Macular change image recognition method, system and equipment and storage medium

Abstract

The invention provides a maculopathy image recognition method, system and device and a storage medium, and belongs to the field of medical image processing, and the method comprises the steps: obtaining an ophthalmology OCT medical image; the method comprises the following steps: on the basis of a Byol dual-network architecture, updating parameters of a target network in the Byol dual-network architecture through an adaptive moving average (ARMA) method to obtain a pre-trained contrast learning network; and inputting the ophthalmology OCT medical image into the pre-trained contrast learning network, and labeling the OCT medical image with the disease. A CSPDarkNet53 structure is used as a main network, a ResBlock-body module in the CSPDarkNet53 structure is replaced by a PCRBlock residual error network module, PSAM attention is introduced into the PCRBlock residual error network module, and a medical image classification network is obtained; and inputting the screened OCT medical image into a medical image classification network to obtain a maculopathy classification recognition result. And under the condition that the existing data volume is not increased, the classification accuracy of the model on different maculopathy types is improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical image processing, and in particular relates to a method, system, device and storage medium for recognizing macular degeneration images. Background Art

[0002] Maculopathy is a common eye disease that primarily affects the macula, the central area of ​​the retina responsible for fine vision. Maculopathy can lead to decreased vision, distorted vision, and even blindness, making it crucial to tailor treatment options based on the type of lesion. Common retinal diseases include Vogt-Koyanagi-Harada syndrome, diabetic macular edema (DME), and age-related macular degeneration (AMD). Retinal optical coherence tomography (OCT) imaging offers non-contact, non-invasive, and high-resolution imaging. Accurately identifying patients using retinal OCT images is a key tool for ophthalmologists in diagnosing maculopathy. However, maculopathy is not easily distinguishable. Manual identification is not only time-consuming and labor-intensive, but also prone to subjective judgment errors.

[0003] Deep learning algorithms enable end-to-end automated learning and are widely used in medical image processing. However, the identification of macular lesions still faces the following challenges: Different types of macular lesions have varying sizes of lesion areas, requiring improved classification accuracy; and the large number of model parameters requires extensive data training, resulting in lengthy training times. Summary of the Invention

[0004] In order to address the shortcomings of existing deep learning models in the identification of macular lesions, the present invention provides a method, system, device and storage medium for macular lesion image recognition.

[0005] In order to achieve the above object, the present invention provides the following technical solutions: A method for recognizing macular degeneration images, comprising the following steps: Acquire ophthalmic OCT medical images; Based on the Byol dual-network architecture, the parameters of the target network in the Byol dual-network architecture are updated by an adaptive sliding average ARMA method to obtain a pre-trained contrastive learning network; the ophthalmic OCT medical image is input into the pre-trained contrastive learning network, and the ophthalmic OCT medical image is labeled with diseases to distinguish between diseased images and normal images; Based on the original CSPDarkNet53 structure, the PCRBlock residual network module is introduced to replace the ResBlock-body module in the original CSPDarkNet53 structure, and PSAM attention is introduced into the PCRBlock residual network module to obtain a medical image classification network; the annotated disease image is input into the medical image classification network to obtain the classification result of macular lesions.

[0006] Preferably, the pre-trained contrastive learning network includes an online network and a target network, the online network consists of an encoder, a converter and a predictor, and the target network includes an encoder and a converter; the encoder and converter of the target network have the same structure as the online network.

[0007] Preferably, the updating of the parameters of the target network in the Byol dual network architecture by the adaptive sliding average ARMA method specifically includes the following steps: Initialize the window and random value, and update the parameters of the target network using the following formula: ; in, , , is white noise, are the parameters of the autoregressive model, , ,.... are the parameters of the moving average model; ptimizer( , , ); ; Among them, optimizer is the optimizer, , is the learning rate; Get the updated parameters of the target network.

[0008] Preferably, the PCRBlock residual network module includes a trunk branch and a residual edge part. The input image is first down-sampled by a 3*3 convolution operation, and then feature transformed by two 1*1 convolutions to output the first tensor. The other path passes through the PCRblock module to output the second tensor. The two tensors are spliced ​​and then subjected to a 1*1 convolution operation for feature fusion.

[0009] Preferably, the medical image classification network divides the input annotated ophthalmic OCT medical image into two paths, the first path performs convolution and multiple PCRBlock residual network processing; the second path performs a 1*1 convolution operation on the annotated ophthalmic OCT medical image, and then directly connects the outputs of the first and second paths through a shortcut.

[0010] Preferably, the PSAM attention includes a channel attention module PAM and a spatial attention module SAM.

