Biometric image processing program, biological image processing device, and biological image processing method
The biological image processing program uses a lightweight bottleneck narrowing Unet with attention blocks to simplify arithmetic processing and enhance segmentation accuracy in biological images.
Patent Information
- Application Number
- JP2023553835
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2021-10-13
- Publication Date
- 2024-12-05
- Estimated Expiration
- 2041-10-13
AI Technical Summary
Denoising and segmenting biological images requires complex arithmetic processing and high-performance hardware resources.
A biological image processing program that adds an attention block for averaging channels to a convolution block, reducing feature points and extracting important features through a lightweight bottleneck narrowing Unet with attention blocks (LWBNA_Unet) for simple arithmetic processing.
Accurately segments biological images with reduced computational requirements, improving segmentation accuracy and efficiency.
Smart Images

Figure 0007594827000001 
Figure 0007594827000002 
Figure 0007594827000003
Abstract
Description
[Technical field]
[0001] The technology described in this specification relates to a biometric image processing program, a biometric image processing device, and a biometric image processing method. [Background technology]
[0002] The captured biological image (in other words, medical image) may be denoised (in other words, image quality improved) and the target area may be segmented from the biological image.
[0003] For example, in Patent Document 1, a medical image of a specific part of a subject is acquired, and a high-quality image is generated from the acquired medical image using a high-quality image engine including a machine learning engine obtained using learning data in which noise corresponding to the state of at least some areas of the medical image is added to at least some areas. [Prior art documents] [Patent documents]
[0004] [Patent Document 1] JP 2020-166813 A [Patent Document 2] JP 2020-163100 A [Patent Document 3] JP 2019-187551 A [Non-patent literature]
[0005] [Non-Patent Document 1] U-Net: Convolutional Networks for Biomedical Image Segmentation; 2015; Cham. Springer International Publishing. [Non-Patent Document 2] Hu J, Shen L, Albanie S, et al. Squeeze-and-Excitation Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 2020;42(8):2011-23 doi: 10.1109 / TPAMI.2019.2913372[published Online First: Epub Date]. Summary of the Invention [Problem to be solved by the invention]
[0006] However, denoising of biological images may require complex computational processing and high-performance hardware resources.
[0007] In one aspect, the technology described in this specification aims to perform accurate segmentation of a biometric image by simple computational processing. [Means for solving the problem]
[0008] In one aspect, the biometric image processing program causes a computer for performing segmentation of a biometric image to perform a process of reducing the size of the input biometric image by adding an attention block for averaging channels to a convolution block, thereby extracting feature points, and outputting information regarding a feature image obtained by segmenting the biometric image including the extracted feature points by contracting the channel with the attention block added to the convolution block. Effect of the Invention
[0009] As one aspect, the segmentation of a biometric image can be performed accurately by simple computational processing. [Brief description of the drawings]
[0010] [Figure 1] FIG. 2 is a diagram for explaining biometric image processing according to an embodiment. [Diagram 2] 2 is a diagram for explaining details of the blood vessel density calculation process shown in FIG. 1. FIG. [Diagram 3] 11A to 11C are diagrams illustrating a segmentation process of a biological image in an embodiment. [Figure 4] FIG. 4A is a flowchart for explaining the processing for the convolution operation block shown in FIG. 3, and FIG. 4B is a flowchart for explaining the processing for the attention block shown in FIG. [Diagram 5] 4A to 4C are diagrams illustrating contraction processing and classification processing in the segmentation processing shown in FIG. [Figure 6] 11A and 11B are diagrams for explaining a process for removing fake FAZ according to an embodiment; [Figure 7] This is a table comparing the performance of each type of deep learning model. [Figure 8] 1 is a graph comparing the Dice similarity coefficients for each type of deep learning model. [Figure 9] 9A is an OCTA image, FIG. 9B is a FAZ region image of GT, FIG. 9C is a FAZ region image of Unet, FIG. 9D is a FAZ region image of Unet_AB, FIG. 9E is a FAZ region image of Unet_AB_128_upsampling_Add, and FIG. 9F is a FAZ region image of LWBNA_Unet. [Figure 10] 1 is a block diagram illustrating an example of a hardware configuration of a biometric image processing apparatus according to an embodiment; [Figure 11] 11 is a block diagram illustrating an example of a software configuration of the biometric image processing device illustrated in FIG. 10. FIG. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0011] Hereinafter, an embodiment will be described with reference to the drawings. However, the embodiment described below is merely an example, and is not intended to exclude various modifications and application of techniques not explicitly stated in the embodiment. In other words, the present embodiment can be modified in various ways without departing from the spirit of the present invention.
