Biological data anomaly detection method based on semantic information fusion

Abstract

The invention discloses a biological data anomaly detection method based on semantic information fusion, and relates to the technical field of computer processing, and the method comprises the following steps: S01, constructing biological entity nodes and relation edges through a depth map neural network, generating an adjacency matrix and a node attribute matrix, aggregating node neighbor information by using a message passing mechanism, and generating a node attribute matrix; a biological network architecture model is obtained; s02, constructing a semantic fusion space based on a large model, analyzing deep semantic information and node association information of nodes, and synthesizing a few abnormal node candidate sets; and S03, adding the composite node candidate sets into the original data set one by one through an edge generator, and inputting the composite node candidate sets into a reinforcement learning module for screening. According to the method, classification and risk assessment of biological data are optimized by utilizing large-model semantic analysis and biological network construction, so that abnormal biological data can be quickly identified through multi-source data fusion and deep learning analysis in the early stage of virus transmission, and accurate detection and intelligent early warning are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of computer processing, and particularly to a method for detecting anomalies in biological data based on semantic information fusion. Background Art

[0002] In the global public health field, traditional biological data analysis methods (such as epidemiological investigations and statistical models) have significant limitations. For example, in the early stage of the COVID-19 outbreak, traditional methods were difficult to identify virus mutations and transmission risks in a timely manner due to their reliance on lagging data and static models. In addition, existing graph neural networks (GNNs) perform poorly in scenarios with class imbalance, and the limited number of minority abnormal nodes restricts feature propagation, causing the model to be prone to bias towards the majority class.

[0003] In the prior art, random data augmentation methods may introduce noise, while manual adjustment of learning strategies relies on experience and is inefficient. Therefore, there is an urgent need for a method for detecting anomalies in biological data that can fuse semantic information, dynamically optimize data distribution, and improve the generalization ability of the model. Summary of the Invention

[0004] In view of the above technical problems, the technical solution adopted by the present invention is a method for detecting anomalies in biological data based on semantic information fusion, and the method includes the following steps:

[0005] S01. Construct biological entity nodes and relationship edges through a deep graph neural network, generate an adjacency matrix and a node attribute matrix, and aggregate node neighbor information using a message passing mechanism to obtain a biological network architecture model;

[0006] S02. Construct a semantic fusion space based on a large model, analyze the deep semantic information and node association information of nodes, and synthesize a candidate set of minority abnormal nodes;

[0007] S03. Add the synthesized node candidate set to the original data set one by one through an edge generator, and input it into a reinforcement learning module for screening to obtain an optimized synthesized node candidate set;

[0008] S04. Adjust the learning order according to the node training loss value, organize the obtained optimized synthesized node candidate set into a training set, and train the biological network architecture model to obtain an iterative biological network architecture model to replace the biological network architecture model in step S01.

[0009] Preferably, in step S01, the biological entity includes a virus or a gene.

[0010] Preferably, obtaining the biological network architecture model in step S01 includes:

[0011] S11. Construct a biological network G=(V, E) with biological entities as nodes and the relationships between entities as undirected edges, where V represents the set of nodes in the network, the nodes represent biological entities, E represents the set of edges in the network, and the edges represent the relationships between entities;

[0012] S12. The adjacency matrix A represents the connection relationship between nodes, and the node attribute matrix X and the label matrix Y represent node features and categories respectively;

[0013] S13. Aggregate the neighbor information of nodes based on the message passing mechanism, and the formula is:

[0014]

[0015] Where, A is the adjacency matrix, I is the identity matrix, D is the degree matrix, l is the number of layers, H is the feature of each layer. For the input layer, H is X, and σ is the non-linear activation function.

[0016] Preferably, in step S02, synthesizing the minority abnormal node candidate set includes the following formula:

[0017]

[0018] Where, δ is a random variable uniformly distributed in the range [0, 1], nn(v) is the nearest neighbor node of the same class, and the synthesized node inherits the source node label.

[0019] Preferably, in step S03, the screening process of the reinforcement learning module includes state, action, transition, and reward, and uses policy gradient training to update the parameters of the policy network. The steps include:

[0020] State: State S t Consists of the embeddings of the nodes in the training set V t and the embeddings of the unlabeled nodes U t . Then, the sum of the node embeddings in V t is used to represent the information of V t U0 is the first node V0=V in V c ; L ;

[0021] Action: Action a t Determines whether the current unlabeled node U C in V t should be included in the current training set V t at time t. a t ∈{0, 1}, where a t =1 means selecting the node U t to supplement the imbalanced training set, while when at = 0 indicates U t Indicates not applicable. a t Determined by the policy function π θ With S t As the input of the state, the multi-layer perceptron MLP:

[0022] a t = P(a t | s t ) = π θ (s t ) = MLP θ (s t );

[0023] Transition: After taking a t , the state of the environment becomes S t+1 , and the state is composed of V t and U t :

[0024]

[0025] When The agent fully traverses the candidate set V c Once, the transition terminates;

[0026] Reward: Evaluate the behavior in a certain state. If u t Can improve the performance of the classifier, a positive reward is assigned, otherwise a negative reward is assigned,

[0027]

[0028] Based on {V t ∪ u t} Train a classifier and evaluate its accuracy on a small balanced validation set.