[0011] Preferably, the classification results of macular lesions include VKH, hole, retinal detachment, AMD and edema.

[0012] The present invention also provides a macular degeneration image recognition system, specifically comprising: The data acquisition module is used to acquire ophthalmic OCT medical images.

[0013] The network preprocessing module is used to update the parameters of the target network in the Byol dual-network architecture through the adaptive sliding average ARMA method to obtain a pre-trained contrast learning network; the ophthalmic OCT medical image is input into the pre-trained contrast learning network, the ophthalmic OCT medical image is labeled with diseases, and the disease image is distinguished from the normal image.

[0014] The lesion classification module is used to introduce the PCRBlock residual network module to replace the ResBlock-body module in the original CSPDarkNet53 structure based on the original CSPDarkNet53 structure, and introduce PSAM attention into the PCRBlock residual network module to obtain a medical image classification network; the annotated disease image is input into the medical image classification network to obtain the classification result of macular lesions.

[0015] The present invention also provides a computer device, comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps described in the method for identifying macular lesions images.

[0016] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is loaded by a processor, it can execute the steps described in the method for recognizing macular lesions images.

[0017] The present invention provides a method for recognizing macular degeneration images, which has the following beneficial effects: The present invention obtains ophthalmic OCT medical images; based on the Byol dual network architecture, the parameters of the target network in the Byol dual network architecture are updated by the adaptive sliding average ARMA method, and the parameters are adaptively updated according to the parameter change trend to improve the training stability and obtain a pre-trained contrast learning network. The ophthalmic OCT medical images are input into the pre-trained contrast learning network for disease labeling, and disease images and normal images are distinguished, which is beneficial for subsequent lesion classification. With the CSPDarkNet53 structure as the backbone network, the PCRBlock residual network module is introduced to replace the ResBlock-body module in the CSPDarkNet53 structure to construct a medical image classification network. Compared with the ResBlock-body module, the PCRBlock residual network introduces PSAM attention, aggregates a wide range of associated information in the image, obtains deep features related to the pathology, and improves the accuracy of classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] To more clearly illustrate the embodiments of the present invention and its design, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort.

[0019] Figure 1 The present invention is a flowchart of a method for recognizing macular degeneration images.

[0020] Figure 2 This is an OCT case image in an embodiment of the present invention.

[0021] Figure 3 Schematic diagram of the process of the AMBY self-supervision method in an embodiment of the present invention.

[0022] Figure 4 This is a medical image classification network diagram in an embodiment of the present invention.

[0023] Figure 5 This is the PCRBlock residual network module in the embodiment of the present invention.

[0024] Figure 6 2 is a structural diagram of PSAM in an embodiment of the present invention.

[0025] Figure 7 2 is a structural diagram of the PAM module in an embodiment of the present invention.

[0026] Figure 8 2 is a structural diagram of the SAM module in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to enable those skilled in the art to better understand the technical solution of the present invention and to be able to implement it, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and are not intended to limit the scope of protection of the present invention.

[0028] Example The present invention provides a method for recognizing macular degeneration images. Figure 1 As shown, the specific steps include: S1. Obtain ophthalmic OCT medical images.

[0029] S2. Construct a self-supervised pre-training AMBY method based on contrastive learning. The AMBY method is specifically based on the Byol dual network architecture and updates the parameters of the target network in the Byol dual network architecture through the ARMA method (i.e., adaptive sliding average method). Figure 3 As shown in Figure 2, the goal of the AMBY method is to learn the visual representation w and then use it in downstream tasks. It is learned through two neural networks: the online network and the target network. The online network consists of an encoder , converter and predictor The target network consists of an encoder and converter , the encoder of the target network , converter It has the same structure as the online network, but the network parameter set x is different. The encoder f can be the backbone of any neural network, such as CSPDarkNet. , converter and predictor Both belong to the structure type of multi-layer perceptron and use fully connected layers. The specific process is to generate two random views of the same input medical image through data enhancement, which includes cropping, rotation, color perturbation, etc. After passing through the encoders of the two networks, different feature vectors w and The specific process includes: randomly selecting two enhanced images, passing through the encoders of the two networks and outputting different feature vectors w and , the feature alignment loss between the online network and the target network in self-supervised learning is:

[0030] ; in, To prevent the gradient from backpropagating to the target network and keep the parameters stable, stable training is achieved through prediction tasks and gradient isolation mechanism, and the prediction of online network Need to approximate the output of the target network The target network provides the regression target for the online network, and the parameters of the target network are According to online network parameters The parameters of the online network are updated through back propagation and gradient descent, and the parameters of the target network are updated through the ARMA method, i.e. adaptive sliding average. The specific expression is as follows:

[0031] ; in, , is white noise, are the parameters of the autoregressive model, , ,.... is the parameter of the moving average model. p is the order of autoregression, which determines how many past observations the model relies on. q is the order of moving average, which determines how many past errors the model relies on. Therefore, the overall update process of the model is:

[0032] ptimizer( , , ); ; Among them, optimizer is the optimizer, , is the learning rate. After training, the encoder is retained and weight parameters.

