Method and system for three-dimensional shape reconstruction from topograms in medical imaging
Reconstructing 3D liver shape from two-dimensional positioning film images through machine learning generation networks, solving the high cost and radiation exposure problems of CT scan dependence, and achieving efficient and accurate 3D shape reconstruction and topological details improvement.
Patent Information
- Application Number
- CN201980080459.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2018-12-05
- Filing Date
- 2019-05-31
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2039-05-31
AI Technical Summary
Existing 3D liver shape reconstruction methods rely on CT scans, resulting in long doctor-patient interaction time, high cost and radiation exposure, while simultaneously difficulty extracting 3D shapes from X-ray images, especially visual blur and information loss due to projection .
Using machine learning to generate networks, including positioning chip encoder and decoder, through training generation adversarial networks or variational automatic encoding networks, reconstruct 3D shapes from two-dimensional positioning chip images, use mask information to improve prediction accuracy, and 3D shapes can be effectively predicted without a mask.
It realizes efficient and accurate reconstruction of 3D liver shape from X-ray images, reducing radiation exposure and cost, while improving the accuracy and topological details of shape prediction, with low errors and suitable for medical imaging systems.
Smart Images

Figure CN113272869B_ABST
Abstract
Description
[0001] Related Applications
[0002] This patent document claims the benefit of the filing date under 35 USC §119(e) of provisional U.S. patent application serial number 62 / 775,440, filed on December 5, 2018, which is hereby incorporated by reference. Background Art
[0003] This embodiment involves viewing true organ shapes. Accurate depiction of internal organs, such as the liver, can allow for more accurate health screening, early diagnosis, and procedure planning that targets specific locations in the human body. Visualization of three-dimensional (3D) organs aids clinical applications, such as surgical planning and post-operative evaluation, as well as pathology detection and disease diagnosis.
[0004] The liver and other organs can exhibit highly heterogeneous shape variations, making the liver difficult to segment. Liver volume segmentation is performed semi-automatically or automatically using statistical shape models, sigmoid-edge modeling, graph cutting or other processes. Automated deep learning-based methods can be used. However, these methods rely on volume information available from computed tomography (CT) scans. Existing delineation tools depict 2D shapes in each slice of a 3D CT volume and combine this set of predictions into a 3D shape. Due to the lack of spatial context, this intermediate slice-by-slice processing may introduce additional sources of error to the overall shape prediction quality. The procedure for obtaining CT scans involves long doctor-patient interaction time, high-cost machines, and exposure to a certain dose of radiation. Practical challenges in obtaining CT scans may hinder the acquisition of accurate 3D organ delineation.
[0005] Topograms, which are 2D images projected from tomographic devices such as X-rays, can be more readily available, less expensive, require less radiation, and are often used by medical professionals. Delineating 3D organ shapes from two-dimensional (2D) X-ray images (e.g., topograms) is difficult due to visual blurring and information loss as a result of projections in X-ray imaging. Extracting 2D shapes from X-rays is particularly complex because projections can result in complex or blurred textures, boundaries, and overlapping anatomical parts. To mitigate these challenges, traditional methods use prior knowledge such as motion patterns or intensity and background analysis in order to perform X-ray segmentation. More recent methods focus on learning segmentation using deep neural networks. For example, generative adversarial networks (GANs) are used to decompose X-ray images into non-overlapping components to improve segmentation quality. Unpaired image-to-image translation techniques are applied to learn to segment X-ray images by observing CT scan segmentations. These methods have achieved results on 2D shape delineation and segmentation tasks, but not in 3D.
[0006] In the field of computer vision, deep generative 3D shape models based on variational autoencoder networks (VAE) and generative adversarial networks (GAN) are trained to generate complex shape topologies. Combined with mappings from image space, these methods are able to infer 3D shape predictions from 2D observations. In order to obtain more detailed and accurate predictions, input annotations (such as landmarks or masks) are often used to guide the synthesis process. For example, 2D landmarks are incorporated for alignment optimization of skin vertex-based human shape models with image observations. Landmark annotations can guide the synthesis of 3D shapes observed in the input image. Landmarks or contours can be incorporated to formulate additional objective terms, thereby improving performance in 3D shape reconstruction and synthesis problems. However, these computer vision-based systems may not handle the noise and complexity from medical scans and fusion well, especially in projection-based medical imaging. Summary of the invention
[0007] Systems, methods and instructions on a computer-readable medium are provided for reconstructing a 3D shape from a localizer. A generating network is machine trained. The generating network includes a localizer encoder for inputting a localizer and a decoder for outputting a 3D shape from the output of the decoder. For training, one or more other encoders such as regularizers for inputting a mask and / or inputting a 3D shape are included. The localizer encoder and decoder are trained together with one or more other encoders output to the decoder. For application, with or without an encoder for a mask and without an encoder for a 3D shape, such as trained localizer encoders and decoders, are used to estimate the 3D shape of the patient based on the input of the localizer of the patient.
