Oral cavity panoramic X-ray image tooth segmentation method based on deep learning

Abstract

The invention discloses an oral cavity panoramic X-ray image tooth segmentation method based on deep learning. The method comprises the following steps: preprocessing data of an oral cavity panoramic X-ray image; designing and constructing an oral cavity panoramic X-ray image segmentation network model; constructing an evaluation system of the instance segmentation model; constructing a self-training iterative model method; training and obtaining an oral tooth segmentation model with the optimal precision on the verification set; inputting a label-free oral panoramic X-ray image into the trained oral tooth segmentation model, and reasoning the image to complete mask segmentation; and the oral tooth reasoning result is visualized. According to the method, the segmentation model with excellent performance can be obtained only by publicly annotating the data set without additional annotation of clinical experts, and the method has the advantages of high efficiency, high precision, good objectivity, stable repeatability and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and medical image processing, and in particular to a method for tooth segmentation of oral panoramic X-ray images based on deep learning. Background Art

[0002] Precise tooth segmentation can help dentists better assess the health of teeth, plan treatment plans, and monitor treatment effects. Oral panoramic images, as a cheap and efficient imaging method commonly used in clinical practice, play an important role in dental diagnosis and treatment.

[0003] However, the current mainstream oral panoramic image segmentation methods in the industry have significant limitations. These solutions mostly rely on a large amount of labeled data, which usually needs to be performed by experienced dental experts, consuming a lot of time and cost. At the same time, manual labeling may also introduce subjectivity and consistency issues, which will affect the training effect of the model. In addition, the fully supervised model is difficult to achieve real-time online updates in a clinical environment, and cannot quickly adapt to new data or new situations, which limits its flexibility and generalization ability in practical applications. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the present invention provides a tooth segmentation method for oral panoramic X-ray images based on deep learning, aiming to improve the accuracy and efficiency of diagnosis and reduce dependence on the personal experience of dentists.

[0005] In order to achieve the above-mentioned invention object, the technical solution adopted to solve the technical problem is as follows:

[0006] A method for tooth segmentation of oral panoramic X-ray images based on deep learning, comprising the following steps:

[0007] Step 1: Model data preprocessing, grayscale value standardization, size unification, and data enhancement processing of the input panoramic data;

[0008] Step 2: Use the Transformer+UNet-based architecture design to build a tooth segmentation network model for oral panoramic X-ray images;

[0009] Step 3: Build an evaluation system for the instance segmentation model and use three evaluation indicators: DICE coefficient, intersection-over-union ratio, and Hausdorff distance to comprehensively evaluate the performance of the model;

[0010] Step 4: In the model training phase, the model is trained using self-supervised and semi-supervised learning strategies. In the main and secondary backbone network parts, the pre-training parameters of ResNet50 come from the large public image dataset ImageNet, and the parameters of other modules are randomly initialized; the input oral panoramic X-ray image data enters the tooth segmentation network model constructed in step 2 after the preprocessing described in step 1. The model is trained using the gradient descent algorithm and combined with a pseudo-label generation mechanism to enable unlabeled data to participate in the training; the performance of the model is evaluated on the validation set by calculating the evaluation indicators defined in step 3, and the optimal parameters of the model are determined based on the performance indicators;

[0011] Step 5: Model network reasoning. In the model reasoning stage, the optimal parameters determined in step 4 are used as the final network parameters of the model. An oral panoramic X-ray image data is input. After the preprocessing described in step 1, the data is sent to the tooth segmentation network model constructed in step 2. The model automatically completes the tooth segmentation and returns the prediction result.

[0012] Furthermore, in step 2, the segmentation network model as a whole is composed of the following parts:

[0013] Main backbone network: ResNet50 is used as the backbone network to extract multi-scale features of the image. The network transfers features at different levels to the subsequent decoder module through skip connections;

[0014] Neck network: The deep extraction network uses a visual Transformer structure to capture global context information. After passing through the projection layer, the features are input into the Transformer module, which consists of 12 Transformer Blocks. Each Block contains a multi-head self-attention mechanism and a feedforward neural network to extract deep features from the input image.