[0033] The model with updated parameters is used to preprocess OCT medical images, learn common features in ophthalmic OCT medical images, screen out disease images and normal images, and annotate disease images.

[0034] S3, using the CSPDarkNet53 structure as the backbone network, introduce the PCRBlock residual network module to replace the ResBlock-body module in the CSPDarkNet53 structure to build a medical image classification network, such as Figure 4 As shown in the figure, the CSPDarkNet53 structure is a general network architecture. Compared to the ResBlock-body module, the PCRBlock residual network introduces the PSAM attention mechanism, which can adaptively aggregate a wide range of contextual information in the image, effectively improving the feature representation capability of image recognition.

[0035] Medical image classification networks divide input disease images into two parts. One part processes the disease image through convolution, while the other part uses shortcut connections to retain more feature information while reducing the computational burden. The residual edges are directly connected to the input and output of the backbone. A shortcut is a connection method that connects the previous and next layers in series.

[0036] Since the network uses a convolution kernel with a stride of 2 and a size of 3×3 for downsampling, although it can extract fine-grained and rich semantic information and increase the network's receptive field, it will cause the loss of feature information in small target areas. Introducing the attention mechanism can enhance the network's representation ability. Introducing the PCRBlock residual network module to replace the ResBlock-body module in the CSPDarkNet53 structure. PCRBlock residual network module, such as Figure 5 As shown in the figure, it consists of a main branch and a residual edge. The input image is downsampled through a 3x3 convolution. Based on the number of channels, two 1x1 convolutions are used to transform features. One path directly outputs a tensor, while the other passes through the PCRblock module to output a tensor. These two tensors are then concatenated and subjected to another 1x1 convolution to output features.

[0037] At the same time, PSAM attention is introduced in the PCRBlock residual network. PSAM attention includes channel attention module (PAM) and spatial attention module (SAM), such as Figure 6 shown.

[0038] The working principle of the channel attention module (PAM) is to use the average pooling downsampling method to downsample the feature map x (B, C, H, W) to obtain a smaller feature map (B, C, , ). The downsampled feature map is normalized, and then a 1×1 depthwise separable convolution is used to generate queries (q), keys (k), and values ​​(v) for calculating the relationship between channels. The query and key are dot-producted, the attention matrix is ​​calculated, and a scaling factor is applied to prevent numerical overflow. The final attention weight is calculated through a gating mechanism, where Dropout is used for regularization. The attention weight is then matrix-multiplied with the value (v) to obtain the weighted feature, which is output as a channel-enhanced feature map. Compared with using common convolution operations to model channel dependencies, PAM has stronger input perception capabilities and can more effectively use features to deepen learning, such as Figure 7 The specific calculation steps are as follows:

[0039] ; ; ; ; ; in, It means that a convolution with a kernel size of n×n completes a pooling operation to adjust the image with a resolution of (H, W) to (H^, W^). Represents the linear projection of query, key, and value generation. In the process of generating queries, keys, and values, depthwise separable convolution is selected, which has fewer parameters and faster running speed than traditional ordinary convolution. , In this step, a complete embedding is organically divided into eight small embeddings. Independent operations are performed in each small embedding. After the operation is completed, they are linked through Concat and then linearly transformed to return to the original shape. In order to enable different sub-features to fully interact with each other, multi-head attention is used in CASA. Since multi-head attention allows the model to observe input data from multiple perspectives at the same time, it can better capture the different dependencies of the input sequence and improve the expressive ability of the model. Each head performs an independent linear transformation, allowing the model to learn different contextual dependencies, improving the model's expressiveness and generalization capabilities.