[0008] In a first aspect, a method for reconstructing a 3D shape from a patient topogram in a medical imaging system is provided. A patient topogram representing a projection through a patient in two dimensions is acquired. In response to inputting the patient topogram into a machine learning generative network, a 3D shape of an object represented in the patient topogram is reconstructed by the machine learning generative network. Information from the 3D shape is displayed.
[0009] The topogram is acquired using an X-ray imager or other projection imager. The 3D shape is reconstructed as the outer surface of the object, such as a surface mesh of the liver. The X-ray image includes a projection of an object such as an organ, and the patient's organ is reconstructed.
[0010] In one embodiment, the generative network of machine learning is trained as a generative adversarial network, or is trained as a network with an encoder and decoder based on variable autoencoding. In some embodiments, the network has a localization sheet encoder configured to receive a patient localization sheet, and the network has been trained with a 3D shape encoder and with a localization sheet encoder, the 3D shape encoder receives a training 3D representation in the training data as input, and the localization sheet encoder receives a training localization sheet in the training data as input. In other embodiments, the network has a localization sheet encoder configured to receive a patient localization sheet, and the network has been trained with a mask encoder, the mask encoder receives a mask in the training data as input. In yet another embodiment, the generative network of machine learning utilizes a 3D shape encoder and is trained with a mask encoder, the 3D shape encoder receives a 3D representation in the training data as input, and the mask encoder receives a mask in the training data as input. For reconstruction, any of these embodiments is reconstructed in response to the input of a patient localization sheet, the patient localization sheet having or not having an object mask from the patient localization sheet.
[0011] In one approach, the displayed information is a 3D rendering of the 3D shape of the object. In another approach, the displayed information is a quantity determined from the 3D shape, such as a volume.
[0012] In a second aspect, a method for performing machine training to reconstruct a 3D representation of an anatomical structure from a first localizer in a medical imaging system is provided. A machine learning architecture is defined having a first encoder configured to receive a second localizer as input and a second encoder configured to receive a mask or a first 3D shape as input. The machine learning architecture is used to machine learn to output a 3D representation in response to the input of the first localizer. The machine learning model generated by the machine learning is stored for later application.
[0013] In one embodiment, the architecture is defined as a decoder configured to receive the output of the first and second encoders and output a second 3D shape in response to receiving the output of the first and second encoders. In one example, the architecture is defined wherein the second encoder is configured to receive the mask and the third encoder is configured to receive the first 3D shape as input. In one method, the architecture is defined as a generative model based on variable autoencoding.
[0014] In a third aspect, a medical imaging system for 3D organ estimation is provided. An X-ray imager is configured to generate an X-ray image of a patient. An image processor is configured to generate an estimate of a 3D organ based on an input of the X-ray image to a machine learning model. The machine learning model is configured to receive as input a view from a single direction, wherein the X-ray image is an X-ray image of a view from a single direction. A display is configured to display information from the 3D organ estimation.
[0015] In one embodiment, a first encoder for localization slices and a second encoder for 3D shapes are used to train the machine learning model. In another embodiment, a first encoder for localization slices and a second encoder for the contours of organs in localization slices are used to train the machine learning model. In yet another embodiment, a first encoder for localization slices, a second encoder for 3D shapes, and a third encoder for the contours of organs in localization slices are used to train the machine learning model.
[0016] Any one or more of the above aspects may be used alone or in combination. Any aspect or method used in one type of claim (e.g., method or system) may be used in another type of claim (e.g., method, system, or computer-readable medium). These and other aspects, features, and advantages will become apparent from the following detailed description of the preferred embodiments read in conjunction with the accompanying drawings. The present invention is defined by the following claims, and nothing in this section should be construed as limiting those claims. Further aspects and advantages of the present invention are discussed below in conjunction with the preferred embodiments, and the aspects and advantages may be claimed later, either independently or in combination. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Components and the figures are not necessarily drawn to scale, emphasis instead being placed on illustrating the principles of the embodiments. Moreover, in the figures, like reference numerals designate corresponding parts throughout the different views.
[0018] Figure 1 is a flow chart of one embodiment of a method for machine training for reconstructing a 3D representation of an anatomical structure from a topogram in a medical imaging system;
[0019] Figure 2 and Figure 3 illustrates an example architecture for machine training of 3D shape estimation from 2D topograms;
[0020] Figure 4 is a flow chart of one embodiment of a method for reconstructing a 3D shape from a patient positioning film in a medical imaging system;
[0021] Figure 5 shows 3D shape changes in the liver from a CT scan;
[0022] Figure 6 Example 3D organ shape reconstructions are shown using example topograms of different architectures in training;
[0023] Figure 7 shows an example 2D projection from the predicted 3D organ shape; and
[0024] Figure 8 is a block diagram of one embodiment of a system for 3D organ estimation from topograms. DETAILED DESCRIPTION
[0025] Reconstructing 3D organ shape from topogram images. For example, the shape of the liver in three dimensions is predicted using a generative model. The 3D organ shape is automatically predicted directly from the topogram images, which are easier to acquire than volumetric CT scans and have limited exposure to radiation during acquisition. To improve prediction accuracy, a 2D mask can be input to the generative model. Automatic or semi-automatic methods for 3D organ shape reconstruction from topograms can be used to accurately predict the shape of the observed 3D organ given a single topogram or projection image. In automatic delineation, the 3D shape is output based solely on the topogram images. In semi-automatic delineation, the user outlines an approximate two-dimensional mask that is used in conjunction with the topogram to obtain a more accurate 3D shape prediction.
