Semi-supervised ultrasonic video focus segmentation method and system

Abstract

The invention discloses a semi-supervised ultrasonic video focus segmentation method and system, and is applied to the technical field of video segmentation, and the method comprises the steps: obtaining ultrasonic video data which comprise marked ultrasonic video data and unmarked ultrasonic video data; inputting the marked ultrasonic video data into a student model for processing, and outputting a first segmentation result; inputting the unmarked ultrasonic video data into the student model and the teacher model for processing, and outputting a second segmentation result; inputting the unmarked ultrasonic video data into the student model and the video segmentation network model for processing, and outputting a third segmentation result; training a student model on the basis of the first segmentation result, the second segmentation result, the third segmentation result and an obtained manual marking segmentation result to obtain a trained student model; wherein the trained student model is used for processing newly input ultrasonic video data, and outputting a lesion area segmentation result.

Description

Technical Field

[0001] This invention relates to the field of video segmentation technology, and in particular to a semi-supervised ultrasound video lesion segmentation method and system. Background Technology

[0002] In clinical medicine, the use of computer technology to automatically and accurately segment lesion areas in ultrasound videos can effectively improve the accuracy of early screening, precise diagnosis, and prognostic prediction of diseases based on ultrasound video imaging. Ultrasound video lesion segmentation combines semantic information of each frame with tumor shape constraints between frames to classify each pixel in the video frame. Compared to static ultrasound image lesion segmentation, ultrasound video lesion segmentation can capture the dynamic information of lesions changing over time, providing doctors with more comprehensive and accurate lesion identification and segmentation results, better meeting the needs of precise clinical diagnosis.

[0003] However, existing ultrasound video lesion segmentation methods require a large amount of labeled data during model training. Ultrasound video annotation relies on radiologists manually completing this qualitative visual annotation based on image information in video frames, combined with physiological and anatomical knowledge. This method is time-consuming, inefficient, tedious, and somewhat subjective, resulting in limited labeled data and failing to meet the needs of large-scale data annotation. Furthermore, the presence of speckle noise, hypoechoic areas, and blurred boundaries between lesions and surrounding tissues in ultrasound video frames makes segmentation of lesion edge regions challenging for existing methods. In addition, to improve lesion region segmentation accuracy, existing ultrasound video segmentation methods typically rely on Transformer networks to capture correlation information between consecutive frames, increasing model training costs and inference time, which does not meet the real-time segmentation requirements of clinical applications.

[0004] To overcome these shortcomings, this application proposes a semi-supervised ultrasound video lesion segmentation method and system. Summary of the Invention

[0005] The purpose of this application is to provide a semi-supervised ultrasound video lesion segmentation method and system, which aims to solve the above-mentioned problems.

[0006] To achieve the above objectives, this application provides the following technical solution:

[0007] This application provides a semi-supervised ultrasound video lesion segmentation method, including:

[0008] Acquire ultrasound video data, wherein the ultrasound video data includes labeled ultrasound video data and unlabeled ultrasound video data;

[0009] The labeled ultrasound video data is input into the student model for processing, and the first segmentation result is output.

[0010] The unlabeled ultrasound video data is input into the student model and the teacher model for processing, and the second segmentation result is output.

[0011] The label-free ultrasound video data is input into the student model and the video segmentation network model for processing, and a third segmentation result is output.

[0012] Based on the first segmentation result, the second segmentation result, the third segmentation result, and the obtained manually labeled segmentation result, the student model is trained to obtain a trained student model; wherein, the trained student model is used to process newly input ultrasound video data and output lesion region segmentation results.

[0013] Furthermore, the step of inputting the labeled ultrasound video data into the student model for processing and outputting the first segmentation result specifically includes the following steps:

[0014] The student model uses Vivim as the base network and is optimized through supervised learning.

[0015] Furthermore, the formula for calculating the supervised loss is as follows:

[0016]

[0017] Where Ls is the supervision loss, L seg For cross-entropy loss and Dice loss, X i For labeled ultrasound video data, Y i For X i The corresponding label values, f1 represents the student model, N is the number of training samples of labeled ultrasound video data, and θ s These are the parameters of the student model.