[0015] Decoder network: It consists of CBAM decoder modules. The CBAM decoder module enhances important features by combining channel and spatial attention to improve the segmentation accuracy of the model. The features output by each layer of decoder are fused with the features of the jump connection through splicing operations, and finally the original resolution of the image is restored through the upsampling layer.

[0016] Head network: includes a segmenter structure for generating tooth segmentation masks, which enables more accurate segmentation through deep supervision of multi-scale features.

[0017] Furthermore, the main backbone network uses ResNet50 as the backbone network, which is responsible for extracting multi-scale features of the image. The network transfers features of different levels to the subsequent decoder module through skip connections, including the following:

[0018] For the feature map F with an initial dimension of H×W×1, after passing through the ResNet backbone network 2 The dimensions are:

[0019] (H / 32)×(W / 32)×D

[0020] Among them, H and W are the initial dimensions of the image, both are 620, and D is the number of channels, which is 512.

[0021] Furthermore, the neck network feature candidate region is generated using a Transformer module to capture global context information, including the following:

[0022] For the feature map F 2 The subsequent mathematical description is as follows:

[0023] Q=F 2 W Q

[0024] K=F 2 W K

[0025] V=F 2 W V

[0026] Among them, W Q , W K , W V represents the fully connected matrix weight, F 2 Represents the feature map after passing through the ResNet backbone network;

[0027] Then after the attention mechanism:

[0028]

[0029] MultiHead(Q,K,V)=Concat(head 1 ,…,head h )W O

[0030] in, Represents the feature map of each head, d k Represents the dimension of Q, K vectors, and W represents the weight of the fully connected matrix;

[0031] Finally, the residual connection is used to avoid gradient disappearance:

[0032] LayerOut = LayerNorm(F 2 +MultiHead(Q,K,V)

[0033] Among them, F 2Represents the feature map after passing through the ResNet backbone network, and LayerNorm represents layer normalization.

[0034] Furthermore, the decoder network feature map uses a CBAM decoder module, which reduces the spatial dimension of the input features through global average pooling and global maximum pooling to generate two different feature vectors:

[0035] Favg c , Fmax c

[0036] Among them, Favg c Represents the vector after average pooling of the upsampled feature map, Fmax c Represents the vector after the maximum pooling of the upsampled feature map;

[0037] The two feature vectors are transformed through a shared fully connected layer, and the channel attention weights are obtained through the Sigmoid activation function. The process is described as:

[0038]

[0039] Among them, σ represents the Sigmoid function, and MLP represents the fully connected layer;

[0040] The features output by the channel attention module are input to the spatial attention module, which generates two two-dimensional feature maps through maximum pooling and average pooling operations on the channel dimension, combines these feature maps through convolutional layers, and finally generates a spatial attention map through the Sigmoid function:

[0041]

[0042] Among them, f 7×7 Represents the convolution layer, and the convolution kernel size is 7×7;

[0043] Finally, the original features are weighted by channel attention weights and spatial attention weights to obtain the output features:

[0044] F′=M c (F)·F

[0045] F″=M s (F′)·F′

[0046] Among them, F represents the feature map before entering the CBAM module, M c (F) represents the channel attention processing first, and then M s (F') Its spatial attention processing.

[0047] Furthermore, the head network structure adopts a deep supervision method to enhance gradient propagation: the feature maps are O 1 ,O 2 ,…,O n ,in, represents the i-th feature map, and the target map is Y∈R H×W×C , the Loss at each stage is as follows:

[0048]

[0049] in, represents the i-th feature map, and the true label map is Y i ∈R H×W×C ,H i ,W i ,C represents the width, height and number of channels of the feature map;

[0050] The overall loss function is:

[0051]

[0052] Among them, λ i Represents the weight of each stage, Represents the loss at each stage.