[0012] In addition, each figure does not necessarily include only the components shown in the figure, but may include other components. In the following drawings, parts with the same reference numerals indicate the same or similar parts unless otherwise specified.
[0013] FIG. 1 is a diagram for explaining biometric image processing according to an embodiment.
[0014] As shown by reference symbol A1, the program accepts input of an optical coherence tomography angiography (OCTA) image of the retina as an original image (INPUT OCTA IMAGE).
[0015] As shown by reference symbol A2, the program detects and segments the foveal avascular zone (FAZ) region from the OCTA image, and calculates FAZ parameters (CALCULATE FAZ PARAMETERS). For example, the FAZ parameters include the area (e.g., 0.4716 mm ) of the FAZ region (FAZ_AREA). 2 ), Circularity Index (e.g., 0.6404), Circumference (3.04111 mm), Upper Half Detailed calculation of perimeter (e.g., 1.6503 mm), Lower half Detailed calculation of perimeter (e.g., 1.3908 mm), and Circumference with angle (e.g., 0-90°: 0.8164 mm, 90-180°: 0.8339 mm, 180-270°: 0.7101 mm, 270-360°: 0.6808 mm) may be calculated. Note that the closer the Circularity Index value is to 1, the more circular the shape is, and the closer the value is to 0, the more linear the shape is. The program determines the center of the FAZ and calculates the perimeters of the upper and lower halves by dividing the boundary of the FAZ into the upper and lower halves.
[0016] As indicated by reference symbol A3, the program performs calculation of vascular density. The details of the calculation process of vascular density will be described later with reference to FIG.
[0017] FIG. 2 is a diagram for explaining details of the blood vessel density calculation process shown in FIG.
[0018] As shown by reference symbol A31, the program uses a trained Deep Learning (DL) model on an original OCTA image to densify the image and obtain a denoised image.
[0019] As shown by reference symbol A32, the program generates a binary image from the denoised image and calculates the total image vascular density %wiVD including the FAZ region (e.g., 40.1203) and the total image vascular density %wiVD_without_FAZ excluding the FAZ region (e.g., 42.3787).
[0020] As shown in the reference symbol A33, since the center of the FAZ region is not exactly at the center point of the image, the program sets the center of the FAZ region as the image center.Then, the program divides the image into upper and lower halves from the center of the FAZ region, and calculates the vascular density of the upper and lower halves (Upper and lower half Vascular) depending on the presence or absence of the FAZ region.In the illustrated example, the vascular density %UH_VD of the upper half including the FAZ region is 39.7751, the vascular density %UH_VD noFAZ of the upper half not including the FAZ region is 42.3611, the vascular density %LH_VD of the lower half including the FAZ region is 38.8605, and the vascular density %LH_VD noFAZ of the lower half not including the FAZ region is 41.1779.
[0021] As shown at reference symbol A34, the program calculates para-foveal vascular density %pfVD (eg, 40.8791).
[0022] As shown by reference symbol A35, the program calculates the foveal vascular density including the FAZ region %fVD (Foveal Vascular density with faz; for example, 28.1084) and the foveal vascular density not including the FAZ region %fVD_without_FAZ (Foveal Vascular density without faz; for example, 38.5901).
[0023] As shown by reference symbol A36, the program calculates the vascular density with angle for each of the six angle regions with and without the FAZ region. For example, the vascular density of the angle region 0-60° including the FAZ region (0-60 with FAZ) is 35.8756, and the vascular density of the angle region 0-60° not including the FAZ region (0-60 without FAZ) is 39.0671.
[0024] FIG. 3 is a diagram for explaining segmentation processing of a biological image in the embodiment.
[0025] One DL model designed for biomedical image segmentation is Unet (see, for example, non-patent literature 1).
[0026] Unet has a reduction path (in other words, an encoder for extracting feature points) and an expansion path with skip connections (in other words, a decoder for image reconstruction).
[0027] The structure of Unet relies on a cascaded recurrent neural network (CNN) to extract regions of interest. The easiest way to promote somitogenesis is to make the network deeper by stacking more layers.
[0028] Generally, making the network deeper involves more parameter multiplications, requiring more computational and storage power.