[0018] Furthermore, the step of inputting the label-free ultrasound video data into the student model and teacher model for processing and outputting the second segmentation result specifically includes the following steps:

[0019] The teacher model uses Vivim as its base network and is a self-ensemble of the average values ​​of the student model.

[0020] The unlabeled data is input into the student model and the teacher model for optimization, and a second segmentation result is output.

[0021] Furthermore, the teacher model is optimized using unsupervised loss, where the formula for calculating unsupervised loss is:

[0022] Lu = L cps1 +L cps2

[0023]

[0024] Where Lu is the unsupervised loss, L seg For cross-entropy loss and Dice loss, X j The data consists of unlabeled ultrasound video data, f1 is the student model, f2 is the teacher model, f3 is the SAM model, and M is the number of training samples for the unlabeled video data.

[0025] Furthermore, the step of inputting the label-free ultrasound video data into the student model and the video segmentation network model for processing and outputting the third segmentation result specifically includes the following steps:

[0026] The video segmentation network model is a SAM model. The unlabeled ultrasound video data is input into the SAM model, and pseudo-labels are output. The SAM model is optimized using unsupervised loss.

[0027] Furthermore, the SAM model includes a spatiotemporal enhancement adapter, which is used to enhance the feature representation of the SAM model for unlabeled ultrasound video data;

[0028] Unlabeled ultrasound video data is input into the spatiotemporal enhancement adapter, and feature dimension reduction is performed through a multilayer perceptron.

[0029] After downsampling the features of unlabeled ultrasound video data, the features in the temporal and spatial dimensions are enhanced and aggregated through temporal and spatial enhancement branches to obtain the spatiotemporally enhanced feature representation.

[0030] This application provides a semi-supervised ultrasound video lesion segmentation system, comprising:

[0031] Acquisition module: Acquires ultrasound video data, wherein the ultrasound video data includes labeled ultrasound video data and unlabeled ultrasound video data;

[0032] Processing module: Inputs the labeled ultrasound video data into the student model for processing and outputs the first segmentation result; inputs the unlabeled ultrasound video data into the student model and the teacher model for processing and outputs the second segmentation result; inputs the unlabeled ultrasound video data into the student model and the video segmentation network model for processing and outputs the third segmentation result.

[0033] Output module: Based on the first segmentation result, the second segmentation result, the third segmentation result, and the acquired manually labeled segmentation result, the student model is trained to obtain a trained student model; wherein, the trained student model is used to process newly input ultrasound video data and output lesion region segmentation results.

[0034] This application provides an apparatus comprising a processor and a memory coupled to the processor, wherein the memory stores program instructions for implementing a semi-supervised ultrasound video lesion segmentation method; the processor is configured to execute the program instructions stored in the memory to implement a semi-supervised ultrasound video lesion segmentation method.

[0035] This application provides a storage medium storing processor-executable program instructions for executing a semi-supervised ultrasound video lesion segmentation method.

[0036] This application provides a semi-supervised ultrasound video lesion segmentation method and system, which has the following advantages:

[0037] The spatiotemporal enhancement adapter proposed in this application adapts the SAM model to label-free ultrasound video data. With minor parameter adjustments, the SAM model can effectively utilize the label-free video data to generate more reliable pseudo-labels. Furthermore, it combines with the video segmentation network Vivim to fully utilize the spatiotemporal correlation information between video frames. Unlike static medical image lesion segmentation, this application can effectively capture the dynamic changes of lesions, saving model training costs while improving the model's segmentation accuracy. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating a semi-supervised ultrasound video lesion segmentation method according to Embodiment 1 of this application;

[0039] Figure 2 This is a schematic diagram of the framework of a semi-supervised ultrasound video lesion segmentation method according to Embodiment 1 of this application;

[0040] Figure 3 This is a schematic diagram of the spatiotemporal enhancement adapter according to Embodiment 1 of this application;

[0041] Figure 4 This is a schematic diagram of a semi-supervised ultrasound video lesion segmentation system according to Embodiment 2 of this application;

[0042] Figure 5 This is a schematic diagram of the device structure in Embodiment 3 of this application;

[0043] Figure 6 This is a schematic diagram of the storage medium structure of Embodiment 4 of this application. Detailed Implementation

[0044] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0045] The following analysis, based on relevant technologies, examines existing solutions.

