Method, system and storage medium for constructing brainwave state database

Abstract

The present invention provides a method, system, and storage medium for constructing a brainwave state database, including the following steps: constructing a brainwave state transition graph using a deep learning method using brainwave data; and storing the constructed brainwave state transition graph in the form of a graph database to describe possible transition paths between various brainwave states. The method for constructing a brainwave state database of the present invention does not rely on expert knowledge, but automatically identifies and extracts brainwave states directly from a large amount of brainwave data, which is more consistent with the situation of brainwave data. The present invention pre-trains a model similar to the BERT structure and an automatic brainwave data segmentation and recognition model, which has low data requirements, and all types of collected brainwave data can be used after being unified in form. Since the brainwave state database constructed by the present invention is constructed unsupervisedly using general data, it has a good auxiliary effect for specific tasks related to certain brainwave states.

Description

Technical Field

[0001] The present invention relates to the technical field of brain wave state analysis, and in particular to a method, system and storage medium for constructing a brain wave state database. Background Art

[0002] Brainwaves, also known as "electroencephalograms," are bioelectric signals generated when neurons transmit information. They are electrical signals produced by the exchange of ions during synaptic activity in pyramidal cells in the cerebral cortex. Brain cells generate corresponding bioelectricity whenever they are active, and this electricity connects them to other brain cells through their dendrites, forming networks of thought. In scientific applications, brainwaves that can be detected by instruments must be generated by the activity of a certain number of neurons. These waves are categorized by frequency into the following types: Δ waves, θ waves, α waves, β waves, and γ waves. Beta waves are further subdivided into SMR waves, βl waves (low β waves), and βh waves (high β waves), depending on the brain's state of control.

[0003] Current research on brainwaves includes analysis of brainwave states and induction. Most brainwave studies require prior knowledge of various states, requiring expert knowledge. This knowledge is often limited to fixed criteria for judging brainwaves, such as classifying them by primary frequency. However, the complexity of brainwaves related to human thought is not simply reflected by frequency. A more rational approach would be to study large amounts of brainwave data, analyzing and recording all or most of its states, and thereby guiding research on brainwave states.

[0004] BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained model in natural language processing. This model is trained using massive amounts of text data. One of its uses is as a text feature extraction tool, similar to the Word2vec model, encoding text. This significantly improves performance across various natural language processing tasks. This model effectively uses massive amounts of task-unrelated text data to create a rich encoding method. Its utilization of data resources within the field offers insights into the limited amount of target data currently found in EEG data research.

[0005] In general, there is currently little research on the technology used to construct brainwave state data. Identifying brainwave states requires the assistance of professionals and is limited to the frequency of brainwave data, but does not consider the various variations in brainwave patterns. Summary of the Invention

[0006] The present invention proposes a method for constructing a brain wave state database, which can solve the above technical problems.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] A method for constructing a brain wave state database comprises the following steps:

[0009] Use brainwave data to construct brainwave state transition graphs through deep learning methods;

[0010] The constructed brain wave state transition graph is stored in the form of a graph database to describe the transition paths between various brain wave states.

[0011] Furthermore, the steps of constructing a brainwave state transition diagram using the brainwave data through a deep learning method are as follows:

[0012] First, a BERT-like encoding model for brainwave data is trained. Based on the brainwave data encoded by this BERT-like encoding model, a model for automatically segmenting and recognizing brainwave data is trained. This model can identify each value in the brainwave data as a different number ranging from 1 to n, where the value of n is determined experimentally. Data values ​​with the same number are in the same brainwave state, and the number also represents the node number of the final graph.

[0013] The graph construction module constructs a brain wave state transition graph based on each state transition path obtained from the brain wave data.

[0014] Furthermore, during the composition process, the state numbers are regarded as nodes. If there is a transition process between states, there is an edge, otherwise there is no edge; the state transition path corresponding to each brain wave data is collected, the same number represents the same state, and the number of occurrences of each edge is counted as the weight of the edge.

[0015] Furthermore, a model capable of automatically dividing and identifying brainwave data is trained based on the brainwave data encoded by the model, and the training steps are as follows:

[0016] (1) Symbolization of brainwave data:

[0017] First, according to the data collected by the acquisition device, it is expressed as X (x1, x2, ... x i ), get the maximum value V possible for each time step of the data max and minimum value V min , the interval [V min , V max ] is evenly divided into N boxes, represented by 1 to N respectively; for the value x in the data i , if V min ≤x i≤V min +(V max -V min ) / N, then x i Recorded as 1; if V min +(V max -V min ) / N ≤x i ≤V min +2*(V max -V min ) / N, then x i Recorded as 2; and so on, encode X into a sequence represented by 1-N;

[0018] (2) After processing the brainwave data into a symbol sequence, simulate the BERT embedding method; where:

[0019] Token Embeddings is a word vector. The first word is the CLS flag, which indicates the data boundary and is used for subsequent classification tasks.