[0026] In one embodiment, the generative shape model includes a 3D shape encoder and decoder and an encoder from 2D observations (either topograms only, or topograms and masks). The 3D shape encoder and decoder form a Variant Autoencoder (VAE) generative model to represent each shape observation using a compact low-dimensional representation. The topogram and optional mask encoders map partial observations from the image (and masks when provided) to the coordinates of the corresponding shape observations. The entire architecture is optimized end-to-end to simultaneously infer shape from topogram image observations and learn the underlying shape space. This allows for simultaneous learning of a generative shape space that covers complex shape variations from 3D supervision and inferring shape from input 2D observations.
[0027] In the example used herein, the organ of interest is the liver. The organ shape prediction method is general and can be used for organs other than the human liver, such as the lungs or kidneys. The prediction method can be used for other anatomical structures rather than specific organs. The prediction method can be used for inserted or non-natural objects, such as stents or joints.
[0028] Figure 1The invention is a flowchart of an embodiment of a method for machine training, which is used to reconstruct a 3D representation of an anatomical structure from a localizer in a medical imaging system. The model is machine trained to predict the 3D shape of an organ based on a single localizer or images from only one viewing direction (e.g., one or more localizers projected along the same axis from the same or different times, such as a single localizer image and a mask created based on the localizer). The generator model is machine trained to predict the 3D shape based on a 2D projection. Alternatively, the machine is trained to predict the 3D shape based on localizers from different viewing directions without the need for iterative reconstruction of tomography. The model can be trained using an encoder for receiving a localizer, a 3D shape and / or a mask and a decoder for outputting a 3D shape based on the value of a feature provided by the encoder.
[0029] The method is performed in the order shown (i.e., from top to bottom or numerically) or in a different order. For example, acts 11, 12, and 13 are performed in any order. Additional, different, or fewer acts may be provided. For example, act 11 is provided without acts 12 and 13. In another example, only one of acts 12 or 13 is provided with act 11 as part of act 10. As yet another example, acts for collecting training data and / or application of a generator model for machine learning are included. The method is performed by a computer, server, workstation, or Figure 8 The system may be executed by a system that has access to a memory or database for storing training data or parameter values of the architecture and / or the generated model.
[0030] In act 10, a machine learning architecture is defined. The programmer programs or configures the architecture. For example, the type of machine learning is selected. The architecture or arrangement of nodes, layers, processes, operations, or other features of the model to be learned is set. The number of layers or units of the network, the type of learning, and other characteristics are controlled by the programmer or user. In other embodiments, one or more aspects (e.g., the number of nodes, the number of layers or units, or the type of learning) are defined and selected by the machine during learning. The architecture is defined using graphical programming, menu selection, or other methods.
[0031] In one embodiment, a generative (e.g., image-to-image) network is defined. Any machine training architecture for outputting a spatial distribution from an input spatial distribution can be used. The architecture is a neural network, such as a fully connected neural network or a convolutional neural network. The architecture (such as in a U-net arrangement) includes one or more encoders and decoders. A convolution to transposed convolution network can be used. One segment of a layer or unit (e.g., an encoder) applies convolution to increase the abstractness or compressibility of the input localizer. The most abstract feature value (e.g., bottleneck) is then output to another segment. Another segment of a layer or unit (e.g., a decoder) then applies a transposed convolution to reduce the abstractness or compressibility, resulting in the output of a 3D shape.
[0032] In one embodiment, the architecture is defined as a Generative Adversarial Network (GAN). The GAN includes a generator, such as an image-to-image or U-net, and a discriminator. The generator includes an encoder (convolutional) network and a decoder (transposed convolutional) network, with connections between the transferred features at the maximum compression or abstraction level from the encoder to the decoder. Skip connections at lower compression levels from the encoder to the decoder may be included. The adversarial network is used in training to learn to distinguish the output of the generator from the ground truth. The generator learns to output more realistic 3D shapes, while the discriminator learns to distinguish better in iterative training. Once the generator is trained, it is used without the discriminator in testing or application. In other embodiments, the generator can be trained without the discriminator.
[0033] The generator is a deep architecture that can include a convolutional neural network (CNN) or a deep belief network (DBN). Other deep networks can be used. CNNs learn a feed-forward mapping function, while DBNs learn a generative model of the data. In addition, while DBNs are fully connected networks (i.e., have different weights for all regions of the image), CNNs use shared weights for all local regions. CNNs are trained entirely with back-propagation for discrimination. DBNs, on the other hand, employ layer-by-layer unsupervised training (e.g., pre-training), followed by discriminative refinement using back-propagation if necessary.
