Real-time 4D-CBCT imaging method based on Unet-Transform model

Abstract

The invention provides a real-time 4D-CBCT imaging method based on a Unet-Transform model, and the method comprises the steps: constructing a comprehensive cascade deep learning network model integrating a Unet network and a Transform network, training the network model, inputting a plurality of CBCT projections containing different respiratory movement amplitudes, which are collected at the same rack angle at the online treatment moment of a patient, into the trained network model, the model outputs three PCA feature tags of a DVF between a predicted phase of breathing motion of a patient at the current moment and a reference phase corresponding to CBCT projection in real time, the PCA feature tags are converted into the DVF through principal component analysis, a reference phase image of an earlier-stage 4D-CT of the patient is deformed under the guidance of the DVF, a corresponding 4D-CBCT image of the breathing phase at the current moment is obtained, and the CBCT image of the breathing motion of the patient at the current moment is obtained. And finally, a 4D-CBCT image sequence capable of covering a complete respiratory cycle is obtained. According to the method, dynamic volume imaging is carried out on the patient during SBRT fractional treatment beam exiting by utilizing a Unt-Transform network model framework, a tool for accurately monitoring movement of tumors in the lung during beam exiting treatment of the patient is provided for clinicians, and technical guarantee is provided for accurate and safe implementation of the lung cancer SBRT.

Description

Technical Field

[0001] The present invention relates to the technical field of image guidance for stereotactic radiotherapy of lung cancer, and in particular to a real-time 4D-CBCT imaging method based on a Unet-Transformer model. Background Art

[0002] With the advancement of deep learning technology, the clinical definition of cone beam computed tomography (CBCT) image guidance in radiotherapy has gradually expanded from 3D imaging for pre-accelerator beam alignment to providing real-time 4D-CBCT images during online treatment. 4D-CBCT is an image guidance system integrated into the head of a linear accelerator for radiotherapy. Currently, methods for providing real-time 4D-CBCT images during online treatment are still in the preclinical research stage. A popular approach first uses a patient's previous 4D-CT scan to perform orthographic projections at a fixed angle simulating the accelerator's CBCT imaging geometry, acquiring a large number of 2D orthographic projections with varying respiratory motion amplitudes. These orthographic projections are then used as a training dataset to train a deep learning network. Once the network is trained, a single CBCT projection corresponding to the patient at the same gantry angle is acquired during online treatment and fed into the trained deep learning network. The network immediately outputs the motion feature label between the patient's respiratory motion phase and the reference phase at the current moment in real time. From this label, the corresponding image DVF (Deformable Vector Fields) can be generated later. Under the guidance of DVF, the patient's previous 4D-CT reference phase image is further deformed to obtain the current respiratory phase image. Since this process is real-time, when 9 to 10 projections of the patient with different respiratory amplitudes and covering the entire respiratory cycle obtained at the gantry angle are continuously input and input into the network in sequence, a complete 4D-CBCT image sequence can be output in real time. However, due to the different physical conditions of clinical patients, the stability and generalization of the aforementioned method of real-time 4D-CBCT images at the time of online treatment in a wide range of clinical samples still need to be further verified.

[0003] Currently, image guidance for stereotactic body radiation therapy (SBRT) for lung cancer typically involves performing an online 4D-CBCT scan before the accelerator delivers the beam, followed by positioning correction based on the online CBCT scan. However, the inventors of this application have discovered through clinical research that the current 4D-CBCT functional module cannot provide real-time 4D reconstructed images during beam delivery (i.e., within fractionated treatment). This results in the physician being unable to determine whether the beam is accurately landing on the moving target area in the lung at the moment of actual beam delivery, and therefore cannot help clinicians monitor the movement trajectory of the patient's lung tumor during beam delivery. Summary of the Invention