[0020] Segment Embeddings is used to distinguish input and output in sequence prediction, which refers to using the data of the previous L1 time steps to predict the data of the next L2 time steps;

[0021] Position Embeddings represents the position encoding that can be learned;

[0022] (3) Use BERT’s network structure to train improved tasks;

[0023] Directly use the BERT network structure to perform the following two training tasks:

[0024] 31) T1: Masked language model

[0025] This step is consistent with the original BERT training. During training, 15% of the tokens are randomly masked, and the final loss function only calculates the masked tokens.

[0026] 32), T2: Sentence prediction

[0027] In this step, sequence prediction is performed, using the data of the first L1 time steps in the brain wave data to predict the data of the next L2 time steps.

[0028] Furthermore, the model for automatically dividing and identifying brain wave data includes:

[0029] Self-attention module: A module implemented using the self-attention mechanism to calculate the similarity score between each data value of the encoded brainwave data. If the length of the brainwave data is L, the output of this module is an L*L matrix; for example, for the value a ij Represents the similarity score between the i-th value and the j-th value;

[0030] Graph convolution layer: Graph convolution is performed on the output of the self-attention module. The output is a graph feature vector for each point in the data. Assuming it is an L*E1 matrix, the i-th point in the original data is represented by a graph feature vector of length E1. This feature has two functions: on the one hand, it is used for state number recognition, and on the other hand, it is used for state feature transformation.

[0031] State number prediction: Based on the output of the graph convolutional layer, identify the state number corresponding to each point in the brainwave data, which is also the number in the composition;

[0032] State feature transformation: Based on the state number and the feature vector obtained by the graph convolution layer, the graph feature vector of each state number in the current processing data is obtained, and the feature corresponding to each state number is obtained through transformation;

[0033] State feature matrix: A global state feature is maintained in this model, that is, all values ​​predicted to have the same number in the data have the same or similar representation after the state feature transformation.

[0034] Furthermore, a model for automatically dividing and identifying brainwave data is trained based on the brainwave data encoded by the model, and the training process is as follows:

[0035] After the brainwave data is encoded by a BERT-like model, it first passes through the self-attention module to obtain the similarity between the value at each position in the data and the values ​​at other positions. Graph convolution is then used to extract features on the similarity matrix to obtain the graph representation features between the values. Based on the graph representation features, the number of each value in the data in the graph is predicted, and the graph representation features of the values ​​with the same number are transformed to obtain the global state features of the same numbered state. This global state feature is then compared with the globally maintained state feature matrix.

[0036] The final loss function is calculated by calculating the similarity between the numbered state features in the globally maintained state feature matrix and the numbered state features obtained in the current data, and taking the negative value as the loss function.

[0037] On the other hand, the present invention also discloses a system for constructing a brain wave state database, which is used to implement a method for constructing a brain wave state database, and is characterized by comprising the following units:

[0038] Data analysis and mapping module, used to construct brainwave state transition diagrams using a large amount of brainwave data through deep learning methods;

[0039] The brainwave state data storage module is used to store the acquired brainwave state transition graph in the form of a graph database, which is used to describe the possible transition paths between various brainwave states, facilitating various applications related to brainwave states;

[0040] Database application module, which relies on various applications implemented by the brainwave state database.

[0041] On the other hand, the present invention further discloses a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the steps of the above method.

[0042] From the above technical solution, it can be seen that the method for constructing the brainwave state database of the present invention has the following beneficial effects:

[0043] The present invention provides a method for automatically constructing a brainwave state change database, which does not rely on expert knowledge and can automatically identify and extract brainwave states directly from a large amount of brainwave data, and is more consistent with the situation of brainwave data.

[0044] The present invention pre-trains a model similar to the BERT structure and an automatic brainwave data segmentation and recognition model, which has low data requirements and can be used after all types of collected brainwave data are unified in form.