[0053] Furthermore, using publicly available labeled datasets Train the model. The initial model training is expressed as follows:

[0054]

[0055] Among them, x i represents the sample, y i represents the label, L(θ) is the supervision loss function, θ is the parameter of the model, and f(·) represents the model;

[0056] Then generate pseudo labels: use the trained model to generate pseudo labels for unlabeled data. Perform inference and generate pseudo labels The mathematical expression is as follows:

[0057]

[0058] Among them, θ 1 Represents the model parameters of version 1;

[0059] Combined dataset: Pseudo-labeled data With labeled dataset Combine to form a new training data set:

[0060]

[0061] in, represents the newly fused dataset;

[0062] Iterative training model: copy the parameters of the previous model and use them on the new training dataset The model is iteratively trained above. For each round of iteration, the model is updated as follows:

[0063]

[0064] Among them, x i and i Represents the training images and label images in the new dataset.

[0065] Preferably, the random initialization method in step 4 is: he_normal, lecun_uniform, glorot_normal, glorot_uniform or lecun_normal.

[0066] Preferably, the gradient descent algorithm in step 4 is: Adam, SGD, MSprop or Adadelta.

[0067] Due to the adoption of the above technical solution, the present invention has the following advantages and positive effects compared with the prior art:

[0068] 1. The semi-supervised tooth segmentation model of the present invention can accurately segment the pathological areas of large and small targets with higher efficiency through multi-scale feature extraction of the main backbone network and feature transformation of the neck network.

[0069] 2. The technical solution of the present invention takes into account the complexity of tooth shape, and the decoding network also needs to understand context information. Therefore, the present invention introduces a dual-dimensional modeling attention head in the network model to better complete mask segmentation through channel space dimension modeling.

[0070] 3. The training step of the technical solution of the present invention can continuously update the model weights based on clinical production data to better adapt to complex and changeable clinical scenarios.

[0071] 4. According to the technical solution of the present invention, based on computer vision and medical image processing technology, it is possible to automatically segment and identify complex dental conditions in oral panoramic X-ray images, which is more efficient and accurate than traditional diagnostic methods that rely on dentists' experience and intuitive judgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0073] Figure 1 It is a flow chart of a method for teeth segmentation of oral panoramic X-ray images based on deep learning according to an embodiment of the present invention;

[0074] Figure 2 is a model network architecture diagram of an embodiment of the present invention;

[0075] Figure 3 is a training flow chart of an embodiment of the present invention;

[0076] Figure 4 Schematic diagram of tooth segmentation result according to an embodiment of the present invention. DETAILED DESCRIPTION

[0077] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0078] like Figure 1-4 As shown, this embodiment discloses a method for teeth segmentation of oral panoramic X-ray images based on deep learning, comprising the following steps:

[0079] Step 1: Model data preprocessing, grayscale value standardization, size unification, and data enhancement processing of the input panoramic data;

[0080] Step 2: Use the Transformer+UNet-based architecture design to build a tooth segmentation network model for oral panoramic X-ray images;

[0081] Furthermore, in step 2, the segmentation network model as a whole is composed of the following parts:

[0082] Main backbone network: ResNet50 is used as the backbone network to extract multi-scale features of the image. The network transfers features at different levels to the subsequent decoder module through skip connections;

[0083] Specifically, for the initial dimension of H×W×1, the feature map F after passing through the ResNet backbone network 2 The dimensions are:

[0084] (H / 32)×(W / 32_×D

[0085] Among them, H and W are the initial dimensions of the image, both are 620, and D is the number of channels, which is 512.

[0086] Neck network: The deep extraction network uses a visual Transformer structure to capture global context information. After passing through the projection layer, the features are input into the Transformer module, which consists of 12 Transformer Blocks. Each Block contains a multi-head self-attention mechanism and a feedforward neural network to extract deep features from the input image.

[0087] Specifically, the neck network feature candidate region generation adopts a Transformer module to capture global context information, including the following:

[0088] For the feature map F 2 The subsequent mathematical description is as follows:

[0089] Q=F 2 W Q

[0090] K=F 2 W K

[0091] V=F 2 W V

[0092] Among them, W Q , W K , W V represents the fully connected matrix weight, F 2 Represents the feature map after passing through the ResNet backbone network;

[0093] Then after the attention mechanism:

[0094]

[0095] MultiHead(Q,K,V)=Concat(head 1 ,…,head h )W O

[0096] in, Represents the feature map of each head, d k Represents the dimension of Q, K vectors, and W represents the weight of the fully connected matrix;

[0097] Finally, the residual connection is used to avoid gradient disappearance:

[0098] LayerOut = LayerNorm(F 2 +MultiHead(Q,K,V)

[0099] Among them, F 2 Represents the feature map after passing through the ResNet backbone network, and LayerNorm represents layer normalization.