[0004] To address the technical problem that existing 4D-CBCT functional modules cannot provide real-time 4D reconstructed images during beam delivery (i.e., within fractionated treatment), resulting in the doctor being unable to determine whether the beam accurately lands on the moving target area of ​​the lung at the actual moment of beam delivery, and therefore unable to help clinicians monitor the movement trajectory of the patient's lung tumor during beam delivery treatment, the present invention provides a real-time 4D-CBCT imaging method based on the Unet-Transformer model. This method utilizes the Unet-Transformer network model framework to perform dynamic volumetric imaging during the delivery of the patient's SBRT fractionated treatment, providing clinicians with a tool to accurately monitor the movement of the patient's lung tumor during beam delivery treatment, and providing technical support for the accurate and safe implementation of SBRT for lung cancer.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] The real-time 4D-CBCT imaging method based on the Unet-Transformer model includes:

[0007] A comprehensive cascade deep learning network model integrating Unet and Transformer networks was constructed and trained. CBCT projections with varying respiratory motion amplitudes acquired at the same gantry angle during the patient's online treatment were input into the trained model. The model then output, in real time, three PCA feature labels representing the difference between the predicted respiratory motion phase and the reference phase of the corresponding CBCT projection.

[0008] The PCA feature labels are converted into DVF through principal component analysis. The conversion formula is as follows:

[0009]

[0010] Among them, F and denote the DVF and mean motion vector field, respectively; p and q denote the principal components and their corresponding PCA feature labels, respectively; i and j denote the respiratory phase and the index of the principal component, respectively;

[0011] Under DVF guidance, the patient's previous 4D-CT reference phase image is deformed to obtain the corresponding current respiratory phase 4D-CBCT image, thereby ultimately obtaining a 4D-CBCT image sequence that can cover the entire respiratory cycle.

[0012] Furthermore, in the integrated cascade deep learning network model structure integrating the Unet network and the Transformer network, a Transformer module is embedded in the Unet network as a calming layer to supplement feature detail extraction. The Unet deep learning network model structure embedded with the Transformer module specifically includes:

[0013] A dual convolution submodule for receiving simulated orthographic projection data, wherein the dual convolution submodule is sequentially connected to four input cascade undersampling modules, namely the first to the fourth, and the fourth input cascade undersampling module is sequentially connected to a Transformer module, and the Transformer module is sequentially connected to four upsampling modules, namely the first to the fourth, and the fourth upsampling module is connected to an output convolution layer via two two-dimensional maximum pooling layers; and

[0014] A skip connection mechanism is added between the dual convolution submodule and the fourth upsampling module, between the first in-cascade undersampling module and the third upsampling module, between the second in-cascade undersampling module and the second upsampling module, and between the third in-cascade undersampling module and the first upsampling module.

[0015] Furthermore, the dual convolution submodule includes two serially connected two-dimensional convolution units, each of which includes a two-dimensional convolution layer, a two-dimensional batch normalization unit and a ReLu activation function connected in sequence.

[0016] Furthermore, in the four in-cascade undersampling modules, each in-cascade undersampling module includes a two-dimensional maximum pooling unit and a dual convolution submodule connected in sequence.

[0017] Furthermore, the Transformer module includes a reorganization layer, a multi-head attention mechanism, a Dropout layer, a feedforward neural network module and a Dropout layer connected in sequence.

[0018] Furthermore, among the four upsampling modules, each upsampling module includes a two-dimensional transposed convolution layer, a feature splicing layer and a double convolution sub-module connected in sequence.

[0019] Furthermore, the output convolution layer includes an adaptive two-dimensional average pooling layer, a flattening layer and a linear layer connected in sequence.

[0020] Furthermore, the training process of the comprehensive cascade deep learning network model integrating the Unet network and the Transformer network is as follows:

[0021] First, the nine groups of DVFs generated between each phase image of the patient's previous 4D-CT image and the 0% reference phase image were subjected to PCA decomposition to obtain the initial nine groups of PCA coefficients, i.e., PCA feature label groups.

[0022] Afterwards, the PCA coefficients covering the patient's maximum respiratory amplitude range were selected and the PCA coefficients were equally divided into 1000 groups, obtaining 1000 groups of PCA coefficients with different respiratory amplitudes. The reference phase images were then deformed in sequence to obtain 1000 groups of 4D-CT with different motion amplitudes.

[0023] Next, according to the accelerator CBCT imaging geometry, these 1,000 sets of 4D-CT images were simulated for orthographic projection at a fixed angle to obtain a large amount of orthographic projection data containing different respiratory motion amplitudes. These orthographic projection data were input into the constructed comprehensive cascade deep learning network model integrating the Unet network and the Transformer network, allowing the network model to fully learn the respiratory motion characteristics in the patient's orthographic projection data.