[0045] Since the brainwave state database constructed by the present invention is constructed unsupervisedly through general data, it has a good auxiliary effect on some specific tasks related to brainwave states. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Schematic diagram of the method framework structure of an embodiment of the present invention;

[0047] Figure 2 Schematic diagram of the process of constructing a brain wave state database according to an embodiment of the present invention;

[0048] Figure 3 This is an example diagram of constructing a brain wave state database according to an embodiment of the present invention;

[0049] Figure 4 It is a schematic diagram of the data structure of an embodiment of the present invention;

[0050] Figure 5This is the embedding process of symbolized data in an embodiment of the present invention;

[0051] Figure 6 This is a schematic diagram of the network structure of the automatic segmentation and recognition model of brain wave data according to an embodiment of the present invention;

[0052] Figure 7 Schematic diagram of the method for using the brain wave database according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0054] like Figure 1 As shown, the method for constructing the brainwave state database described in this embodiment includes a data analysis and mapping module, a brainwave state database, and a database application system. The main functions and implementation methods of each part are as follows:

[0055] Data analysis and mapping module: The main function is to use a large amount of brain wave data to construct a brain wave state transition diagram through deep learning methods.

[0056] Brainwave state database: stores the brainwave state transition graph obtained in (1) in the form of a graph database. This database mainly describes the possible transition paths between various brainwave states, facilitating various applications related to brainwave states.

[0057] Database application system: This module mainly relies on the brain wave state database to implement various applications.

[0058] The following are specific instructions:

[0059] 1. Construction of brainwave state database

[0060] The construction of the brainwave state database is mainly based on deep learning methods. A large amount of brainwave data is used to train a model that can automatically convert brainwave data into a series of brainwave state transition paths. The brainwave state transition paths corresponding to all the data are integrated to construct a brainwave state transition diagram.

[0061] The specific process is as follows Figure 2As shown, the basic idea is to first train a BERT-like encoding model for EEG data. Based on the EEG data encoded by this BERT-like encoding model, a model for automatic segmentation and recognition of EEG data is trained. This automatic segmentation and recognition model identifies each EEG data value as a unique number ranging from 1 to n, where n is determined experimentally. Data values ​​with the same number represent the same EEG state, and the number also represents the node number of the resulting graph. The graph construction module constructs an EEG state transition graph based on each state transition path obtained from the EEG data.

[0062] An example is shown in Figure 3. In the figure, Data1 (V1, V2…V16) represents an EEG signal. This data is encoded using a BERT-like encoding model, which converts the one-dimensional signal data into a two-dimensional matrix. Each value, Vi, is then transformed. After encoding, the encoded data is passed through the EEG data automatic segmentation and recognition model, where each value is assigned a number: 1112222223333444. This means that V1 in Data1 is numbered 1, V2 is numbered 1, and V3 is numbered 1. V1, V2, and V3 are in EEG state 1, while V4 through V9 are in EEG state 2. Similarly, the corresponding state transformation path for Data1 is 1->2->3->4.

[0063] For the graph construction process, the state numbers are regarded as nodes. If there is a transition process between states, there is an edge, otherwise there is no edge. The state transition path corresponding to each brain wave data is collected. The same number represents the same state. At the same time, the number of occurrences of each edge is counted as the edge weight. Assuming that the transition path corresponding to Data2 is: 2->5->4->7, and the transition path corresponding to Data3 is 3->6->7, then the constructed graph is as follows Figure 4 As shown;

[0064] Figure 4 In the graph constructed based on the three data, since each edge appears only once, their weights are all 1. After the state transition paths corresponding to all the data are synthesized through the above steps, the resulting graph can be stored in the graph database, and the brain wave state database is constructed.

[0065] The following details the training process of the BERT-like encoding model and the EEG data automatic segmentation and recognition model used in this process.

[0066] 1.1. BERT-like encoding model training

[0067] This paper simulates the training method of the BERT model in natural language processing and improves the training task. There is no specific data requirement. Since collecting EEG data alone is a huge workload, data from various existing public EEG datasets can be used.

[0068] The details are as follows:

[0069] (1) Symbolization of brainwave data:

[0070] First, according to the data collected by the acquisition device, it is expressed as X (x1, x2, ... x i ), get the maximum value V possible for each time step of the data max and minimum value V min , the interval [V min , V max ] is evenly divided into N boxes, represented by 1 to N. For the value x in the data i , if V min ≤x i ≤V min +(V max -V min ) / N, then x i Recorded as 1; if V min +(V max -V min ) / N ≤x i ≤V min +2*(V max -V min ) / N, then x i Recorded as 2; and so on, encode X into a sequence represented by 1-N.

[0071] (2) The embedding process of symbolic data is as follows Figure 5 Specifically, the brainwave data is processed into a symbol sequence and then the BERT embedding method is simulated.

[0072] Token Embeddings is a word vector. The first word is the CLS flag, which indicates the data boundary and can be used for subsequent classification tasks.

[0073] Segment Embeddings are used to distinguish inputs and outputs in sequence prediction, which involves using data from the previous L1 time steps to predict data from the next L2 time steps. This is similar to sentence prediction in BERT.

