take a picture of the user's eyes i1, and then the fundus camera will take a picture of the user's retina i2. After the image acquisition is completed, the sensing component will output all the images to the control module.

[0037] After receiving the image information, the control module uses its built-in pupil segmentation model f1, based on the human eye image i1, to calculate the segmentation mask o1 = f1(i1). Based on this, the intelligent chip uses Python's cv2.Canny method to perform edge detection on o1, obtaining a list of edge points. The average coordinates of these edge points are calculated to obtain the center O of the ellipse, which is the center point of the pupil. This point is then projected onto the observation port plane of this invention and connected to the base of the laser generator. The direction of this line is used as the laser's illumination angle, and the control circuit adjusts the orientation of the base to change the illumination angle. The circular laser beam P emitted by the laser generator is translated along this line until its center coincides with the intersection point. The projection of P onto the mask plane is the area P' of the laser beam irradiated on the laser mask.

[0038] Subsequently, the integrated segmentation model f2 based on the retinal image built into the control module calculates the integrated segmentation mask o2 = f2(i2) for the retinal image i2, and locates the macular mask i2_blo; it calculates the portion P_blo corresponding to the region i2_blo in i2 = P'*i2_blo; the regression model f3 built into the control module further processes o2 to obtain the light intensity and light frequency (I,f) = f3(o2) of the laser suitable for the user; the control group controls the light-shielding component according to P_blo, and controls the laser component according to I, f, and the illumination angle.

[0039] like Figure 3As shown, the base for fixing the laser generator in the laser assembly includes: a concave hemispherical base with a built-in movable magnet controlled by an electric field and a smooth hemisphere with a built-in magnet located at the bottom of the laser emitter. The movable magnet in the base is connected to a control module, which outputs a signal containing irradiation angle information to adjust the orientation of the laser generator in a small range in two degrees of freedom (excluding rotation) to achieve a suitable irradiation angle. The laser emitter itself emits a stable single-frequency parallel laser (which can be regarded as a surface light source), and the intensity and frequency of the laser are adjusted by the control module.

[0040] like Figure 4 As shown, when the shielding component uses a liquid crystal mask, it specifically includes: a control circuit and several sub-units. Each sub-unit includes a pair of shielding plates and a liquid crystal sub-unit disposed therebetween. The polarization directions of the pair of shielding plates are perpendicular to each other and are respectively connected to the control circuit. The control circuit controls the orientation of the liquid crystal in each sub-unit by energizing and de-energizing each sub-unit, thereby changing the ability of the liquid crystal to change the polarization direction of light, thus changing the light transmittance of the mask itself, so as to achieve the purpose of controlling the laser irradiation area.

[0041] like Figure 5 As shown, when the shielding assembly adopts a mechanical mask, it specifically includes: a control circuit, a shielding plate with several small holes, and a shielding plate movably disposed on each small hole, wherein: the shielding plate is connected to the control circuit to realize lifting and lowering, thereby allowing the small hole to transmit light or block light, so as to achieve the purpose of controlling the laser irradiation area.

[0042] Compared with existing technologies, this invention applies an AI model to low-intensity laser irradiation treatment for myopia and amblyopia to avoid irradiating the macula, thereby preventing macular degeneration caused by laser irradiation and making the treatment safer; it can intelligently select light intensity and light frequency more suitable for the user, making the treatment more effective.

[0043] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. An adaptive myopia and amblyopia treatment device, characterized in that, include: A pair of optical lens groups, a blocking component, a laser component, an optical sensing component, and a control module connected to the blocking component, the laser component, and the optical sensing component are arranged symmetrically in sequence. The control module calculates the illumination area based on the image acquired by the optical sensing component and outputs control commands to the blocking component and the laser component to control the light intensity, frequency, illumination angle, and illumination area. The images acquired by the optical sensing component include pupil images and retinal images; The shielding component is a liquid crystal mask or mechanical mask connected to the control module. Under the control of the control module, the area through which the laser passes is adjusted to achieve the effect of shielding the macula area. The control module includes: a pupil image segmentation unit, a retinal image segmentation unit, a mask control unit, and a laser control unit. Specifically: the pupil image segmentation unit obtains the pupil position in the human eye image using a pupil segmentation AI model; the retinal image segmentation unit obtains the positions of blood vessels, the optic disc, and the macula in the retinal image using a comprehensive retinal image segmentation AI model; the mask control unit controls the light-shielding mask of the occlusion component based on the macula position information obtained by the retinal image segmentation unit, ensuring that laser irradiation avoids the macula region; the laser control unit calculates the corresponding irradiation angle based on the pupil position information obtained by the pupil image segmentation unit and the positions of blood vessels, the optic disc, and the macula in the retina obtained by the retinal image segmentation unit, generates control commands, and outputs them to the occlusion component; it also calculates the corresponding laser intensity and frequency based on the retinal information, generates control commands, and outputs them to the laser component, thereby adaptively adjusting the laser irradiation angle, intensity, and frequency. The pupil segmentation AI model includes an encoder, a feature processor, and a decoder, wherein: the encoder extracts five 16×16×1024 feature vectors from the input pupil image; the feature processor generates a soft attention vector based on the five feature vectors and fuses them to obtain a 16×16×1024 feature vector; and the decoder generates a pupil segmentation mask image based on the feature vector. The retinal image comprehensive segmentation AI model includes an encoder, a feature processor, and a decoder, wherein: the encoder extracts five 16×16×1024 feature vectors from the input retinal image; the feature processor generates a soft attention vector based on the five feature vectors and fuses them to obtain a 16×16×1024 feature vector; and the decoder generates a retinal segmentation mask based on the feature vector. The encoders in the pupil segmentation AI model and the retinal image comprehensive segmentation AI model both contain five convolutional modules and a max pooling layer, wherein each convolutional module consists of a convolutional layer-batch processing layer-ReLU activation function-convolutional layer-batch processing layer-ReLU activation function; The decoders in the pupil segmentation AI model and the retinal image comprehensive segmentation AI model both include four upper convolutional modules, four residual convolutional layers, and one convolutional layer with a 1×1 kernel. Each upper convolutional module consists of an upsampling layer, a convolutional layer, a batch processing layer, and a ReLU activation function. The feature processors in both the pupil segmentation AI model and the retinal image comprehensive segmentation AI model employ a dilated spatial pooling pyramid (ASPP) module with attention engineering. This module performs the following operations on the input vector:

1. Global average pooling + upsampling to the original size; 2. Convolution with a 1×1 kernel, stride of 1, no padding, and no dilation; 3. Convolution with a 3×3 kernel, stride of 1, padding of 6, and dilation of 6; 4. Convolution with a 3×3 kernel, stride of 1, padding of 12, and dilation of 12; 5. A 3×3 convolution kernel with a stride of 1, padding of 18, and dilation of 18 is used. The five vectors obtained above are concatenated together at the channel level and passed through a convolutional layer with a 3×3 kernel, padding of 1, stride of 1, 5120 input channels, and 1024 output channels to obtain a 16×16×1024 vector as a soft attention vector. This vector is then multiplied by the original feature vector to obtain a 16×16×1024 feature vector.

2. The adaptive myopia and amblyopia treatment device according to claim 1, characterized in that, The optical sensing component includes: an optical camera for capturing images of the human eye and a fundus camera for capturing images of the retina, wherein: the optical camera outputs the captured images of the human eye to the control module for locating the pupil position, thereby adjusting the orientation of the laser component; the fundus camera outputs the retinal images to the control module for identifying and segmenting the position of the macular region, and combining the human eye images to confirm and optimize the light intensity and light frequency, as well as the area that the laser mask needs to block.

3. The adaptive myopia and amblyopia treatment device according to claim 1, characterized in that, The pupil segmentation AI model is trained as follows: 3000 human eye images with pupil masks are used as a sample set and randomly divided into training and test sets at a ratio of 9:

1. Then, a dataset is generated using the Dataset class in the PyTorch package of Python, and the Dataloader class is used to optimize the model training. Afterwards, the training set data is processed in batches through the model to obtain results, and the accuracy and cross-entropy are calculated using the results and the labels of the training set. The cross-entropy is used as the loss function in the optimization process, and the optim.Adam class of the Adam algorithm built into PyTorch is used to optimize the model, finally obtaining a high-precision pupil segmentation model.

4. The adaptive myopia and amblyopia treatment device according to claim 1, characterized in that, The aforementioned retinal image comprehensive segmentation AI model is trained as follows: 3000 human eye images with retinal masks are used as a sample set, randomly divided into training and test sets at a ratio of 9:

1. A dataset is then generated using the Dataset class in the PyTorch package of Python, and the Dataloader class is used to optimize the model training. The training set data is then processed in batches through the model to obtain results. The accuracy and L2 loss are calculated using the results and the labels of the training set, with the L2 loss serving as the loss function during the optimization process. Finally, the optim.Adam class of the Adam algorithm built into PyTorch is used to optimize the model, resulting in a high-precision retinal comprehensive segmentation model.

5. The adaptive myopia and amblyopia treatment device according to claim 1, characterized in that, The laser control unit incorporates a regression model and a control circuit connected to the laser component; The regression model includes three convolutional layers, two average pooling layers, and one fully connected layer. The convolutional layers extract a 256×256×64 feature vector from the input 256×256×1 retinal segmentation mask. The feature vector is then transformed into a 64×64×64 feature vector by the average pooling layer, then into a 64×64×256 feature vector by the convolutional layer, downsampled to a 1×256 feature vector by the global average pooling layer, and finally obtained two values, light intensity and light frequency, by the fully connected layer.

6. The adaptive myopia and amblyopia treatment device according to claim 5, characterized in that, The regression model is trained as follows: 3000 images with light intensity and frequency are used as a sample set, randomly divided into training and test sets at a ratio of 9:

1. A dataset is generated using the Dataset class in the PyTorch package of Python, and the Dataloader class is used to optimize the model training. The training set data is then processed in batches through the model to obtain results, and the L1 loss is calculated using the results and training set labels. The L1 loss is used as the loss function in the optimization process, and the optim.Adam class of the built-in Adam algorithm in PyTorch is used to optimize the model, ultimately obtaining a high-precision regression model.

Images