[0040] The Spatial Attention Module (SAM) fuses the multi-scale features output by ASPP to generate a spatial attention matrix. The matrix models the spatial relationship between any two pixels in the feature and serves as the input for attention calculation. Matrix multiplication is performed between the attention matrix and the original features, and element-wise summation is performed on the result matrix of the previous two steps and the original features to obtain the final representation reflecting the long-range context. Figure 8 The specific calculation steps are as follows:

[0041] = ; ; ; ; ; First, the input feature map passes through the ASPP structure. The ASPP structure has four branches. Branch one is a 1×1 convolution, which can capture local details. Branches two and three are 3×3 hole convolutions. Branch four is global average pooling and upsampling, which can capture global context. is the feature output by each branch i, is the output feature of the global branch, Feature concatenation is performed. The ASPP output channels are compressed to C / r (r=8 for the compression ratio) via 1×1 convolutions. The query and key are then compressed using the Q and K formulas. The matrices are then transformed, with Q and K reshaped into 2D matrices. Inter-position correlations are calculated via matrix multiplication, and softmax is applied along the rows for normalization. Weighted aggregation with the original value is then performed to transform the matrix back to the original spatial dimensions. This module enhances the perception of distant context and small objects and uses residual connections to gradually activate the attention mechanism during training, enhancing network stability and reducing network complexity.

[0042] S4. Input the images with disease after screening into the medical image classification network to obtain the recognition results of macular lesions. The recognition results include VKH, hole, retinal detachment, AMD and edema. Figure 2 Shown are example images of various case images and normal images.

[0043] Introducing contrastive learning as pre-training in the network improves network performance, effectively addressing the challenges of small feature differences in medical images and the difficulty of learning them. This is shown in Table 1. Comparative experiments demonstrate the effectiveness of contrastive learning in feature representation after introducing MoCo (Momentum Contrastive Learning), SimCLR (Contrastive Learning Framework), and the proposed AMBY method. Pre-training with the AMBY method boosts the algorithm's mean average performance (MAP) to 88.7%, a 4% to 9% improvement in accuracy compared to other contrastive learning models. This significantly improves the network's detection accuracy. The proposed AMBY method is based on an asymmetric dual-network structure. The predictor in the online network incorporates more encoded information in the network's subsequent computational stages. The dual networks utilize different parameter update methods, and the proposed method uses only positive samples, effectively avoiding model collapse.

[0044] Table 1 Learning results of AMBY and contrastive learning as pre-training methods To verify the effectiveness of the PSAM attention mechanism, we conducted experiments on algorithms pre-trained using the AMBY method. This is shown in Table 2. Adding channel and spatial attention modules effectively enhances the model's ability to capture key features. With the integration of these modules, the number of parameters gradually increases, and the FPS decreases, but the mAP significantly improves. In this experiment, the PSAM attention mechanism enabled more accurate diagnostics in medical image recognition. In practical applications, this speed loss can be offset simply by pairing it with high-performance hardware.

[0045] Table 2 Comparison of the effectiveness of PSAM mechanisms To verify the superiority of the PSAM mechanism over other attention mechanisms, we introduced other attention mechanisms for comparative experiments. As shown in Table 3, the PSAM attention mechanism significantly outperforms the CBAM attention mechanism in mAP, achieving a 13.2% improvement in mAP compared to the worst CBAM attention mechanism. Incorporating the PSAM attention mechanism into the residual network improves the network's representational capabilities while highlighting key features. This effectively addresses the challenges of small medical image objects, minimal feature variance, difficult feature extraction, and significant loss of feature information through network propagation. The proposed PSAM attention mechanism demonstrates significant advantages in high-precision scenarios.

[0046] Table 3 Comparison results of PSAM and other attention mechanisms The present invention also provides a macular degeneration image recognition system, specifically comprising: The data acquisition module is used to acquire ophthalmic OCT medical images.

[0047] The network preprocessing module is used to update the parameters of the target network in the Byol dual-network architecture through the adaptive sliding average ARMA method to obtain a pre-trained contrastive learning network; the ophthalmic OCT medical images are input into the pre-trained contrastive learning network, and the ophthalmic OCT medical images are labeled with diseases to distinguish between disease images and normal images.

[0048] The lesion classification module is used to replace the ResBlock-body module in the original CSPDarkNet53 structure with the PCRBlock residual network module, and to introduce PSAM attention into the PCRBlock residual network module to obtain a medical image classification network. The annotated disease image is input into the medical image classification network to obtain the classification results of macular lesions.

[0049] Each module in the aforementioned macular degeneration image recognition system can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a computer device memory in the form of software, so that the processor can call and execute the corresponding operations of each module.

[0050] The present invention also provides a computer device comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of an embodiment of a method for recognizing macular lesions. The specific implementation method can be found in the method embodiment and will not be described in detail here.