[0029] The architecture of the lightweight DL model designed for ophthalmic image segmentation is shown in Figure 3. The segmentation process shown in Figure 3 utilizes a Unet with an Attention Block in the shrink, expand and bottleneck paths.
[0030] In the layer shown by code B1, the size of the input image is 320×320 with 3 channels, while the size remains 320×320 but the number of channels is increased to 128. In the layer shown by code B2, the size is compressed to 160×160, in the layer shown by code B3, the size is compressed to 80×80, in the layer shown by code B4, the size is compressed to 40×40, and in the layer shown by code B5, the size is compressed to 20×20. Note that the number of channels remains 128 in each layer.
[0031] Model parameters are trimmed by fixing the number of feature channels / filters to 128 in both erosion and augmentation paths, replacing concatenation (used for skip connections) with Add layers, and transposing 2D convolutional layers (conv2D, activation-sigmoid) with upsampling.
[0032] Each convolution block (Conv-Block) in Figure 3 is made from a layer of 3x3 convolutions with the same padding, followed by batch normalization and rectified linear units (relu) for activation, and then a decidual layer after maxpooling / upsampling.
[0033] To focus attention on the feature map layer in both the contraction and expansion paths, we add an attention block created by modifying the squeeze and excitation (SE) block (see, for example, Non-Patent Document 2).
[0034] The attention block performs global averaging of channels, followed by a fully connected layer, activation by relu and sigmoid functions, and multiplication of the resulting weights to each channel. The combination of relu and sigmoid activation limits the attention weights of channels to between 0.5 and 1.0. In other words, channels with low priority are suppressed. Here, low priority channels are channels that do not contain feature points such as the outline of the FAZ region or blood vessels. The biometric image that contains many feature points only after suppressing channels that do not contain feature points may be called a feature image.
[0035] That is, low-priority channels are suppressed by half instead of eliminating them completely, and the network may decide their importance in the next layer. In Unet-type models, the refinement of high-level feature maps at the centers of contraction and expansion paths is also important in the accurate reconstruction of segmented images.
[0036] In Non-Patent Document 2 and the like, the feature map at the bottleneck was not considered. Therefore, in this embodiment, focusing on the bottleneck, successive narrowing of channels is introduced. Here, only the functions necessary for the decoder to reconstruct the necessary divided images are allowed. This is done by reducing the number of channels by half in successive layer blocks. For example, in the bottleneck (i.e., intermediate block) shown by reference character B5 in FIG. 3, the number of channels is reduced from 128 to 16 in four stages. Fine-tuning is performed using skip connections followed by attention blocks, and the overall feature map before upsampling is further emphasized.
[0037] In the expansion path from the intermediate block to the generation of the output image, the size is gradually increased while the number of channels of the image is maintained at 128, in contrast to the contraction path of the input image, so that the output image has the same size as the input image, 320 x 320, and three channels.
[0038] In what follows, we demonstrate that Bottleneck Narrowing (BNA) with Attention Block is remarkably effective in identifying fake FAZ regions detected by the network in poor-quality OCTA images, and helps improve the overall segmentation accuracy of FAZ.
[0039] In this embodiment, we exploit image size reduction to extract important feature points and channel reduction at the bottleneck to reconstruct accurate segmented images by the encoder network. In this embodiment, we call this network Lightweight Bottleneck Narrowing Unet with Attention Block (LWBNA_UNet).
[0040] FIG. 4A is a flowchart for explaining the processing for the convolution operation block shown in FIG. 3, and FIG. 4B is a flowchart for explaining the processing for the attention block shown in FIG.
[0041] In FIG. 4A, the convolution operation block is transposed to a 2D convolution operation layer (conv 2D, activation-sigmoid) by upsampling (step S11).
[0042] The convolution operation block is subjected to batch normalization (step S12).
[0043] The convolution operation block activates relu (Activation-relu) (step S13), and the processing for the convolution operation block ends.
[0044] In FIG. 4B, the attention block performs global average pooling of channels (step S21).
[0045] The attention block densifies the channel (step S22).
[0046] The attention block performs relu+sigmoid activation (ACTIVETION relu + SIGMOID) (step S23).
[0047] The activated attention block is multiplied by the original attention block (step S24), and the processing for the attention block is then completed.
[0048] FIG. 5 is a diagram for explaining the contraction process and the classification process in the segmentation process shown in FIG.