[0034] A network is defined as a number of sequential feature units or layers. The order is used to indicate the general flow of output feature values from one layer to the input to the next layer. Information from the next layer is fed to the next layer, and so on, until the final output. The layers may be forward-feed only or may be bidirectional, including some feedback to previous layers. The nodes of each layer or unit may be connected to all or only a subset of the nodes of the previous or subsequent layers or units.
[0035] Deep architectures are defined as learning features at different levels of abstraction based on input images with or without preprocessing, rather than pre-programming features and trying to associate features with attributes. The features are learned to reconstruct lower level features (i.e., features at a more abstract or compressed level). For example, features are learned to reconstruct a localizer. For the next unit, features are learned to reconstruct the features of the previous unit, providing more abstraction. Each node of the unit represents a feature. Different units are provided for learning different features.
[0036] In a unit or layer, any number of nodes is provided. For example, 100 nodes are provided. Later or subsequent units may have more, less, or the same number of nodes. In general, for convolutions, subsequent units have more abstraction. For transposed convolutions for reconstruction, the level of abstraction is reversed. Each unit or layer reduces the level of abstraction or compression.
[0037] The machine uses any building block to learn the features of the nodes. For example, an autoencoder (AE) or restricted Boltzmann machine (RBM) method is used. The AE transforms the data linearly and then applies nonlinear rectification, such as a sigmoid function. The objective function of the AE is the expected mean square error between the input image and the image reconstructed using the learned features. The AE can learn the features that lead to the best reconstruction using stochastic gradient descent or other methods by the machine. A variational AE (VAE) can be used. The objective function of the RBM is an energy function. The exact calculation of the likelihood term associated with the RBM is intractable. Therefore, approximate algorithms such as contrast divergence based on k-step Gibb sampling or others are used to train the RBM to reconstruct the image from the features. Sparse or denoising techniques (e.g., sparse denoising AE (SDAE)) can be employed. Enforcing sparsity within the hidden layer (i.e., activating only a small number of units in the hidden layer at a time) can also regularize the network. In other embodiments, at least one unit is a convolution with ReLU activation, or a batch normalization (BN+LeakyRu+convolution) with a ReLU activation followed by a convolution layer. Max pooling, upsampling, downsampling, and / or softmax layers or units may be used. Different units may be of the same or different types.
[0038] Any now known or later developed architecture may be used. Support vector machines, Bayesian networks, other neural networks, or other arrangements may be used. Other neural networks may be used.
[0039] In one embodiment, generating a shape model is defined to include one or more encoders based on 2D observations. Figure 2 and Figure 3 Two embodiments of a defined architecture for learning and applying are shown.An encoder 24 receives the position patch 21 as input. Figure 2 and Figure 3 Another 3D shape encoder 26 is shown which receives the 3D shape 22 as input. Figure 3 Another mask encoder 23 is shown that receives the mask 20 as a 2D input. In other embodiments, the mask encoder 23 is not provided with the 3D shape encoder 26. The mask encoder 23 is shown as learning (i.e., Figure 3 In other embodiments, it can be used only in learning and not in the test pipeline. Figure 2 and Figure 3 Of these two, the 3D shape encoder 26 is only used in the learning pipeline and not in the testing pipeline. Figure 2 and Figure 3 The learning pipeline in is all the architecture shown (e.g., the test pipeline and the encoder 26 shown outside the test pipeline).
[0040] The 3D shape encoder 26 receives as input a 3D shape 22, such as a ground truth 3D shape from a CT scan. The 3D shape encoder 26 is used as a regularizer in training and is not used in applications (e.g., a test pipeline) because the 3D shape is unknown to the input. The generative model has a 3D shape encoder 26 and a decoder 28 that is shared by all encoders 23, 24, 26 by receiving bottleneck feature values 27 from different encoders 23, 24, 26. The 3D shape encoder 26 learns to map the 3D shape of the organ to their coordinates in the latent space, and the decoder 28 learns to reconstruct the shape from their coordinates. The generative model receives a set of examples As input, is the example shape 22, and is the corresponding topographic image 21 observed. The generative model G = (Q, P) includes the encoding component Q. According to the prior distribution The shape is distributed in a random low-dimensional space s Mapped to its potential coordinates z ,and The potential coordinates z Mapping back to shape space S .
[0041] The 3D shape encoder 26 maps observations represented by a 64×64×64 voxel grid to a 200-dimensional latent vector z Other sizes of voxel grids and / or vectors can be used. For the variational autoencoder model, the normal distribution parameters are defined as and , but other parameter values may also be used. According to an example embodiment, the architecture of the 3D shape encoder 26 has five convolutional layers with outputs of sizes 64, 128, 256, 512, 200, a kernel size of 4 for each layer, and padding sizes of 1, 1, 1, 1, and 0. The convolutional layers are separated by batch normalization layers and ReLU layers. Other sizes, numbers of layers, and / or layer types may be used.