[0024] Compared with the existing technology, the real-time 4D-CBCT imaging method based on the Unet-Transformer model provided by the present invention utilizes the Unet-Transformer network model framework to perform dynamic volumetric imaging during the beam delivery of a patient's SBRT fractionated treatment. This method can achieve complete real-time 4D-CBCT reconstruction of beam delivery moments within a radiotherapy fraction without changing the existing accelerator CBCT clinical scanning protocol, providing a reliable means for real-time monitoring imaging within fractionated SBRT treatment for lung cancer. Furthermore, the method uses only three PCA feature labels (numerical labels) to directly represent the DVF, a data volume with a huge amount of data, thereby greatly reducing the computational load of the network and ensuring the real-time performance of the entire prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the Unet-Transformer network model structure provided by the present invention.

[0026] Figure 2 It is a schematic diagram comparing the image generated by the Unet-Transformer network model provided by the present invention and the gold standard image of the XCAT digital phantom.

[0027] Figure 3 It is a schematic diagram comparing the clinical prediction result image of the Unet-Transformer network model provided by the present invention and the gold standard image. DETAILED DESCRIPTION

[0028] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below with reference to specific illustrations.

[0029] The present invention provides a real-time 4D-CBCT imaging method based on a Unet-Transformer model, comprising:

[0030] A comprehensive cascade deep learning network model integrating the Unet and Transformer networks (referred to as the Unet-Transformer model) was constructed and trained. CBCT projections with varying respiratory motion amplitudes acquired at the same gantry angle during the patient's online treatment were input into the trained comprehensive cascade deep learning network model. The network model output, in real time, three PCA feature labels (corresponding to the first three feature labels of the PCA) representing the difference between the predicted respiratory motion phase of the patient at the current moment and the reference phase of the corresponding CBCT projection.

[0031] The PCA feature labels are converted into DVF through principal component analysis (PCA). The conversion formula is as follows:

[0032]

[0033] Among them, F and denote the DVF and mean motion vector field, respectively; p and q denote the principal components and their corresponding PCA feature labels (PCA coefficients), respectively; i and j denote the index of the respiratory phase and principal component, respectively;

[0034] Under DVF guidance, the patient's previous 4D-CT reference phase image is deformed to obtain the corresponding current respiratory phase 4D-CBCT image, thereby ultimately obtaining a 4D-CBCT image sequence that can cover the entire respiratory cycle.

[0035] As a specific example, please refer to Figure 1 As shown, in the integrated cascade deep learning network model structure integrating the Unet network and the Transformer network, a Transformer module is embedded in the Unet network as a calming layer to supplement feature detail extraction. The Unet deep learning network model structure embedded with the Transformer module specifically includes:

[0036] A double convolution submodule DoubleConv is used to receive simulated orthographic projection data of size 1*384*512, and the double convolution submodule is sequentially connected to the first to fourth input cascade undersampling modules Down1 to Down4, and the fourth input cascade undersampling module Down4 is connected to the Transformer module, and the Transformer module is sequentially connected to the first to fourth four upsampling modules Up1 to Up4, and the fourth upsampling module Up4 is connected to the output convolution layer OutConv through two two-dimensional maximum pooling layers MaxPool2d. The model finally outputs 3 PCA coefficient prediction values; and

[0037] Skip connections are added between the double convolution submodule DoubleConv and the fourth upsampling module Up4, between the first cascade undersampling module Down1 and the third upsampling module Up3, between the second cascade undersampling module Down2 and the second upsampling module Up2, and between the third cascade undersampling module Down3 and the first upsampling module Up1. By adding the skip connection mechanism, the effective fusion and transmission of features at different levels are ensured, avoiding the gradient vanishing or gradient exploding problems.

[0038] The integrated cascade deep learning network model structure provided in the above embodiment, which integrates the Unet network and the Transformer network, enhances the network's feature extraction capability for motion-sensitive areas while taking into account spatiotemporal continuity, further improving the network's real-time 4D-CBCT volume imaging prediction accuracy and achieving results superior to those obtained by modeling each network module separately.