[0051] Furthermore, the present invention provides a non-transitory computer-readable storage medium containing instructions, wherein the storage medium stores a computer program. For example, this may be a memory device containing instructions, wherein the instructions are executable by a processor of a computer device to perform the above-described method. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device. When executed by the processor, this computer program can implement the steps in an embodiment of a method for recognizing macular lesions images. Specific implementation methods can be found in the method embodiments and are not further detailed here.

[0052] Those skilled in the art will appreciate that embodiments of the present invention may provide methods, systems, or computer program products. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0053] The present invention is described with reference to flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0054] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0055] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1The steps for the function specified in one or more boxes.

[0056] It should be pointed out that the specific implementation methods described above can enable those skilled in the art to understand the invention more comprehensively, but do not limit the invention in any way. Therefore, although the present specification and examples have described the invention in detail, those skilled in the art should understand that the invention can still be modified or replaced by equivalents; and all technical solutions and improvements that do not deviate from the spirit and scope of the invention are included in the scope of protection of the patent for the invention. Any figure mark in the claims should not be regarded as limiting the claims involved. Any simple change or equivalent replacement of the technical solution that can be obviously obtained by any person familiar with the art within the technical scope disclosed in the present invention falls within the scope of protection of the present invention.

Claims

1. A method for recognizing macular degeneration images, characterized in that: The following steps are involved: Acquire ophthalmic OCT medical images; Based on the Byol dual-network architecture, the parameters of the target network in the Byol dual-network architecture are updated by an adaptive sliding average ARMA method to obtain a pre-trained contrastive learning network; the ophthalmic OCT medical image is input into the pre-trained contrastive learning network, and the ophthalmic OCT medical image is labeled with diseases to distinguish between diseased images and normal images; Based on the original CSPDarkNet53 structure, the PCRBlock residual network module is introduced to replace the ResBlock-body module in the original CSPDarkNet53 structure, and the PSAM attention is introduced into the PCRBlock residual network module to obtain a medical image classification network; The labeled disease image is input into the medical image classification network to obtain a classification result of macular degeneration.

2. The method for recognizing macular degeneration images according to claim 1, wherein: The pre-trained contrastive learning network includes an online network and a target network. The online network is composed of an encoder, a converter and a predictor. The target network includes an encoder and a converter. The encoder and converter of the target network have the same structure as the online network.

3. The method for recognizing macular lesions according to claim 1, wherein: The method of updating the parameters of the target network in the Byol dual network architecture by the adaptive sliding average ARMA method specifically includes the following steps: Initialize the window and random value, and update the parameters of the target network using the following formula: ； in, , , is white noise, are the parameters of the autoregressive model, , ,.... are the parameters of the moving average model; ptimizer( , , ); ； Among them, optimizer is the optimizer, , is the learning rate; Get the updated parameters of the target network.

4. The method for recognizing macular degeneration images according to claim 1, wherein: The PCRBlock residual network module consists of a trunk branch and a residual side part. The input image is first downsampled by a 3*3 convolution operation, and then the feature is transformed by two 1*1 convolutions to output the first tensor. The other path passes through the PCRblock module to output the second tensor. The two tensors are spliced ​​and then subjected to another 1*1 convolution operation for feature fusion.

5. The method for recognizing macular degeneration images according to claim 1, wherein: The medical image classification network divides the input annotated ophthalmic OCT medical image into two paths, the first path undergoes convolution and multiple PCRBlock residual network processing; The second path is to perform a 1*1 convolution operation on the annotated ophthalmic OCT medical image, and then directly connect the outputs of the first and second paths through a shortcut.

6. The method for recognizing macular degeneration images according to claim 1, wherein: The PSAM attention includes a channel attention module PAM and a spatial attention module SAM.

7. The method for recognizing macular degeneration images according to claim 1, wherein: The classification results of maculopathy include VKH, hole, retinal detachment, AMD and edema.

8. A macular degeneration image recognition system, characterized in that: include: A data acquisition module, used to acquire ophthalmic OCT medical images; A network preprocessing module is configured to update the parameters of the target network in the Byol dual-network architecture using an adaptive sliding average ARMA method to obtain a pretrained contrastive learning network; input the ophthalmic OCT medical image into the pretrained contrastive learning network, perform disease labeling on the ophthalmic OCT medical image, and distinguish between diseased images and normal images; A lesion classification module is used to replace the ResBlock-body module in the original CSPDarkNet53 structure with the PCRBlock residual network module, and to introduce PSAM attention into the PCRBlock residual network module to obtain a medical image classification network. The labeled disease image is input into the medical image classification network to obtain a classification result of macular degeneration.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is loaded into a processor, it is capable of executing the steps of the method according to any one of claims 1 to 7.

Images