[0100] Decoder network: It consists of a CBAM (Convolutional Block Attention Module) decoder module (attention mechanism). The CBAM decoder module enhances important features by combining channel and spatial attention to improve the segmentation accuracy of the model. The features output by each layer of the decoder are fused with the features of the jump connection through the splicing operation, and finally the original resolution of the image is restored through the upsampling layer;

[0101] Specifically, the decoder network feature map uses a CBAM decoder module, which reduces the spatial dimension of the input features through global average pooling and global maximum pooling to generate two different feature vectors:

[0102] Favg c ,Fmax c

[0103] Among them, Favg c Represents the vector after average pooling of the upsampled feature map, Fmax c Represents the vector after the maximum pooling of the upsampled feature map;

[0104] The two feature vectors are transformed through a shared fully connected layer, and the channel attention weights are obtained through the sigmoid activation function. The process is described as:

[0105]

[0106] Among them, σ represents the Sigmoid function, and MLP represents the fully connected layer;

[0107] The features output by the channel attention module are input to the spatial attention module, which generates two two-dimensional feature maps through maximum pooling and average pooling operations on the channel dimension, combines these feature maps through convolutional layers, and finally generates a spatial attention map through the sigmoid function:

[0108]

[0109] Among them, f 7×7 Represents the convolution layer, and the convolution kernel is 7×7;

[0110] Finally, the original features are weighted by channel attention weights and spatial attention weights to obtain the output features:

[0111] F′=M c (F)·F

[0112] F″=M s (F′)·F′

[0113] Among them, F represents the feature map before entering the CBAM module, M c (F) represents the channel attention processing first, and then M s (F)') its spatial attention processing;

[0114] Head network: includes a segmenter structure for generating tooth segmentation masks, which enables more accurate segmentation through deep supervision of multi-scale features.

[0115] Specifically, the head network structure adopts a deep supervision method to enhance gradient propagation: the feature maps are O 1 ,O 2 ,…,O n ,in, represents the i-th feature map, and the target map is Y∈R H×W×C , the Loss at each stage is as follows:

[0116]

[0117] in, represents the i-th feature map, and the true label map is Y i ∈R H×W×C ,H i ,W i ,C represents the width, height and number of channels of the feature map;

[0118] The overall loss function is:

[0119]

[0120] Among them, λ i Represents the weight of each stage, Represents the loss at each stage.

[0121] Step 3: Build an evaluation system for the instance segmentation model and use three evaluation indicators: DICE Similarity Coefficient (DSC), Intersection over Union (IoU), and Hausdorff Distance to comprehensively evaluate the performance of the model.

[0122] Step 4: In the model training phase, the model is trained using self-supervised and semi-supervised learning strategies. In the main and secondary backbone network parts, the pre-training parameters of ResNet50 come from the large public image dataset ImageNet, and the parameters of other modules are randomly initialized; the input oral panoramic X-ray image data enters the tooth segmentation network model constructed in step 2 after the preprocessing described in step 1. The model is trained using the gradient descent algorithm and combined with a pseudo-label generation mechanism to enable unlabeled data to participate in the training; the performance of the model is evaluated on the validation set by calculating the evaluation indicators defined in step 3, and the optimal parameters of the model are determined based on the performance indicators;

[0123] Furthermore, the model training process is as follows Figure 3 As shown, pseudo labels are generated and trained 4 times. Using a publicly available labeled dataset Train the model. The initial model training is expressed as follows:

[0124]

[0125] Among them, x i represents the sample, y i represents the label, L(θ) is the supervision loss function, θ is the parameter of the model, and f(·) represents the model;

[0126] Then generate pseudo labels: use the trained model to generate pseudo labels for unlabeled data. Perform inference and generate pseudo labels The mathematical expression is as follows:

[0127]

[0128] Among them, θ 1 Represents the model parameters of version 1;

[0129] Combined dataset: Pseudo-labeled data With labeled dataset Combine to form a new training data set:

[0130]

[0131] in, represents the newly fused dataset;

[0132] Iterative training model: copy the parameters of the previous model and use them on the new training dataset The model is iteratively trained above. For each round of iteration, the model is updated as follows:

[0133]

[0134] Among them, x iand i Represents the training images and label images in the new dataset.