[0039] As a specific example, please refer to Figure 1 As shown, the dual convolution submodule DoubleConv includes two two-dimensional convolution units connected in series, and each two-dimensional convolution unit includes a two-dimensional convolution layer Conv2d, a two-dimensional batch normalization unit BatchNorm2d and a ReLu activation function connected in sequence. Among them, the two-dimensional convolution layer Conv2d is used to extract motion features in the input projection image, the two-dimensional batch normalization unit BatchNorm2d is used to process data in batches, and the ReLu activation function is used to introduce nonlinear factors into the neural network model, so that the model can learn and express more complex function mapping relationships. The dual convolution submodule DoubleConv provided in this embodiment can further extract features and enhance the expression ability of the network.

[0040] As a specific example, please refer to Figure 1As shown, each of the four cascaded undersampling modules Down1 to Down4 includes a two-dimensional maximum pooling unit MaxPool2d and a double convolution submodule DoubleConv, connected in sequence. The two-dimensional maximum pooling unit MaxPool2d is used to reduce the size of the feature map while retaining important feature information, and the double convolution submodule DoubleConv is used to further extract features and enhance the network's expressive power. By using the four cascaded undersampling modules provided in this embodiment, the image size can be continuously reduced, and shallow feature information can be extracted.

[0041] As a specific example, please refer to Figure 1 As shown, the Transformer module includes a sequentially connected reshape layer, a multi-head attention mechanism, a Dropout layer, a feedforward neural network module FFN, and a Dropout layer. The reshape layer is used to compress the hidden representation of each sequence in the three-dimensional tensor into one dimension to facilitate subsequent fully connected layer processing. In the projection data [B, C, H, W] received by the reshape layer, B represents Batch, i.e., the amount of projection data processed in each batch, and C / H / W represents the size of each input projection, i.e., 1x384x256. The multi-head attention mechanism is used to process each position in the sequence in parallel, which enables the model to effectively capture long-range dependencies. At the same time, by focusing on different parts of the sequence in different representation subspaces, multi-head attention can enrich the model's representation capabilities. The Dropout layer is used to prevent overfitting, and the residual link can retain features. The residual link joint layer normalization LayerNorm can enhance network stability and prevent gradient explosion. The feedforward neural network module FFN is used to implement the self-attention mechanism. Self-attention is a type of attention mechanism that allows the model to focus on any position in the sequence while processing the current input. Through feedforward neural networks, the Transformer can capture long-range dependencies in the input sequence and better understand the semantic information in the input sequence. The Transformer module provided in this embodiment is embedded in the Unet network architecture and acts as a calming layer, which can effectively supplement feature detail extraction.

[0042] As a specific example, please refer to Figure 1As shown, in the four upsampling modules Up1 to Up4, each upsampling module includes a two-dimensional transposed convolution layer ConvTranspose2d, a feature splicing layer (torch.cat) and a double convolution submodule DoubleConv connected in sequence; wherein, the two-dimensional transposed convolution layer ConvTranspose2d is used for the transposed convolution operation to amplify and restore the initial feature map, and provides flexibility in combination with learnable parameters, the feature splicing layer (torch.cat) is used to splice multiple tensors along a specific dimension to form a larger tensor, and the double convolution submodule DoubleConv is used to further extract features and enhance the expression ability of the network. The four series upsampling modules provided in this embodiment can well complete the feature splicing of the projection data details.

[0043] As a specific example, please refer to Figure 1 As shown, the output convolution layer OutConv includes an adaptive two-dimensional average pooling layer AdaptiveAvgPool2d, a flattening layer Flatten, and a linear layer Linear connected in sequence; wherein, the adaptive two-dimensional average pooling layer AdaptiveAvgPool2d is used for adaptive average pooling operation, uniformly adjusting the feature maps of various sizes output by the previous layer to the required size, the flattening layer Flatten is used to reduce the amount of calculation and encode the global information of the image into each small pacth, and the linear layer Linear is used to map the output of the model to the vocabulary space of the prediction target or generated text. The output convolution layer OutConv provided in this embodiment is used to generate the final prediction result. The output channel depends on the task requirements. This model sets the number of channels to 3, which represents the output of three predicted PCA coefficients.