[0029] Preferably, there is one multi-layer perceptron MLP in the actions.

[0030] Preferably, the biological node classification process of adjusting the learning order according to the node training loss value in step S04 in combination with loss-aware learning includes: using the magnitude of the loss function of each node during the training process as an index of the perception difficulty of the model to learn the node, ranking the nodes according to the magnitude of the loss function, and then training them in sequence according to the ranking order.

[0031] The present invention has at least the following beneficial effects:

[0032] 1. Through large model semantic analysis and network model construction, key features in biological data can be accurately extracted, significantly improving the accuracy of biological data anomaly detection, especially excelling in identifying risk points related to biosafety;

[0033] 2. By precisely identifying potential risk points and classifying them, it can better support biosafety monitoring, promptly discover potential biosafety hazards, and enhance the effectiveness of biosafety management;

[0034] 3. Optimize the data analysis and processing process, reduce computational complexity, improve the operation efficiency of the model, make the processing process of biological data more efficient, and adapt to the analysis requirements of large-scale data;

[0035] 4. Through network node optimization and adjustment, the adaptability of the model to different types of biological data is enhanced, enabling it to maintain high accuracy and stability when facing diverse and highly biodiverse data. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0037] Figure 1 It is a flowchart framework of a method for detecting anomalies in biological data based on semantic information fusion provided in Embodiment 1 of the present invention;

[0038] Figure 2 It is a flowchart of biological data mining and classification based on semantic information fusion provided in Embodiment 1 of the present invention;

[0039] Figure 3 It is a flowchart of reinforcement learning provided in Embodiment 1 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0041] It should be noted that the terms "first", "second", etc. in the description, claims and above-mentioned drawings of the present invention are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention described here can be implemented in an order other than those illustrated or described here. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or server comprising a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0042] Embodiment 1

[0043] This embodiment provides a method for detecting abnormal biological data based on semantic information fusion. The method includes the following steps, as Figure 1 shown:

[0044] S01. Construct biological entity nodes and relationship edges through a deep graph neural network, generate an adjacency matrix and a node attribute matrix, and use the message passing mechanism to aggregate node neighbor information to obtain a biological network architecture model;

[0045] Specifically, in step S01 of the above embodiment, the biological entities include viruses or genes.

[0046] Further, obtaining the biological network architecture model in step S01 includes:

[0047] S11. Construct a biological network G=(V, E) with biological entities as nodes and the relationships between entities as undirected edges, where V represents the set of nodes in the network, the nodes represent biological entities, E represents the set of edges in the network, and the edges represent the relationships between entities;

[0048] S12. The adjacency matrix A represents the connection relationship between nodes, and the node attribute matrix X and the label matrix Y represent node features and categories respectively;

[0049] S13. Aggregate the neighbor information of nodes based on the message passing mechanism. The formula is:

[0050]

[0051] where, A is the adjacency matrix, I is the identity matrix, D is the degree matrix, l is the number of layers, H is the feature of each layer. For the input layer, H is X, and σ is the non-linear activation function.

[0052] Among the above technologies, in the semi-supervised node classification task based on graph neural networks (GNNs), the class imbalance problem will seriously affect the learning effect of the model. Existing studies usually assume that the number of samples in each category is relatively balanced, but in practical applications, the number of nodes in the minority category (the number is lower than the average) is often far less than that in the majority category, such as fraud accounts in transaction fraud detection networks or unpopular research fields in academic citation networks. In this case, the training process of GNN is easily affected by data distribution bias, resulting in the model being more inclined to predict the majority category and having weak recognition ability for the minority category. Since GNN relies on the information of neighboring nodes for feature propagation, when minority category nodes are scarce, their category features are difficult to propagate effectively, further exacerbating the imbalance of classification performance. In order to solve this problem, synthesizing a minority of abnormal nodes becomes a necessary strategy.