[0046] In practical applications, traditional fully supervised image segmentation methods rely on large amounts of labeled data. However, the annotation of medical images is typically done by professional radiologists, which is time-consuming and costly. Therefore, semi-supervised learning techniques have become an important research direction in recent years to address the problem of insufficient labeled data. By combining a small amount of labeled data with a large amount of unlabeled data, semi-supervised learning not only reduces annotation costs but also effectively improves the model's generalization ability and segmentation accuracy.

[0047] The most similar method to this application is SemiSAM (Exploring SAM for Enhancing Semi-Supervised Medical Image Segmentation with Extremely Limited Annotations), which aims to improve the accuracy of semi-supervised medical image segmentation under conditions of limited labeled data. Specifically, when the amount of labeled data is extremely limited, the model cannot learn sufficient domain knowledge, making it difficult to segment complex lesion structures. Furthermore, semi-supervised segmentation frameworks based on consistency learning often lead to similar erroneous information for most samples, thus misleading the learning of student models.

[0048] SemiSAM employs a semi-supervised learning framework based on consistency loss. It first trains a student model using a small amount of labeled data and then generates pseudo-labels based on unlabeled data. In this framework, SAM is introduced as an external source of supervision. The student model's output serves as input cues to SAM, generating corresponding pseudo-labels. Subsequently, the prediction consistency loss between the student model's output and the pseudo-labels generated by SAM is calculated as additional supervision, further enhancing the student model's learning process.

[0049] While SemiSAM has achieved some improvement in segmentation accuracy, it is mainly applied to static medical image segmentation and lacks in-depth exploration of dynamic video data, failing to fully utilize the spatiotemporal correlation between frames. Furthermore, its pseudo-label generation process is relatively complex, increasing inference time and not meeting the needs of real-time segmentation in clinical practice. Therefore, this invention investigates how to achieve real-time and accurate segmentation of lesions in ultrasound videos using semi-supervised learning techniques with limited labeled data.

[0050] [1] Y. Zhang, Y. Cheng, and Y. Qi, "Semisam: Exploring sam for enhancing semi-supervised medical image segmentation with extremely limited annotations," arXiv preprint arXiv:2312.06316, 2023.

[0051] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0052] Example 1

[0053] Please see Figure 1 This is a flowchart illustrating a semi-supervised ultrasound video lesion segmentation method according to Embodiment 1 of this application; the steps include:

[0054] S1: Acquire ultrasound video data, wherein the ultrasound video data includes labeled ultrasound video data and unlabeled ultrasound video data.

[0055] S2: Input the labeled ultrasound video data into the student model for processing and output the first segmentation result.

[0056] In this embodiment, the student model uses Vivim as the base network and is optimized through supervised learning. The formula for calculating the supervised loss is:

[0057]

[0058] Where Ls is the supervision loss, L seg For cross-entropy loss and Dice loss, X i For labeled ultrasound video data, Y i For X i The corresponding label values, f1 represents the student model, N is the number of training samples of labeled ultrasound video data, and θ s These are the parameters of the student model.

[0059] S3: Input the unlabeled ultrasound video data into the student model and teacher model for processing, and output the second segmentation result.

[0060] In this embodiment, the teacher model uses Vivim as its base network and is a self-ensemble of the student model's average values. The unlabeled data is input into the student model and the teacher model for combined optimization, outputting a second segmentation result.