[0074] Position Embeddings represent learnable position encodings. Their parameters can be learned in subsequent training tasks.

[0075] (3) Use BERT’s network structure to train the improved task.

[0076] Here we use the BERT network structure directly without any changes. We perform the following two training tasks:

[0077] 31) T1: Masked language model

[0078] This step is consistent with the original BERT training. During training, 15% of the tokens are randomly masked, and the final loss function only calculates the masked tokens.

[0079] 32), T2: Sentence prediction

[0080] Since there are no so-called sentences in the EEG data, we use the common sequence prediction task instead. That is, in this step, sequence prediction is performed, using the data of the first L1 time steps in the EEG data to predict the data of the last L2 time steps.

[0081] 1.2. Automatic segmentation and recognition model of brain wave data.

[0082] There is no requirement for data when training this model, and it can usually be completed through public brain wave datasets. The specific network structure is as follows Figure 6 As shown in the dotted box, the functions of each module in the above figure are as follows:

[0083] Self-attention module: This module is implemented using the self-attention mechanism. It mainly calculates the similarity score between each data value of the encoded brain wave data. If the length of the brain wave data is L, the output of this module is an L*L matrix. For example, for the value a ij Represents the similarity score between the i-th value and the j-th value.

[0084] Graph convolution layer: Graph convolution is performed on the output of the self-attention module. The resulting output is a graph feature vector for each point in the data. Assuming it is an L*E1 matrix, the i-th point in the original data is represented by a graph feature vector of length E1. This feature has two functions: first, it is used for state number identification, and second, it is used for state feature transformation.

[0085] State number prediction: Based on the output of the graph convolutional layer, identify the state number (that is, the number in the composition) corresponding to each point in the brain wave data.

[0086] State feature transformation: Based on the state number and the feature vector obtained by the graph convolution layer, the graph feature vector of each state number in the currently processed data is obtained, and the features corresponding to each state number are obtained through transformation.

[0087] State feature matrix: This state feature matrix can be learned through training. Its main function is to maintain a global state feature in the model, that is, to ensure that all values ​​predicted to have the same number in the data have the same or similar representation after state feature transformation.

[0088] The specific training process is as follows: After the brainwave encoding data is encoded by a BERT-like model, it first passes through the self-attention module to obtain the similarity between the value at each position in the data and the values ​​at other positions. Graph convolution is then used to extract features on the similarity matrix to obtain the graph representation features between the values. Based on the graph representation features, the number of each value in the data in the graph is predicted. The graph representation features of the values ​​with the same number are transformed to obtain the global state features of the same numbered state, and this global state feature is compared with the globally maintained state feature matrix. The final loss function is calculated by calculating the similarity between the numbered state features in the globally maintained state feature matrix and the numbered state features obtained in the current data, and taking the negative value of this similarity as the loss function.

[0089] The final brainwave data automatic segmentation and recognition model only needs to include three parts: the self-attention module, the graph convolution layer, and the state number prediction.

[0090] The following describes how to use the brain wave database

[0091] This corresponds to the database application system part, and the usage is as follows:

[0092] Optimal brainwave change path selection: This is achieved through a graph-based optimal path selection algorithm. Its main function is to obtain the optimal brainwave change process from the initial state to the final state given an initial state and a final state, such as Figure 7 As shown, when it is necessary to induce brain waves to change from state 1 to state 6, the best transformation process may be 1->2->3->6.

[0093] The BERT-like encoding model for brain wave data in this embodiment can be replaced by other encoding methods, or can be directly obtained through learning in the brain wave data automatic segmentation and recognition model.

[0094] In another aspect, the present invention further discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of any of the above methods.

[0095] On the other hand, the present invention further discloses a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of any of the above methods.

[0096] In another embodiment provided by the present application, a computer program product including instructions is also provided, which, when executed on a computer, enables the computer to execute the steps of any one of the methods in the above embodiments.

[0097] It is understandable that the system provided by the embodiment of the present invention corresponds to the method provided by the embodiment of the present invention, and the explanation, examples and beneficial effects of the relevant contents can refer to the corresponding parts of the above method.

[0098] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When executed, the program can include the processes of the above-described method embodiments. Any reference to memory, storage, database, or other media used in the various embodiments provided herein may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM).