[0044] As a specific embodiment, the training process of the comprehensive cascade deep learning network model integrating the Unet network and the Transformer network is as follows:

[0045] First, nine groups of DVFs generated between each phase image of the patient's previous 4D-CT images and the 0% reference phase image (4D-CT has 10 phases, and there is a DVF between each phase and the reference phase) were subjected to PCA decomposition to obtain the initial nine groups of PCA coefficients, i.e., PCA feature label groups.

[0046] Afterwards, the PCA coefficients covering the patient's maximum respiratory amplitude range were selected and the PCA coefficients were equally divided into 1000 groups, obtaining 1000 groups of PCA coefficients with different respiratory amplitudes. The reference phase images were then deformed in sequence to obtain 1000 groups of 4D-CT with different motion amplitudes.

[0047] Next, according to the accelerator CBCT imaging geometry, these 1000 sets of 4D-CT were simulated for orthographic projection at a fixed angle, such as 0 degrees, to obtain a large amount of orthographic projection data (DRRs, Digital Reconstructed Radiography) containing different respiratory motion amplitudes. These orthographic projection data DRRs were input into the constructed integrated cascade deep learning network model (Unet-Transformer model) integrating the Unet network and the Transformer network, allowing the network model to fully learn the respiratory motion characteristics in the patient's orthographic projection data, thereby completing the network model training.

[0048] After training is completed, CBCT projections of the patient at a certain respiratory phase during online treatment are collected at the same angle and input into the trained network model. The network model can output three PCA feature labels corresponding to the DVF between the current respiratory phase and the reference phase in real time. After converting this PCA feature label into DVF through principal component analysis, the patient's previous 4D-CT reference phase image is deformed under the guidance of DVF to obtain the corresponding 4D-CBCT image of the current respiratory phase.

[0049] By continuously inputting several CBCT projections with different motion amplitudes at the same angle, the network model can output corresponding different groups of PCA coefficients in real time, thereby ultimately obtaining a 4D-CBCT image sequence that can cover the entire respiratory cycle. Since the entire process does not require excessive computing resources, it is visually real-time.

[0050] The present invention conducts digital simulation experiments based on XCAT digital phantoms. The images generated by the Unet-Transformer model and the gold standard images of the XCAT digital phantom are respectively as follows: Figure 2 As shown. Figure 2 It can be seen that although the images generated by the Unet-Transformer model do not completely match the gold standard images, they are very close to the gold standard images overall. Therefore, it can provide clinicians with a tool to accurately monitor the movement of lung tumors during beam treatment, providing technical support for the precise and safe implementation of SBRT for lung cancer.

[0051] The method of the present invention was used to conduct clinical trials on real patients, and the clinical trial data results are as follows: Figure 3 As shown in (b), the images (a) and (b) in the figure contain three views: transverse, sagittal and coronal. Figure 3 As can be seen, the clinical prediction result image (b) of the Unet-Transformer network model is very close to the gold standard image (a). Therefore, it can provide clinicians with a tool to accurately monitor the movement of lung tumors during beam treatment, providing technical support for the precise and safe implementation of SBRT for lung cancer.

[0052] Compared with the existing technology, the real-time 4D-CBCT imaging method based on the Unet-Transformer model provided by the present invention utilizes the Unet-Transformer network model framework to perform dynamic volumetric imaging during the beam delivery of a patient's SBRT fractionated treatment. This method can achieve complete real-time 4D-CBCT reconstruction of beam delivery moments within a radiotherapy fraction without changing the existing accelerator CBCT clinical scanning protocol, providing a reliable means for real-time monitoring imaging within fractionated SBRT treatment for lung cancer. Furthermore, the method uses only three PCA feature labels (numerical labels) to directly represent the DVF, a data volume with a huge amount of data, thereby greatly reducing the computational load of the network and ensuring the real-time performance of the entire prediction.