[0042] The decoder 28 architecture mirrors the architecture of the 3D shape encoder 26. The decoder 28 may mirror the other encoders 24, 23, or may have a different arrangement than a mirroring of any of the encoders 23, 24, 26 in other embodiments. The decoder 28 is defined as receiving the output 27 of the encoders 23, 24 and / or 26 and outputting a predicted 3D shape 29 in response to receiving or inputting an output value (e.g., output 27). The decoder 28 is defined as a transposed convolutional network.
[0043] The topogram or 2D image encoder 24 receives the topogram 21 as input. For training, it receives the topogram I of the training data. For application, it receives the patient topogram. The topogram encoder 24 learns to map the two-dimensional projection observations to the coordinates of the corresponding shape. Given a generative model G , learn the localizer image encoder 24 so that for each observation ,image i Mapped to coordinate positions , so that the reconstructed shape is 29, With the reference truth shape 22, s As close as possible.
[0044] In one embodiment, the topogram encoder 24 obtains a 1×256×256 topogram image and outputs a 200-dimensional latent shape vector The patch encoder 24 is defined as having five convolutional layers with numbers of outputs 64, 128, 256, 512, 200, kernel sizes 11, 5, 5, 5, 8, and strides 4, 2, 2, 2, 1, separated by batch normalization and rectified linear units (ReLU). Other numbers of layers, input resolutions, output, sizes, strides, or layer types may be used.
[0045] The mask encoder 23 receives as input a mask 20. The mask 20 is formed from the topogram image 21 and therefore has the same view direction or projection angle as the topogram 21. The mask 20 is a binary designation or pixel labeling that represents or does not represent an object for which a 3D shape 29 is to be predicted. In other embodiments, the mask 20 is the boundary or outline (i.e., contour) of the object in the topogram 21. Other types of inputs that can be encoded using a neural network (e.g., landmarks) can also be used instead of or with the mask to improve prediction accuracy.
[0046] The mask encoder 23, if used, may or may not be used in the application and is used in the training. Figure 2 An example in which the mask encoder 23 is not used is shown. Figure 3 A mask encoder 23 is shown for both testing and learning. The mask encoder 23 and the 2D image encoder 24 are joint topogram and mask encoders 12 that predict the potential coordinates of the organ shape given the 2D mask 20 and topogram 21. When mask information is provided, it helps to generate more accurate predictions. For each observation , given positioning piece i and mask ,in is an orthogonal projection operator that jointly learns the localizer and mask encoder 23, 24 to output , so that and s As close as possible. In an alternative embodiment, Figure 2 In the definition architecture of , a mask encoder 23 is used instead of a 3D shape encoder 26. In other alternative embodiments, the mask encoder 23 is used for learning instead of testing.
[0047] In one embodiment, the positioning piece and the mask encoder I 2 is defined as having a localizer encoder 24 branch, a mask encoder 23 branch, and a common combiner network 25, such that observations are mapped to a common latent coordinate The combiner network 25 is a network of convolutional layers, but may have other arrangements. The patch encoder 24 branch has the same architecture as the patch encoder 24 discussed above, and maps the patch 21 to an intermediate 200-dimensional feature The mask encoder 23 branch receives a 1×64×64 binary mask image and learns to map it to a 200-dimensional vector using 5 convolutional layers. , the 5 convolutional layers have kernel sizes of 3, 3, 3, 3, and strides of 4, 2, 2, 2, 2, 2, which are separated by batch normalization and rectified linear units (ReLU). Other numbers of layers, sizes, strides, and / or layer types can be used. The output of encoders 23, 24, and , are then concatenated and run through a combiner network 25, which is used to predict the joint 200-dimensional latent coordinates The combiner network 25 may use other layer types, sizes, and / or numbers of layers.
[0048] Reference again Figure 1, a processor (e.g., an image processor) performs machine learning using a machine learning architecture to output a 3D representation 29 in response to an input of a topogram 21. For machine training, training data is acquired. The training data includes pairs of topograms 21 with reference truth shapes 22 of objects (e.g., organs, such as a liver) represented in the topogram 22. Samples of the training data may also include masks, such as manually or automatically generated masks 20 for the topograms 21. The training data forms a database of shape and X-ray (two-dimensional observation) pairs.
[0049] For training, various optimizers can be used, such as Adadelta, SGD, RMSprop, or Adam. The weights of the network are randomly initialized, but another initialization can be used. End-to-end training is performed, but one or more features can be set. Various encoders 23, 24, 26 and decoders 28 with or without combiners 25 are trained together or jointly, rather than sequentially. A combination or single loss function that compares the predicted 3D shape 29 to the ground truth is used to update various parts of the architecture. Batch normalization, dropout, and data augmentation are not used, but batch normalization, dropout, and data augmentation can be used (for example, using batch normalization and dropout). During optimization, different distinguishing features are learned. Features that provide 3D shape indications from localizers or localizers and masks are learned.