[0061] The teacher model is optimized using unsupervised loss, where the formula for calculating unsupervised loss is:

[0062] Lu = L cps1 +L cps2

[0063]

[0064] Where Lu is the unsupervised loss, L seg For cross-entropy loss and Dice loss, X j The data consists of unlabeled ultrasound video data, f1 is the student model, f2 is the teacher model, f3 is the SAM model, and M is the number of training samples for the unlabeled video data.

[0065] The teacher and student models have the same network structure, but different weights are used for initialization to handle network perturbations. The total loss function consists of supervised loss and unsupervised loss, as shown in the following formula:

[0066] L=Ls+λLu

[0067] Where Ls is the supervised loss, Lu is the unsupervised loss, and λ is the regularization parameter used to balance the supervised and unsupervised learning losses.

[0068] S4: Input the unlabeled ultrasound video data into the student model and the video segmentation network model for processing, and output the third segmentation result.

[0069] In this embodiment, the video segmentation network model is a SAM model. The unlabeled ultrasound video data is input into the SAM model, and pseudo-labels are output. The SAM model is optimized using unsupervised loss.

[0070] Please see Figure 3 This is a schematic diagram of the spatiotemporal enhancement adapter according to Embodiment 1 of this application. The SAM model also includes a spatiotemporal enhancement adapter, which is used to enhance the feature representation of unlabeled ultrasound video data by the SAM model. During training, only the adapter's parameters are adjusted. The spatiotemporal enhancement adapter is placed after the multi-head attention mechanism in the SAM model. The spatiotemporal enhancement adapter first reduces the feature dimension of the unlabeled ultrasound video data through a multilayer perceptron. This operation reduces the computational load of subsequent algorithms, improves the training and inference speed of the video segmentation network model, and allows the video segmentation network model to focus more on feature representations that are beneficial for lesion region segmentation.

[0071] Specifically, the features obtained by downsampling unlabeled ultrasound video data are then augmented in both the temporal and spatial dimensions through parallel temporal and spatial augmentation branches. The temporal augmentation branch applies three different 3D convolutions to obtain the query, key, and value from the downsampled features. A dot product is then performed on the query and key to obtain a similarity matrix between the two feature vectors. This similarity matrix is ​​then normalized to ensure that the output weight matrix conforms to a probability distribution. Finally, a dot product is performed between the normalized weight matrix and the value to obtain the enhanced feature representation in the spatial dimension.

[0072] To prevent overemphasis on spatial feature representation while neglecting other information, this branch concludes with a residual connection query to further supplement lost details. The temporal augmentation branch aims to enable the video segmentation network model to better model the complex temporal relationships in ultrasound videos.

[0073] Specifically, the temporal augmentation branch decomposes the convolutional kernel of each frame into a base weight and a calibration weight. The base weight is shared by all video frames, while the calibration weight is adaptively generated based on the feature representation of the input video frame. The generation of the calibration weight takes into account not only the information of the current frame but also the temporal context information of the current frame, which is decomposed into temporal and spatial temporal context information.

[0074] To ensure efficiency in this generation process, calibration weights are generated based on temporal and spatial contextual information. The temporal augmentation branch first obtains efficient frame features through a series of convolutional operations. Then, the frame features are multiplied by the base weights to obtain the convolution kernel parameters for each frame. The calibrated convolution kernel is then convolved with the input features to obtain feature augmentation in the temporal dimension. Finally, features in the temporal and spatial dimensions are aggregated using a method similar to that of the spatial augmentation branch to obtain a spatiotemporally augmented feature representation.

[0075] S5: Based on the first segmentation result, the second segmentation result, the third segmentation result, and the obtained artificially labeled segmentation result, the student model is trained to obtain a trained student model; wherein, the trained student model is used to process newly input ultrasound video data and output lesion region segmentation results.