[0053] S02: Build a semantic fusion space based on the big model, analyze the deep semantic information and node association information of the nodes, and synthesize a small number of abnormal node candidate sets;

[0054] Specifically, in step S02, a small number of abnormal node candidate sets are synthesized, including the following formula:

[0055]

[0056] Among them, δ is a random variable uniformly distributed in the range [0, 1], nn(v) is the nearest neighbor node of the same type, and the synthesized node inherits the label of the source node.

[0057] In the above technology, by combining the large model to analyze the deep semantic information and node association information of the nodes in the semantic fusion space, a candidate set of a few abnormal nodes is generated. In this way, the existing graph structure and text information can be fully utilized to extract the characteristic patterns of minority category nodes, thereby ensuring the rationality and category consistency of the synthesized nodes in the process of data enhancement. Compared with traditional data supplementation methods, this method can not only generate minority category nodes more accurately, but also optimize their distribution in the graph network, making them more consistent with the topological structure of real data, thereby effectively improving the classification performance of GNN in category imbalance scenarios.

[0058] S03, adding the synthetic node candidate set to the original data set one by one through the edge generator, and inputting it into the reinforcement learning module for screening to obtain an optimized synthetic node candidate set;

[0059] Specifically, the screening process of the reinforcement learning module in step S03 includes states, actions, transitions, rewards, and using policy gradient training to update the parameters of the policy network, and the steps include:

[0060] Status: Status S t From the training set node V tEmbedded and unlabeled node U t For the embedding composition, it is used in V t The sum of the node embeddings in V is used to represent V t 's information U0 is the first node in V c V0 = V L ;

[0061] Action: Action a t Determine at time t whether the currently unlabeled node U in V C should be included in the current training set V t where a t ∈ {0, 1}, and a t = 1 means selecting node U t to supplement the imbalanced training set, while a t = 0 means U t is not applicable. a t is determined by the policy function π t with the state S θ as the input. The multi-layer perceptron MLP (there is one multi-layer perceptron MLP in the action): t a

[0062] a t = P(a t | s t ) = π θ (s t ) = MLP θ (s t );

[0063] Transition: After taking a t , the state of the environment becomes S t+1 , and the state is composed of V t and U t :

[0064]

[0065] When the agent completely traverses the candidate set V c once, the transition terminates;

[0066] Reward: Evaluate the behavior in a certain state. If u t can improve the performance of the classifier, assign a positive reward, otherwise assign a negative reward

[0067]

[0068] Train a classifier based on {V t ∪ u t} and evaluate its accuracy on a small balanced validation set.

[0069] In the above technology, simple random data augmentation may introduce noise, affecting the generalization ability of the model. Therefore, a reinforcement learning mechanism can be combined to screen and optimize the synthetic nodes to ensure their rationality and effectiveness. Thus, the present invention screens the candidate set through reinforcement learning, which can not only improve the classification performance of the GNN on class-imbalanced data, but also enhance the adaptability of the model in anomaly detection and extremely data-scarce scenarios, thereby improving the robustness and reliability of the overall task.

[0070] S04. Adjust the learning order according to the node training loss value, and organize the obtained optimized synthetic node candidate set into a training set to train the biological network architecture model, so as to obtain an iterative biological network architecture model to replace the biological network architecture model in step S01.

[0071] Specifically, the process of classifying biological nodes by adjusting the learning order according to the node training loss value in step S04 and combining loss-aware learning includes: using the magnitude of the loss function of each node during the training process as an indicator of the perceived difficulty for the model to learn the node, ranking the nodes according to the magnitude of the loss function, and then training them in sequence according to the ranking order.

[0072] In the above technology, after the reinforcement learning determines the final candidate set, the data set is updated because the topological structure of the graph is intricate, making it difficult to ensure that various models can effectively evaluate the difficulty and quality of nodes only based on topological features. To address this challenge, the present invention uses the magnitude of the loss function of each node during the training process as an indicator of the perceived difficulty for the model to learn the node. First, this method is more closely integrated with the paradigm of machine learning, eliminating the need for manual adjustment or empirical adjustment to adapt to the curriculum learning model. Second, only the loss function needs to be modified without introducing additional computational overhead. To optimize the training, the nodes are ranked according to the magnitude of their loss functions, reflecting their learning difficulty. Starting from the nodes showing lower loss values, we gradually introduce nodes with higher values to ensure a balanced approach. This prevents the improper influence of noisy nodes, thereby enhancing the robustness of the model.