[0049] As shown in D1, the block input to the segmentation process is shrunk to X-pix _* Y-pix * F, where M is 1 or more depending on the image size and "-pix" represents pixels. As shown in D2, the block is further shrunk to X-pix / 2 * Y-pix / 2 * 2*M*F, where F is 8 or more depending on the image size. As shown in D3, the block is further shrunk to X-pix / 4 * Y-pix / 4 * 4*M*F. As shown in D4, the block is further shrunk to X-pix / 8 * Y-pix / 8 * 8*M*F.
[0050] In the bottleneck (i.e., intermediate block) shown in code D5, the number of channels is reduced to 8*M*F, 4*M*F, 2*M*F, and M*F in four stages. Then, as shown in code D6, the FC1 layer and the FC2 layer are output. This output may be classified into Category-1, Category-2, Category-3, and so on. In the example (EXAMPLE 1) shown in code D61, the output is classified into three categories: "Category-1 bad", "Category-2 medium", and "Category-3 good". In addition, in the example (EXAMPLE 2) shown in code D62, the output is classified into two categories: "Category-1 DH-" and "Category-3 DH+". Note that DH- indicates no bleeding, and DH+ indicates absence or bleeding.
[0051] FIG. 6 is a diagram for explaining the process of removing fake FAZ according to an embodiment.
[0052] As shown by reference symbol C1, an unknown image for the DL model is input.
[0053] As shown in symbol C2, the Unet Trained DL model is referenced.
[0054] As shown by reference C3, the FAZ in the center of the image is corrected, and the fake FAZ in the lower left of the image is removed, resulting in the image shown by reference C4.
[0055] Then, as indicated by reference symbol C5, the FAZ parameters are calculated.
[0056] Figure 7 is a table comparing the performance of different types of deep learning models.
[0057] In this embodiment, we tested the segmentation of FAZ in OCTA images using the lightweight DL model (LWBNA_Unet). For comparison, we also constructed a standard Unet model with and without attention blocks.
[0058] In Figure 7, we summarize the total number of parameters in the models along with the size of the output file generated after training. Here, the name "Unet" represents the traditional Unet architecture. "Unet_AB" uses the contraction and augmentation paths of the traditional Unet architecture, plus a channel attention block. The models average 128 or 64 and have a constant number of filters (128 or 64) per 2D recurrent layer. Some models use upsampling, adding layers instead of transposed convolutional operations and concatenation layers, reducing the number of trainable parameters.
[0059] In Fig. 7, the performance of the trained DL models was tested on a test dataset of 145 OCTA images. The Dice similarity coefficient (D) was calculated between the manually segmented images and the images predicted by the DL models. The maximum contour D is calculated by selecting the largest segmented region in the predicted image (when multiple FAZ regions exist). The standard deviation is the standard deviation. All models were trained for a fixed number of epochs (500).
[0060] Figure 8 is a graph comparing the Dice similarity coefficient for each type of deep learning model. Figure 9A is an OCTA image, Figure 9B is a FAZ region image of GT, Figure 9C is a FAZ region image of Unet, Figure 9D is a FAZ region image of Unet_AB, Figure 9E is a FAZ region image of Unet_AB_128_upsampling_Add, and Figure 9F is a FAZ region image of LWBNA_Unet.
[0061] From the results in Figure 7, it is difficult to determine which model is the best for segmenting the FAZ. Therefore, the subdivision of the FAZ region predicted by the DL models is analyzed using box plots and shown in Figure 8. Each model has many outliers, and their spread is the smallest for Unet, Unet_AB, Unet_AB_upsampling, and LWBNA_Unet. The smallest number of outliers with D ≤ 0.94 (18) is for Unet_AB, whereas these are the same for Unet and LWBNA_Unet (21). In all three models, 14 outliers are common with the same average D of about 0.917. The main reason for the low D values for these 14 outliers is the unclear FAZ boundary due to the blurred images and the presence of too many scanlines. Accurate manual segmentation of the FAZ in these images is difficult even for humans. The remaining outliers (i.e., uncommon values) of each model mainly have D in the range of 0.91–0.94, again due to blurry images. Our observations show that both Unet_AB and LWBNA_Unet can accurately segment the FAZ.
[0062] The OCTA image shown in Figure 9A has a complex FAZ boundary due to its small size and the presence of relatively strong powder-like noise in the FAZ region. The D obtained from Unet (Figure 9C), Unet_AB_128_upsampling_Add (Figure 9E) is significantly lower than that of LWBNA_Unet (Figure 9F), Unet_AB (Figure 9D).