[0135] Preferably, the random initialization method in step 4 is: he_normal, lecun_uniform, glorot_normal, glorot_uniform or lecun_normal.

[0136] Preferably, the gradient descent algorithm in step 4 is: Adam (Adaptive Moment Estimation), SGD (Stochastic Gradient Descent), MSprop or Adadelta.

[0137] Step 5: Model network inference. In the model inference stage, the optimal parameters determined in step 4 are used as the final network parameters of the model. A panoramic oral X-ray image data is input. After the preprocessing described in step 1, the data is sent to the tooth segmentation network model constructed in step 2. The model automatically completes the tooth segmentation and returns the prediction result. The visualization result is as follows: Figure 4 As shown, Figure 4 It is the visualization result of the panoramic image after segmentation network inference.

[0138] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for tooth segmentation in oral panoramic X-ray images based on deep learning, characterized in that: The following steps are involved: Step 1: Model data preprocessing: grayscale value standardization, size unification, and data enhancement processing of the input panoramic data; Step 2: Use the Transformer+UNet-based architecture design to build a tooth segmentation network model for oral panoramic X-ray images; Step 3: Build an evaluation system for the instance segmentation model and use three evaluation indicators: DICE coefficient, intersection-over-union ratio, and Hausdorff distance to comprehensively evaluate the performance of the model; Step 4: In the model training phase, the model is trained using self-supervised and semi-supervised learning strategies. In the main and secondary backbone network parts, the pre-training parameters of ResNet50 come from the large public image dataset ImageNet, and the parameters of other modules are randomly initialized; the input oral panoramic X-ray image data enters the tooth segmentation network model constructed in step 2 after the preprocessing described in step 1. The model is trained using the gradient descent algorithm and combined with a pseudo-label generation mechanism to enable unlabeled data to participate in the training; the performance of the model is evaluated on the validation set by calculating the evaluation indicators defined in step 3, and the optimal parameters of the model are determined based on the performance indicators; Step 5: Model network reasoning. In the model reasoning stage, the optimal parameters determined in step 4 are used as the final network parameters of the model. An oral panoramic X-ray image data is input. After the preprocessing described in step 1, the data is sent to the tooth segmentation network model constructed in step 2. The model automatically completes the tooth segmentation and returns the prediction result.

2. The method for teeth segmentation of oral panoramic X-ray images based on deep learning according to claim 1, characterized in that: In step 2, the segmentation network model consists of the following parts: composition: Main backbone network: ResNet50 is used as the backbone network to extract multi-scale features of the image. The network transfers features at different levels to the subsequent decoder module through skip connections; Neck network: The deep extraction network uses a visual Transformer structure to capture global context information. After passing through the projection layer, the features are input into the Transformer module, which consists of 12 Transformer Blocks. Each Block contains a multi-head self-attention mechanism and a feedforward neural network to extract deep features from the input image. Decoder network: It consists of CBAM decoder modules. The CBAM decoder module enhances important features by combining channel and spatial attention to improve the segmentation accuracy of the model. The features output by each layer of decoder are fused with the features of the jump connection through splicing operations, and finally the original resolution of the image is restored through the upsampling layer. Head network: includes a segmenter structure for generating tooth segmentation masks, which enables more accurate segmentation through deep supervision of multi-scale features.

3. The method for teeth segmentation of oral panoramic X-ray images based on deep learning according to claim 2, characterized in that: The main backbone network uses ResNet50 as the backbone network, which is responsible for extracting multi-scale features of the image. The network transfers features of different levels to the subsequent decoder module through skip connections, including the following: For an initial dimension of H×W×1, the dimension of the feature map F2 after passing through the ResNet backbone network is: (H / 32)×(W / 32_×D Among them, H and W are the initial dimensions of the image, both are 620, and D is the number of channels, which is 512.