[0099] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0100] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for constructing a brain wave state database, characterized in that: The following steps are included: Use brainwave data to construct brainwave state transition graphs through deep learning methods; Storing the constructed brainwave state transition graph in the form of a graph database to describe the transition paths between various brainwave states; The steps of using brain wave data to construct a brain wave state transition diagram through a deep learning method are as follows: First, a BERT-like encoding model for brainwave data is trained. Based on the brainwave data encoded by this BERT-like encoding model, a model for automatically segmenting and recognizing brainwave data is trained. This model can identify each value in the brainwave data as a different number ranging from 1 to n, where the value of n is determined experimentally. Data values ​​with the same number are in the same brainwave state, and the number also represents the node number of the final graph. The model for automatically dividing and identifying brain wave data includes: Self-attention module: A module implemented using the self-attention mechanism to calculate the similarity score between each data value of the encoded brainwave data. If the length of the brainwave data is L, the output of this module is an L*L matrix; for the value of the matrix a ij represents the similarity score between the i-th value and the j-th value; Graph convolution layer: Graph convolution is performed on the output of the self-attention module. The output is a graph feature vector for each point in the data. This graph feature vector has two functions: on the one hand, it is used for state number recognition, and on the other hand, it is used for state feature transformation; State number prediction: Based on the output of the graph convolutional layer, identify the state number corresponding to each point in the brainwave data, which is also the number in the composition; State feature transformation: Based on the state number and the feature vector obtained by the graph convolution layer, the graph feature vector of each state number in the current processing data is obtained, and the feature corresponding to each state number is obtained through transformation; State feature matrix: A global state feature is maintained in the model, that is, all values ​​predicted to have the same number in the data have the same or similar representation after the state feature transformation; The graph construction module constructs a brain wave state transition graph based on each state transition path obtained from the brain wave data.

2. The method for constructing a brainwave state database according to claim 1, wherein: In the process of graph construction, the state numbers are regarded as nodes. If there is a transition process between states, there is an edge, otherwise there is no edge; The state transition path corresponding to each piece of brain wave data is collected. The same number represents the same state. The number of occurrences of each edge is counted as the weight of the edge.

3. The method for constructing a brainwave state database according to claim 1, wherein: The brainwave data encoded by the BERT encoding model is used to train a model for automatically segmenting and recognizing brainwave data. The training steps are as follows: (1) Symbolization of brainwave data: First, according to the data collected by the acquisition device, it is expressed as X (x1, x2,…x i ), get the maximum value V of each time step of the data max and minimum value V min , the interval [V min , V max ] is evenly divided into N boxes, represented by 1 to N respectively; for the value x in the data i , if V min ≤x i ≤V min +(V max -V min ) / N, then x i Recorded as 1; if V min +(V max -V min ) / N ≤x i ≤V min +2*(V max -V min ) / N, then x i Recorded as 2; and so on, encode X into a sequence represented by 1-N; (2) After processing the brainwave data into a symbol sequence, simulate the BERT embedding method; where: Token Embeddings are word vectors, and the CLS flag indicates the data boundary, which is used for subsequent classification tasks; Segment Embeddings is used to distinguish input and output in sequence prediction, which refers to using the data of the previous L1 time steps to predict the data of the next L2 time steps; Position Embeddings represents the learned position encoding; (3) Use BERT’s network structure to train improved tasks; Directly use the BERT network structure to perform the following two training tasks: 31) T1: Masked language model During training, 15% of tokens are randomly masked, and the final loss function only calculates the masked tokens. 32), T2: Sentence prediction In this step, sequence prediction is performed, using the data of the first L1 time steps in the brain wave data to predict the data of the next L2 time steps.

4. The method for constructing a brain wave state database according to claim 1, wherein: The brainwave data encoded based on the BERT encoding model is used to train a model for automatically segmenting and recognizing brainwave data. The training process is as follows: After the brainwave data is encoded by a BERT-like model, it first passes through the self-attention module to obtain the similarity between the value at each position in the data and the values ​​at other positions. Graph convolution is then used to extract features on the similarity matrix to obtain the graph representation features between the values. Based on the graph representation features, the number of each value in the data in the graph is predicted, and the graph representation features of the values ​​with the same number are transformed to obtain the global state features of the same numbered state. This global state feature is then compared with the globally maintained state feature matrix. The final loss function is calculated by calculating the similarity between the numbered state features in the globally maintained state feature matrix and the numbered state features obtained in the current data, and taking the negative value as the loss function.

5. A system for constructing an EEG state database, for implementing the method for constructing an EEG state database according to any one of claims 1 to 4, characterized in that: Includes the following units, Data analysis and mapping module, used to construct brainwave state transition diagrams using a large amount of brainwave data through deep learning methods; The brainwave state data storage module is used to store the acquired brainwave state transition graph in the form of a graph database, which is used to describe the transition paths between various brainwave states and facilitate various applications related to brainwave states; Database application module, which relies on various applications implemented by the brainwave state database.

6. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 4.

Images