[0050] For training Figure 2 or Figure 3 The training phase involves jointly training the generative model (3D shape encoder 26 and decoder 28) with the 2D observation encoder (topogram encoder 24 or topogram encoder 24, mask encoder 23 and combiner 25) in an end-to-end process. During testing, only 2D observations are necessary for 3D shape prediction.
[0051] The optimizer minimizes an error or loss such as mean square error (MSE), Huber loss, L1 loss, or L2 loss. In one embodiment, the loss function of the generative model is composed of the reconstruction loss for variational autoencoder training. L rec and distribution loss L dist composition. L rec is the binary cross entropy (BCE) error, which measures the ground truth shape and predicted shape For example, L rec It is represented by:
[0052]
[0053] Where N = 64 3 . Other sizes or resolutions may be used for N. L dist It is mandatory The distribution loss of the potential distribution matching its previous distribution ,in The image encoder loss is the binary cross entropy (BCE) loss The loss of the combined mask encoder 23 and the positioning piece encoder 24 I 2 is the binary cross entropy (BCE) error Additional mask losses such as those given by:
[0054] .
[0055] This mask loss ensures that the input mask k and the projected mask of the predicted shape (Right now, ) to match. Any of the encoders can use other loss functions.
[0056] For combined training, all components of the system are optimized together in an end-to-end training process using a combined objective, such as given by:
[0057] ,
[0058] in is the weight applied to each loss type. If the topographic mask encoder 23, 24 is trained, then , and when training only the localization slice encoder 24, . is the reconstruction loss of VAE, and is the 2D-3D reconstruction loss. The model can be trained without the shape encoder, i.e., and . Other combined loss functions can be used.
[0059] exist Figure 1 In action 18, the processor stores the trained model in a memory. Various parameter values of the architecture as defined are determined during training. These values and the architecture are stored in the memory. The stored model can be used by the same or different processors to predict a 3D shape based on an input localizer or an input localizer and a mask. The stored machine learning model can be copied in different systems for application to localizers by different systems. For example, different workstations, computers, servers, medical scanners (e.g., X-ray imagers), or other medical imaging systems apply the machine learning model to localizers of different patients.
[0060] Figure 4 An embodiment of a method for reconstructing a 3D shape from a patient topogram in a medical imaging system is shown. A machine learning model, such as a Figure 1 A machine learning generative network trained to be appropriate for a given patient is discussed. A topogram (such as a single topogram) with or without a mask is input to the machine learning model, which outputs a prediction of the 3D shape of the object represented in the topogram.
[0061] The method is performed in the order shown (e.g., from top to bottom or numerically), but other orders may also be used. Additional, different, or fewer actions may be provided. For example, action 44 is not provided. In another example, an action of manually or automatically generating a mask from a positioning film is included.
[0062] Figure 4 The method is implemented by a medical imaging system. The medical imaging system can be a computer, a server, a workstation, or a medical imager or scanner. For example, an X-ray imager implements the action. In another example, in action 40, an X-ray imager or processor acquires a localizer; in action 42, the same or a different processor reconstructs; and in action 44, a display device or screen displays. Other devices may also be used.
[0063] In act 40, a single patient topogram is acquired. Multiple topograms may be acquired from different times and / or different view directions (e.g., orientation of the x-ray source relative to the patient). A topogram is an x-ray image or other image representing a projection (e.g., an integral) along one dimension to a 2D image (e.g., N×M, where N and M are integers greater than one).
[0064] Acquisition is performed by scanning the patient. For example, an X-ray imager generates X-rays directed at the patient. A detector on the opposite side of the X-ray source detects the X-rays, providing an indication of the X-ray intensity for each pixel of the detector. In other embodiments, acquisition is from transmission over a computer network or loading from a memory. For example, an X-ray image is acquired from a picture archiving and communication system (PACS).
[0065] The topogram of the patient includes a projection of an object of interest, such as the liver. The entire object in 2D is represented in the topogram (i.e., the projection of the object into 2D). Alternatively, only a portion of the object is represented. The topogram may represent other objects or devices, such as other organs, bones, or inserted devices. Alternatively, the topogram is processed to remove information from other objects.
[0066] A mask can be generated from the topogram. For example, segmentation is applied. The machine learning model segments the object, identifying locations in the topogram that represent the object. Other segmentations such as those based on thresholds, gradients, or another algorithm can be used. In other embodiments, the user manually inputs the segmentation, such as tracing the boundaries of the object in the topogram.
[0067] In action 42, the image processor reconstructs the 3D shape of the object represented in the patient topogram, such as reconstructing the patient's liver or other organs in 3D form (i.e., a representation of the spatial extent over three dimensions). The 3D shape is reconstructed as the outer surface of the object. The outer surface resulting from the topogram representation of the object is estimated. The outer surface can be a continuous surface, a voxel representation, or a grid. Alternatively, the 3D shape is reconstructed as a solid (e.g., an outer surface and an interior).