[0076] In summary, this embodiment 1 processes the input ultrasound video data to obtain lesion region segmentation results, thereby reducing the dependence of the video segmentation network model on labeled data during training and improving the speed and accuracy of computer-aided ultrasound video lesion segmentation. Unlike the traditional SemiSAM method, this application adds a spatiotemporal enhancement adapter to the SAM model. By fine-tuning the adapter parameters, the feature representation of the SAM model on unlabeled ultrasound video data is enhanced, resulting in more reliable pseudo-labels. Simultaneously, to better segment lesion regions in ultrasound video data and enhance the video segmentation network model's ability to represent the features of the target segmentation region and surrounding tissues, this application uses Vivim, which has lower computational complexity, as the base network for both the student and teacher models. Compared to traditional static image segmentation networks, Vivim can fully utilize the temporal dimension information in the video data, improving the segmentation accuracy of lesion regions.

[0077] Example 2

[0078] Please see Figure 4 This is a schematic diagram of a semi-supervised ultrasound video lesion segmentation system according to Embodiment 2 of this application; the specific content includes:

[0079] Acquisition module: Acquires ultrasound video data, wherein the ultrasound video data includes labeled ultrasound video data and unlabeled ultrasound video data;

[0080] Processing module: Inputs the labeled ultrasound video data into the student model for processing and outputs the first segmentation result; inputs the unlabeled ultrasound video data into the student model and the teacher model for processing and outputs the second segmentation result; inputs the unlabeled ultrasound video data into the student model and the video segmentation network model for processing and outputs the third segmentation result.

[0081] Output module: Based on the first segmentation result, the second segmentation result, the third segmentation result, and the acquired manually labeled segmentation result, the student model is trained to obtain a trained student model; wherein, the trained student model is used to process newly input ultrasound video data and output lesion region segmentation results.

[0082] In this embodiment, experiments were conducted on this application. Clinical breast nodule ultrasound video data with a resolution of 300×200 was used, including 144 sets of labeled ultrasound video data (9770 frames) and 288 sets of unlabeled ultrasound video data (12390 frames). The invention was then tested on 45 sets of 1978 frames of data. For quantitative comparison, five convincing indicators were used: Jaccard, Dice, Precision, Recall, and frame rate. The experimental results are shown below:

[0083]

[0084]

[0085] As shown in the table, this application outperforms existing methods in both segmentation accuracy and speed. It not only offers higher segmentation accuracy but also faster segmentation speed, meeting the real-time segmentation needs in clinical practice.

[0086] Example 3

[0087] Please see Figure 5 This is a schematic diagram of the device structure in Embodiment 3 of this application. The device 50 includes a processor 51 and a memory 52 coupled to the processor 51.

[0088] The memory 52 stores program instructions for implementing the above-described semi-supervised ultrasound video lesion segmentation method.

[0089] The processor 51 is used to execute program instructions stored in the memory 52 to implement a semi-supervised ultrasound video lesion segmentation.

[0090] The processor 51 can also be referred to as a CPU (Central Processing Unit).

[0091] Processor 51 may be an integrated circuit chip with signal processing capabilities. Processor 51 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor may be a microprocessor or any conventional processor.

[0092] Example 4

[0093] Please see Figure 6This is a schematic diagram of the storage medium in Embodiment 4 of this application. The storage medium in this embodiment stores a program file 61 capable of implementing all the above methods. This program file 61 can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or devices such as computers, servers, mobile phones, and tablets.

[0094] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0095] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0096] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.

[0097] Of course, the present invention may have many other embodiments. Based on this embodiment, other embodiments obtained by those skilled in the art without any creative effort are all within the scope of protection of the present invention.

Claims

1. A semi-supervised ultrasound video lesion segmentation method, characterized in that, include: Acquire ultrasound video data, wherein the ultrasound video data includes labeled ultrasound video data and unlabeled ultrasound video data; The labeled ultrasound video data is input into the student model for processing, and the first segmentation result is output. The unlabeled ultrasound video data is input into the student model and the teacher model for processing, and a second segmentation result is output. The label-free ultrasound video data is input into the student model and the video segmentation network model for processing, and a third segmentation result is output. Based on the first segmentation result, the second segmentation result, the third segmentation result, and the obtained manually labeled segmentation result, the student model is trained to obtain a trained student model; wherein, the trained student model is used to process newly input ultrasound video data and output lesion region segmentation results.