[0073] As can be seen from the above, through large model semantic analysis and network model construction, key features in biological data can be accurately extracted, significantly improving the accuracy of biological data anomaly detection, especially in identifying risk points related to biosafety; secondly, by accurately identifying potential risk points and classifying them, it can better support biosafety monitoring, timely discover potential biosafety hazards, and enhance the effectiveness of biosafety management; furthermore, the above embodiments optimize the data analysis and processing process, reduce the computational complexity, and improve the operation efficiency of the model, making the processing process of biological data more efficient and meeting the analysis requirements of large-scale data. Moreover, through the optimization and adjustment of network nodes, the present invention enhances the adaptability of the model to different types of biological data, enabling it to maintain a high accuracy and stability when facing diverse and highly biodiverse data.

[0074] In the first embodiment, these data are first cleaned and standardized, and key features of virus samples (such as protein structure, mutation sites, etc.) are extracted through semantic analysis technology. A virus transmission association network is constructed using a deep graph neural network (DGL). Nodes represent virus samples, and edges represent the evolutionary relationship or co-occurrence pattern between viruses. For example, if two virus samples have similar mutation characteristics, a connection will be established. The connection strength between viruses is calculated through an adjacency matrix, and the node features are optimized using a message passing mechanism (MP). In virus monitoring data, the sample size of some abnormal virus variants may be much less than that of common virus strains, making it difficult for traditional classification models to learn their features. The present invention combines large model semantic fusion analysis to generate a candidate set of a small number of abnormal virus samples based on the protein structure and mutation pattern of the virus, ensuring that these synthetic samples can truly reflect the characteristics of rare virus variants. Secondly, reinforcement learning is used to screen the synthetic abnormal virus nodes to ensure that only the most representative abnormal virus samples are retained. The reinforcement learning agent selects the virus strains most likely to pose a biosafety risk based on factors such as virus mutation patterns and clinical manifestations, and optimizes the classification performance of the model. Finally, the model optimizes the classification performance of abnormal viruses through loss-aware learning. Since some virus samples may show a large loss value (i.e., high learning difficulty) during training, the present invention adjusts the learning strategy through a loss-aware mechanism, making it easier for the model to identify these difficult-to-classify virus variants, thereby improving the robustness of biological data anomaly detection. After the detection is completed, the system generates a risk assessment report of virus variants and provides it to the disease control center or biosafety agency for epidemic prediction and response strategy formulation. For example, when an abnormal transmission characteristic or key mutation site of a certain virus strain is detected, the system will automatically mark the virus strain as high risk and recommend further experimental verification or preventive and control measures.

[0075] Embodiment 2

[0076] An embodiment of the present invention provides a non-transitory computer-readable storage medium, in which at least one instruction or at least one program segment is stored, and at least one instruction or at least one program segment is loaded and executed by a processor to implement the steps:

[0077] Construct biological entity nodes and relationship edges through a deep graph neural network, generate an adjacency matrix and a node attribute matrix, and use a message passing mechanism to aggregate node neighbor information to obtain a biological network architecture model;

[0078] Construct a semantic fusion space based on a large model, analyze the deep semantic information and node association information of the nodes, and synthesize a candidate set of a few abnormal nodes;

[0079] Add the synthesized node candidate set to the original dataset one by one through an edge generator and input it into a reinforcement learning module for screening to obtain an optimized synthesized node candidate set;

[0080] Adjust the learning order according to the node training loss value, organize the obtained optimized synthesized node candidate set into a training set, and train the biological network architecture model to obtain an iterative biological network architecture model to replace the original biological network architecture model.

[0081] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memories (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memories. Volatile memories can include random access memories (RAM) or external cache memories. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0082] Those skilled in the art can clearly understand that, for the convenience and conciseness of description, only the above-mentioned division of each functional unit and module is used as an example. In practical applications, the above functions can be allocated to different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above.

[0083] Embodiment III

[0084] An embodiment of the present invention provides an electronic device, including a processor and a memory. At least one instruction or at least one program segment is stored in the memory, and the at least one instruction or the at least one program segment is loaded and executed by the processor to implement the steps:

[0085] Construct biological entity nodes and relationship edges through a deep graph neural network, generate an adjacency matrix and a node attribute matrix, and use a message passing mechanism to aggregate node neighbor information to obtain a biological network architecture model;

[0086] Construct a semantic fusion space based on a large model, analyze the deep semantic information and node association information of nodes, and synthesize a candidate set of a few abnormal nodes;

[0087] Add the synthesized node candidate set to the original data set one by one through an edge generator and input it into a reinforcement learning module for screening to obtain an optimized synthesized node candidate set;

[0088] Adjust the learning order according to the node training loss value, organize the obtained optimized synthesized node candidate set into a training set, and train the biological network architecture model to obtain an iterative biological network architecture model to replace the original biological network architecture model.