[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A real-time 4D-CBCT imaging method based on the Unet-Transformer model, characterized by: include: A comprehensive cascade deep learning network model integrating Unet and Transformer networks was constructed and trained. CBCT projections with varying respiratory motion amplitudes acquired at the same gantry angle during the patient's online treatment were input into the trained model. The model then output, in real time, three PCA feature labels representing the difference between the predicted respiratory motion phase and the reference phase of the corresponding CBCT projection. The PCA feature labels are converted into DVF through principal component analysis. The conversion formula is as follows: Among them, F and denote the DVF and mean motion vector field, respectively; p and q denote the principal components and their corresponding PCA feature labels, respectively; i and j denote the respiratory phase and the index of the principal component, respectively; Under DVF guidance, the patient's previous 4D-CT reference phase image is deformed to obtain the corresponding current respiratory phase 4D-CBCT image, thereby ultimately obtaining a 4D-CBCT image sequence that can cover the entire respiratory cycle.

2. The real-time 4D-CBCT imaging method based on the Unet-Transformer model according to claim 1, characterized in that: In the comprehensive cascade deep learning network model structure integrating the Unet network and the Transformer network, a Transformer module is embedded in the Unet network as a calming layer to supplement feature detail extraction. The Unet deep learning network model structure embedded with the Transformer module specifically includes: A dual convolution submodule for receiving simulated orthographic projection data, wherein the dual convolution submodule is sequentially connected to four input cascade undersampling modules, namely the first to the fourth, and the fourth input cascade undersampling module is sequentially connected to a Transformer module, and the Transformer module is sequentially connected to four upsampling modules, namely the first to the fourth, and the fourth upsampling module is connected to an output convolution layer via two two-dimensional maximum pooling layers; and A skip connection mechanism is added between the dual convolution submodule and the fourth upsampling module, between the first in-cascade undersampling module and the third upsampling module, between the second in-cascade undersampling module and the second upsampling module, and between the third in-cascade undersampling module and the first upsampling module.

3. The real-time 4D-CBCT imaging method based on the Unet-Transformer model according to claim 2, characterized in that: The dual convolution submodule includes two serially connected two-dimensional convolution units, each of which includes a two-dimensional convolution layer, a two-dimensional batch normalization unit and a ReLu activation function connected in sequence.

4. The real-time 4D-CBCT imaging method based on the Unet-Transformer model according to claim 2, characterized in that: In the four in-cascade undersampling modules, each in-cascade undersampling module includes a two-dimensional maximum pooling unit and a dual convolution submodule connected in sequence.

5. The real-time 4D-CBCT imaging method based on the Unet-Transformer model according to claim 2, characterized in that: The Transformer module includes a reorganization layer, a multi-head attention mechanism, a Dropout layer, a feedforward neural network module and a Dropout layer connected in sequence.

6. The real-time 4D-CBCT imaging method based on the Unet-Transformer model according to claim 2, characterized in that: Among the four upsampling modules, each upsampling module includes a two-dimensional transposed convolution layer, a feature splicing layer and a double convolution submodule connected in sequence.

7. The real-time 4D-CBCT imaging method based on the Unet-Transformer model according to claim 2, characterized in that: The output convolution layer includes an adaptive two-dimensional average pooling layer, a flattening layer and a linear layer connected in sequence.

8. The real-time 4D-CBCT imaging method based on the Unet-Transformer model according to claim 1, characterized in that: The training process of the comprehensive cascade deep learning network model integrating the Unet network and the Transformer network is as follows: First, the nine groups of DVFs generated between each phase image of the patient's previous 4D-CT image and the 0% reference phase image were subjected to PCA decomposition to obtain the initial nine groups of PCA coefficients, i.e., PCA feature label groups. Afterwards, the PCA coefficients covering the patient's maximum respiratory amplitude range were selected and the PCA coefficients were equally divided into 1000 groups, obtaining 1000 groups of PCA coefficients with different respiratory amplitudes. The reference phase images were then deformed in sequence to obtain 1000 groups of 4D-CT with different motion amplitudes. Next, according to the accelerator CBCT imaging geometry, these 1,000 sets of 4D-CT images were simulated for orthographic projection at a fixed angle to obtain a large amount of orthographic projection data containing different respiratory motion amplitudes. These orthographic projection data were input into the constructed comprehensive cascade deep learning network model integrating the Unet network and the Transformer network, allowing the network model to fully learn the respiratory motion characteristics in the patient's orthographic projection data.

Images