2. The semi-supervised ultrasound video lesion segmentation method according to claim 1, characterized in that, The step of inputting the labeled ultrasound video data into the student model for processing and outputting the first segmentation result specifically includes the following steps: The student model uses Vivim as the base network and is optimized through supervised learning.

3. The semi-supervised ultrasound video lesion segmentation method according to claim 2, characterized in that, The formula for calculating the supervised loss is as follows: Where Ls is the supervision loss, L seg For cross-entropy loss and Dice loss, X i For labeled ultrasound video data, Y i For X i The corresponding label values, f1 represents the student model, N is the number of training samples of labeled ultrasound video data, and θ s These are the parameters of the student model.

4. The semi-supervised ultrasound video lesion segmentation method according to claim 1, characterized in that, The step of inputting the label-free ultrasound video data into the student model and teacher model for processing and outputting the second segmentation result specifically includes the following steps: The teacher model uses Vivim as its base network and is a self-ensemble of the average values ​​of the student model. The unlabeled data is input into the student model and the teacher model for optimization, and a second segmentation result is output.

5. The semi-supervised ultrasound video lesion segmentation method according to claim 1, characterized in that, The teacher model is optimized using unsupervised loss, where the formula for calculating unsupervised loss is: Lu=L cps1 +L cps2 Where Lu is the unsupervised loss, L seg For cross-entropy loss and Dice loss, X j The data consists of unlabeled ultrasound video data, f1 is the student model, f2 is the teacher model, f3 is the SAM model, and M is the number of training samples for the unlabeled video data.

6. The semi-supervised ultrasound video lesion segmentation method according to claim 1, characterized in that, The step of inputting the label-free ultrasound video data into the student model and the video segmentation network model for processing and outputting the third segmentation result specifically includes the following steps: The video segmentation network model is a SAM model. The unlabeled ultrasound video data is input into the SAM model, and pseudo-labels are output. The SAM model is optimized using unsupervised loss.

7. The semi-supervised ultrasound video lesion segmentation method according to claim 6, characterized in that, The SAM model includes a spatiotemporal enhancement adapter, which is used to enhance the feature representation of the SAM model for unlabeled ultrasound video data. Unlabeled ultrasound video data is input into the spatiotemporal enhancement adapter, and feature dimension reduction is performed through a multilayer perceptron. After downsampling the features of unlabeled ultrasound video data, the features in the temporal and spatial dimensions are enhanced and aggregated through temporal and spatial enhancement branches to obtain the spatiotemporally enhanced feature representation.

8. A system for a semi-supervised ultrasound video lesion segmentation method according to claim 1, characterized in that, include: Acquisition module: Acquires ultrasound video data, wherein the ultrasound video data includes labeled ultrasound video data and unlabeled ultrasound video data; Processing module: Inputs the labeled ultrasound video data into the student model for processing and outputs the first segmentation result; inputs the unlabeled ultrasound video data into the student model and the teacher model for processing and outputs the second segmentation result; inputs the unlabeled ultrasound video data into the student model and the video segmentation network model for processing and outputs the third segmentation result. Output module: Based on the first segmentation result, the second segmentation result, the third segmentation result, and the acquired manually labeled segmentation result, the student model is trained to obtain a trained student model; wherein, the trained student model is used to process newly input ultrasound video data and output lesion region segmentation results.

9. A device, characterized in that, The device includes a processor and a memory coupled to the processor, wherein the memory stores program instructions for implementing a semi-supervised ultrasound video lesion segmentation method according to any one of claims 1-7; the processor is used to execute the program instructions stored in the memory to implement a semi-supervised ultrasound video lesion segmentation method.

10. A storage medium, characterized in that, The device stores processor-executable program instructions for performing a semi-supervised ultrasound video lesion segmentation method according to any one of claims 1-7.

Images