[0063] FIG. 10 is a block diagram illustrating an example of a hardware configuration of a biometric image processing apparatus 1 according to an embodiment.
[0064] As shown in FIG. 10, the biological image processing apparatus 1 has a server function and includes a CPU 11, a memory unit 12, a display controller 13, a storage device 14, an input interface (IF) 15, an external recording medium processing unit 16, and a communication IF 17.
[0065] The memory unit 12 is an example of a storage unit, and is illustratively a Read Only Memory (ROM) and a Random Access Memory (RAM). Programs such as a Basic Input / Output System (BIOS) may be written in the ROM of the memory unit 12. The software programs in the memory unit 12 may be appropriately read and executed by the CPU 11. The RAM of the memory unit 12 may be used as a temporary recording memory or a working memory.
[0066] The display control unit 13 is connected to a display device 131 and controls the display device 131. The display device 131 is a liquid crystal display, an organic light-emitting diode (OLED) display, a cathode ray tube (CRT), an electronic paper display, or the like, and displays various information for an operator, etc. The display device 131 may be combined with an input device, for example, a touch panel. In this embodiment, the display device 131 displays information related to the segmented feature image. Note that the output of information related to the segmented feature image may be achieved by a printing device (not shown) or the like.
[0067] The storage device 14 is a storage device with high IO performance, and may be, for example, a dynamic random access memory (DRAM), an SSD, a storage class memory (SCM), or an HDD.
[0068] The input IF 15 may be connected to input devices such as a mouse 151 and a keyboard 152, and may control the input devices such as the mouse 151 and the keyboard 152. The mouse 151 and the keyboard 152 are examples of input devices, and an operator performs various input operations via these input devices.
[0069] The external recording medium processing unit 16 is configured so that the recording medium 160 can be attached. The external recording medium processing unit 16 is configured so that, when the recording medium 160 is attached, the information recorded on the recording medium 160 can be read. In this example, the recording medium 160 is portable. For example, the recording medium 160 is a flexible disk, an optical disk, a magnetic disk, a magneto-optical disk, a semiconductor memory, or the like.
[0070] The communication IF 17 is an interface for enabling communication with an external apparatus.
[0071] The CPU 11 is an example of a processor, and is a processing device that performs various controls and calculations. The CPU 11 executes an Operating System (OS) and programs loaded into the memory unit 12, thereby implementing various functions.
[0072] The device for controlling the overall operation of the biometric image processing device 1 is not limited to the CPU 11, and may be, for example, any one of MPU, DSP, ASIC, PLD, and FPGA. The device for controlling the overall operation of the biometric image processing device 1 may be a combination of two or more of CPU, MPU, DSP, ASIC, PLD, and FPGA. Note that MPU is an abbreviation for Micro Processing Unit, DSP is an abbreviation for Digital Signal Processor, and ASIC is an abbreviation for Application Specific Integrated Circuit. Also, PLD is an abbreviation for Programmable Logic Device, and FPGA is an abbreviation for Field Programmable Gate Array.
[0073] FIG. 11 is a block diagram illustrating an example of the software configuration of the biometric image processing device 1 illustrated in FIG.
[0074] The CPU 11 of the biological image processing device 1 has functions as an extraction processing unit 111, an output processing unit 112, and a blood vessel density calculation unit 113.
[0075] The extraction processing unit 111 extracts feature points by reducing the image size of the input biometric image by adding an attention block for averaging channels to a convolution block. The extraction processing unit 111 may also extract feature points by preferentially reducing the outline of the segmentation target portion included in the biometric image or an area that does not include a blood vessel portion.
[0076] The output processing unit 112 outputs information about a feature image obtained by segmenting a biometric image including the extracted feature points by contracting a channel with an attention block added to the convolution block. After contracting the channel, the output processing unit 112 may expand the feature image to the size of the input biometric image and output the feature image.
[0077] The blood vessel density calculation unit 113 calculates the blood vessel density for each of one or more regions of the eyeball based on information related to the characteristic image.
[0078] This makes it possible to accurately perform segmentation of a biometric image through simple calculation processing.
[0079] The disclosed technology is not limited to the above-described embodiments, and various modifications can be made without departing from the spirit of the embodiments. The configurations and processes of the embodiments can be selected as necessary, or can be combined as appropriate.