[0089] As mentioned above, it is only a preferred embodiment of the present invention and does not impose any form of limitation on the present invention. Although the present invention has been disclosed above with a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the equivalent embodiments by using the disclosed technical content within the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A method for detecting anomalies in biological data based on semantic information fusion, characterized in that, The method includes the following steps: S01. Construct biological entity nodes and relationship edges through a deep graph neural network, generate an adjacency matrix and a node attribute matrix, and aggregate node neighbor information using a message passing mechanism to obtain a biological network architecture model; S02. Construct a semantic fusion space based on a large model, analyze the deep semantic information and node association information of nodes, and synthesize a candidate set of a few abnormal nodes; S03. Add the synthesized node candidate set to the original dataset one by one through an edge generator, and input it into a reinforcement learning module for screening to obtain an optimized synthesized node candidate set; S04. Adjust the learning order according to the node training loss value, organize the obtained optimized synthesized node candidate set into a training set, and train the biological network architecture model to obtain an iterative biological network architecture model to replace the biological network architecture model in step S01.

2. The method for abnormal detection of biological data based on semantic information fusion according to claim 1, wherein, In step S01, the biological entity includes a virus or a gene.

3. The method for abnormal detection of biological data based on semantic information fusion according to claim 1, wherein Obtaining the biological network architecture model in step S01 includes: S11. Construct a biological network G=(V, E) with biological entities as nodes and the relationships between entities as undirected edges, where V represents the set of nodes in the network, the nodes represent biological entities, E represents the set of edges in the network, and the edges represent the relationships between entities; S12. The adjacency matrix A represents the connection relationship between nodes, and the node attribute matrix X and the label matrix Y represent node features and categories respectively; S13. Aggregate the neighbor information of nodes based on the message passing mechanism, and the formula is: Among them, A is the adjacency matrix, I is the identity matrix, D is the degree matrix, l is the number of layers, H is the feature of each layer. For the input layer, H is X, and σ is the non-linear activation function.

4. A method for abnormal detection of biological data based on semantic information fusion according to claim 1, characterized in that Synthesizing the candidate set of a few abnormal nodes in step S02 includes the following formula: where δ is a random variable uniformly distributed in the range [0, 1], nn(v) is the nearest neighbor node of the same type, and the synthesized node inherits the source node label.

5. A method for abnormal detection of biological data based on semantic information fusion according to claim 1, characterized in that, The screening process of the reinforcement learning module in step S03 includes state, action, transition, and reward, and uses policy gradient training to update the parameters of the policy network. The steps include: Status: Status S t Composed of the training set nodes V t Embedding and unlabeled nodes U t Embedding, and is used to represent V t By the sum of the node embeddings in V t Information U0 is the first node in V c The node V0 = V L ; Action: Action a t Determine at time t, V C The currently unmarked node U in t Whether it should be included in the current training set V t in a t ∈{0,1}, where a t =1 means selecting node U t To supplement the imbalanced training set, while in a t =0 means U t Indicates inapplicability. a t Is determined by the policy function π θ With S t As the input of the state, the multi-layer perceptron MLP: a t = P(a t | s t ) = π θ (s t ) = MLP θ (s t ); Transition: After taking a t the state of the environment becomes S t+1 The state consists of V t and U t : When the agent has fully traversed the candidate set V c once, the conversion terminates; Reward: Evaluate the behavior in a certain state. If u t can improve the performance of the classifier, then assign a positive reward; otherwise, assign a negative reward. Train a classifier based on {V t ∪u t} and evaluate its accuracy on a small balanced validation set.

6. The method for abnormal detection of biological data based on semantic information fusion according to claim 5, characterized in that, The multi-layer perceptron MLP in the action is one.

7. A method for abnormal detection of biological data based on semantic information fusion according to claim 1, characterized in that The biological node classification process of adjusting the learning order according to the node training loss value and combining loss-aware learning in step S04 includes: using the size of the loss function of each node during the training process as an indicator of the perception difficulty of the model to learn the node, ranking the nodes according to the size of the loss function, and then training them in order according to the ranking order.

8. A non-transitory computer-readable storage medium storing at least one instruction or at least one program segment therein, characterized in that, The at least one instruction or the at least one program is loaded and executed by a processor to implement the steps of the method for detecting abnormal biological data based on semantic information fusion as described in any one of claims 1-7.

9. An electronic device, characterized in that, It includes a processor and a memory. At least one instruction or at least one program is stored in the memory. The at least one instruction or the at least one program is loaded and executed by a processor to implement the steps of the method for detecting abnormal biological data based on semantic information fusion as described in any one of claims 1-7.

Images