4. The method for teeth segmentation of oral panoramic X-ray images based on deep learning according to claim 2, characterized in that: The neck network feature candidate region generation adopts the Transformer module to capture global context information, including the following: The subsequent mathematical description of the feature map F2 is as follows: Q=F2W Q K=F2W K V=F2W V Among them, W Q , W K , W V represents the fully connected matrix weight, and F2 represents the feature map after passing through the ResNet backbone network; Then after the attention mechanism: MultiHead(Q,K,V)=Concat(head1,…,head h )W O in, Represents the feature map of each head, d k Represents the dimension of Q, K vectors, and W represents the weight of the fully connected matrix; Finally, the residual connection is used to avoid gradient disappearance: LayerOut=LayerNorm(F2+MultiHead(Q,K,V)) Among them, F2 represents the feature map after passing through the ResNet backbone network, and LayerNorm represents layer normalization.

5. The method for teeth segmentation of oral panoramic X-ray images based on deep learning according to claim 2, characterized in that: The decoder network feature map uses the CBAM decoder module, which reduces the spatial dimension of the input features through global average pooling and global maximum pooling to produce two different feature vectors: Favg c ,Fmax c Among them, Favg c Represents the vector after average pooling of the upsampled feature map, Fmax c Represents the vector after the maximum pooling of the upsampled feature map; The two feature vectors are transformed through a shared fully connected layer, and the channel attention weights are obtained through the Sigmoid activation function. The process is described as: Among them, σ represents the Sigmoid function, and MLP represents the fully connected layer; The features output by the channel attention module are input to the spatial attention module, which generates two two-dimensional feature maps through maximum pooling and average pooling operations on the channel dimension, combines these feature maps through convolutional layers, and finally generates a spatial attention map through the Sigmoid function: Among them, f 7×7 Represents the convolution layer, and the convolution kernel size is 7×7; Finally, the original features are weighted by channel attention weights and spatial attention weights to obtain the output features: F’=M c (F)·F F”=M s (F’)·F’ Among them, F represents the feature map before entering the CBAM module, M c (F) represents the channel attention processing first, and then M s (F') Its spatial attention processing.

6. The method for teeth segmentation of oral panoramic X-ray images based on deep learning according to claim 2, characterized in that: The head network structure adopts deep supervision method to enhance gradient propagation: the feature maps are O1, O2, ..., O n , the Loss at each stage is as follows: in, Represents the i-th feature map, and the true label map is Y i ∈R H×W×C ,H i ,W i ,C represents the width, height and number of channels of the feature map; The overall loss function is: Among them, λ i represents the weight of each stage, Represents the loss at each stage.

7. The method for teeth segmentation of oral panoramic X-ray images based on deep learning according to claim 2, characterized in that: Using publicly available labeled datasets Train the model. The initial model training is expressed as follows: Among them, x i represents the sample, y i represents the label, L(θ) is the supervision loss function, θ is the parameter of the model, and f(·) represents the model; Then generate pseudo labels: use the trained model to generate pseudo labels for unlabeled data. Perform inference and generate pseudo labels The mathematical expression is as follows: Among them, θ1 represents the model parameters of version 1; Combined dataset: Pseudo-labeled data With labeled dataset Combine to form a new training data set: in, represents the newly fused dataset; Iterative training model: copy the parameters of the previous model and use them on the new training dataset The model is iteratively trained above. For each round of iteration, the model is updated as follows: Among them, x i and i Represents the training images and label images in the new dataset.

8. The method for teeth segmentation of oral panoramic X-ray images based on deep learning according to claim 1, characterized in that: The random initialization method in step 4 is: he_normal, lecun_uniform, glorot_normal, glorot_uniform or lecun_normal.

9. The method for teeth segmentation of oral panoramic X-ray images based on deep learning according to claim 1, characterized in that: The method of the gradient descent algorithm in step 4 is: Adam, SGD, MSprop or Adadelta.

Images