[0080] In the above-mentioned embodiment, an example of segmenting the FAZ region from a biological image of an eyeball has been described, but the present invention is not limited thereto. The segmentation process in the embodiment can be used to segment a specific part from various biological images by changing the DL model applied to the LWBNA_UNet in the embodiment. For example, it can be used to segment a tumor such as a cancer cell from an image of an organ of a living body. In this case, an image of an organ having a tumor can be applied to the LWBNA_UNet in the embodiment as a DL model. [Explanation of symbols]
[0081] 1: Biometric image processing device 11: CPU 12: Memory section 13: Display control section 14:Storage device 15: Input IF 16: External recording medium processing unit 17: Communication Interface 100: Circuit diagram 101: npn transistor 102: Voltage VGG 103: Voltage VDD 104: Resistance RD 111: Extraction processing unit 112: Output processing section 113: Blood vessel density calculation section 131:Display device 151: Mouse 152: Keyboard 160: Recording media
Claims
1. A computer for performing segmentation of a biological image, comprising: For the input biometric image, an attention block for averaging channels is added to a convolution block to reduce the size of the biometric image and extract feature points; With the attention block added to the convolution block, the channel is contracted to output information about a feature image obtained by segmenting the biometric image including the extracted feature points. A biometric image processing program that executes the processing.
2. extracting the feature points by preferentially reducing an area that does not include an outline of a portion to be segmented or a blood vessel portion included in the biological image; The biometric image processing program according to claim 1 , which causes the computer to execute a process.
3. After shrinking the channel, the channel is expanded to the size of the input biometric image, and the feature image is output. The biometric image processing program according to claim 1 , which causes the computer to execute processing.
4. calculating a blood vessel density for each of one or more regions in the biological image based on information about the characteristic image; 4. The biometric image processing program according to claim 1, which causes the computer to execute a process.
5. the biometric image is an image of an eyeball, outputting information about the feature image obtained by segmenting the foveal avascular zone of the eyeball; 5. The biometric image processing program according to claim 1, which causes the computer to execute a process.
6. 1. A biometric image processing apparatus for performing segmentation of a biometric image, comprising: For the input biometric image, an attention block for averaging channels is added to a convolution block to reduce the size of the biometric image and extract feature points; With the attention block added to the convolution block, the channel is contracted to output information about a feature image obtained by segmenting the biometric image including the extracted feature points. A biometric image processing device comprising a processor.
7. The processor extracts the feature points by preferentially reducing an area that does not include an outline of a segmentation target portion included in the biological image or a blood vessel portion. The biometric image processing device according to claim 6 .
8. The processor contracts the channel, and then expands the channel to the size of the input biometric image to output the feature image. The biological image processing device according to claim 6 or 7.
9. The processor calculates a blood vessel density for each of one or more regions in the biological image based on information about the characteristic image. The biometric image processing device according to any one of claims 6 to 8.
10. the biometric image is an image of an eyeball, The processor outputs information about the feature image that segments the foveal avascular zone of the eye. The biometric image processing device according to any one of claims 6 to 9.
11. A computer for performing segmentation of a biological image, comprising: For the input biometric image, an attention block for averaging channels is added to a convolution block to reduce the size of the biometric image and extract feature points; With the attention block added to the convolution block, the channel is contracted to output information about a feature image obtained by segmenting the biological image. A biometric image processing method for performing processing.
12. extracting the feature points by preferentially reducing an area that does not include an outline of a portion to be segmented or a blood vessel portion included in the biological image; The method of claim 11 , wherein processing is performed by the computer.
13. After shrinking the channel, the channel is expanded to the size of the input biometric image, and the feature image is output. The method for processing a biometric image according to claim 11 or 12, wherein processing is performed by the computer.
14. calculating a blood vessel density for each of one or more regions in the biological image based on information about the characteristic image; The biometric image processing method according to any one of claims 11 to 13, wherein processing is performed by the computer.
15. the biometric image is an image of an eyeball, outputting information about the feature image obtained by segmenting the foveal avascular zone of the eyeball; The biometric image processing method according to any one of claims 11 to 14, wherein processing is performed by the computer.
Citation Information
Patent Citations
Conditional generative adversarial remote sensing image target segmentation method containing multi-level channel attention
CN111259906A
Network training method and device, image processing method and device, electronic equipment and storage medium
CN111310764A
Blood vessel segmentation method, device and equipment for retinal image and storage medium
CN111882566A
Cell nucleus segmentation method based on attention learning
CN112446892A
Image processing device, image processing method, and program
JP2019187551A