[0068] The topogram with or without the mask is input into a machine-learned generative network, which outputs a 3D shape. For example, the topogram and mask at the same resolution are input to the generative network as two channels. The generative network is applied as an input feature vector with whatever input it is trained to use, such as (a) only a single topogram or (b) only a single topogram and a single mask.
[0069] One or more encoders compute values for bottleneck and other features in response to inputs to the one or more encoders. A decoder generates a 3D shape from the bottleneck feature or the bottleneck and other features. Skip connections may be used so that a 3D shape may be generated from values for features in the skip connections and values for bottleneck features. An encoder (e.g., a topographic encoder) or multiple encoders (e.g., a topographic and mask encoder) are trained in conjunction (e.g., jointly) with a 3D shape and / or mask encoder that is not used in the application.
[0070] In response to the input, the generation network outputs a synthetic estimate of the 3D shape of the object. Any parameterization of the shape may be used, such as a voxel, mesh, or continuous surface representation. In an alternative embodiment, the 3D shape is output as an animation of a 3D rendering of the object (e.g., a sequence of 2D views rendered from different viewing directions).
[0071] In action 44, an image processor generates an image and a display displays the image. The image includes information from the 3D shape. The image is generated as a visual representation of the 3D shape. The image is generated using volume rendering or another 3D rendering. Alternatively or additionally, the image includes alphanumeric text, graphics, or another indicator of a quantity characterizing the 3D shape, such as a measure of the volume or surface smoothness or variation of the 3D shape.
[0072] The topogram and / or mask may be included in the image or displayed simultaneously with the information. Other patient information may also be displayed.
[0073] The image can be output to a display, patient medical record, and / or report. Information from the 3D shape can be used for diagnosis, prognosis, treatment planning, or other purposes.
[0074] In one embodiment of training and application, , , as well as , if a mask is provided as input, or (i.e., for topogram-only methods). All models were trained using the Adam optimizer with a learning rate of 0.0001 over 250 epochs and a batch size of 32. The training data was for the task of estimating the 3D shape and predicting the volume of the human liver. Automatic liver segmentation is a very complex problem due to its heterogeneous and diffuse shape. In less complex methods, the 3D shape of the liver is estimated from 2D topogram images and (optionally) a 2D mask. Voxel grids are used as the base representation and the results are visualized using either 2D projections or 3D meshes obtained using marching cubes. The impact of the shape context provided by the mask observations is studied by evaluating a baseline where the 3D shape is predicted directly from the mask.
[0075] For experimental evaluation, 2129 abdominal CT scans (e.g., 3D volumetric images of the abdomen covering the liver) were collected from several different hospital sites. Liver shapes were segmented from the CT scans using a volumetric segmentation method, and topograms and masks were extracted via 2D projections. Example renderings of 3D shapes from the dataset are shown in Figure 5 1554 scans were used for training and 575 scans were used for testing.
[0076] Given a learned generative model of liver shape including an image encoder that estimates a latent space vector given a topogram image (and mask, if given) - a 3D liver shape is output. This 3D shape is projected back onto the topogram image plane to perform a 2D delineation to test accuracy. Visually delineating the exact shape from the topogram is particularly difficult due to visual ambiguities such as color contrast and blurred boundaries. Using Figure 2 or Figure 3 A trained architecture that can automatically predict 3D shapes from localizers. Given 2D mask annotations, the predictions can be refined or made more accurate.
[0077] Figure 6 A qualitative assessment is shown. Figure 6In , the 3D reconstruction results are visualized. The first column is a visualization of the input topogram, the second column is a visualization of the ground truth 3D shape (surface or volume rendering of the 3D shape), and the third column is a visualization of the result of the topogram-only method (i.e., using Figure 2 The fourth column is a visualization of the results of the patch+mask method (i.e., using Figure 3 ), and the fifth and sixth columns are visualizations of the projection masks for the two methods, overlaid with the ground truth masks. Each row corresponds to a different example.
[0078] The two proposed (i.e., using Figure 2 and Figure 3 The trained) method is able to capture significant variations in the observed shapes, such as Figure 6 The prominent dome on the right lobe of Example 1 and Figure 6 The shape of the left lobe in Example 5 of . Compared with the topogram-only method, the topogram+mask method is able to convey more topological details, such as the elongated internal tip in Examples 1 and 4, the protrusion of the left lobe in Examples 2 and 3, and the overall topology in Example 5, where the mask-based method corrects the hole artifacts introduced by the topogram-only method. Overall, the 3D surface from the prediction of the mask-based method is visually closer to the ground truth.
[0079] Figure 7 The projection of the 3D shape prediction directly onto the input topogram is shown. The ground truth projection is shown in the first row, the topogram-only prediction projection is shown in the second row, and the topogram+mask projection is shown in the third row. By predicting the 3D shape of the organ, an accurate 2D segmentation of the input topogram via the projected image can be generated and displayed or used for computational quantities. The displayed image allows visualization of the corresponding inferred 2D segmentation. The shape reconstruction network (only in both the topogram and topogram+mask methods) learns to emphasize characteristic parts of the organ shape, such as the curves in the right lobe and the inner tip.
[0080] In quantitative evaluation, several metrics can be used to quantitatively compare 3D shape reconstructions. Figure 2 Machine learning generative model (in the application, only the localization slice), Figure 3 The machine learning generative model of the Figure 2 , where a mask encoder 23 is used instead of a topogram encoder 24, learning to reconstruct the 3D shape directly from the mask without providing a topogram image). The volume-based intersection over union (IoU) metric and the Dice coefficient are used. The surface-based Hausdorff distance metric is used. Table 1 shows the results:
[0081] .
[0082] The topogram+mask method outperforms the topogram-only method by all metrics, but especially by the Hausdorff distance, which is very sensitive to shape variations, such as the critical case of incorrect tip or protrusion presence predictions.
[0083] The mask-only approach can be used to show whether the provided mask provides too much context, rendering the problem of 3D shape prediction a much easier task. In Table 1, the performance of this baseline and two methods that receive a topogram as input are compared. The mask-only approach fails to achieve the same quality of results as the topogram-based approach, generating significantly lower mean IoU and Dice error, and much larger Hausdorff error. The topogram image contains important information such as shape layout, which is complementary to the context extracted from the mask, and thus both inputs can be used for high-quality reconstruction.
[0084] Automatic volume measurement of major organs is often important for medical analysis (e.g., diagnosis, prognosis, or planning). The predicted 3D shape can be used to directly measure the organ volume. The organ volume is given from the topogram. Table 2 shows a comparison of volume errors.
[0085] .
[0086] The volume error is the difference of the proposed method to the volume of the voxelized 3D segmentation of the liver, which is obtained from the segmentation of 3DCT as the ground truth. Given the 3D shape prediction, the predicted volume is measured as the number of voxels in the generated shape. The number of voxels is converted to milliliters (mL) using the scan configuration parameters, which indicate the volume of each voxel. The volume error prediction used is ,in and are the predicted organ volume and the ground truth organ volume, respectively.
[0087] On average, the error in predicting liver volume using the topogram + mask method was 6%, and the error in predicting liver volume using the topogram-only method was 10%. The mask-based method alone could not accurately predict the volume because it could not predict the correct 3D topology.
[0088] The above qualitative and quantitative comparisons are based on generative networks using machine learning. By including a discriminator in adversarial training, the same architecture using the same training data can be trained as a GAN (e.g., 3D VAE-GAN). Compared to VAE-based methods without a discriminator, the discriminator in this baseline typically encourages more consistent predictions, thereby discouraging the generation of more diverse shape topologies. Quantitatively, without a discriminator in training, GAN achieves lower quality results than both VAE-based methods. Table 3 shows the difference in results for volumetric quantities and 3D shape error.
[0089] .
[0090] 3D-GAN methods suffer from surface-based errors and volumetric errors due to their tendency to predict average shapes without considering the input.
[0091] Reconstructing 3D organ shape from topograms is a very challenging problem in medical imaging. Among other challenges, this is a difficult problem because, in addition to predicting the topology of occluded and invisible parts of the 3D organ, the input X-ray image may also contain projection artifacts that the reconstruction method needs to deal with. Despite the visual ambiguity in this type of image, it is possible to predict the 3D organ shape directly from the topogram. It is also possible to improve the quality of the prediction by providing supplementary 2D shape information in the form of a mask. By using the 3D shape ground truth as input to define the shape space in training, an accurate 3D shape estimate is provided.
[0092] Figure 8 A medical imaging system for 3D organ estimation is shown. The medical imaging system includes a display 80, a memory 84, and an image processor 82. The display 80, the image processor 82, and the memory 84 may be part of a medical imager 86, a computer, a server, a workstation, or other medical system for image processing of medical images from patient scans. A workstation or computer without a medical imager 86 may be used as a medical imaging system.
[0093] Additional, different or fewer components may be provided. For example, a computer network is included for remote 3D shape generation of a locally captured topogram. As another example, a user input device (e.g., a keyboard, button, slider, dial, trackball, mouse or other device) is provided for user mask input or creation from a displayed topogram.
[0094] Medical imager 86 is a medical diagnostic scanning system that is configured to generate projection (i.e., localizer) images of a patient. In one embodiment, medical imager 86 is an X-ray imager for generating X-ray images of a patient. An X-ray source and a detector are positioned on opposite sides of at least a portion of patient 88. The X-ray source generates X-rays that pass through patient 88 and irradiate the detector. The X-ray imager reads the X-ray intensity at different positions from the detector. The reading at each position represents the contribution of tissue from the patient's body along the ray from the X-ray source to that position, providing a projection of the folded third dimension. Other projection imagers may be used.
[0095] The medical imager 86 is configured to generate an X-ray image as a representation of a volumetric object within a patient in two dimensions of the projection. The configuration uses settings of one or more parameters, such as X-ray source voltage, table positioning and / or range of motion, gantry positioning and / or range of motion, focus, field of view, collimation, detector thresholds, image processing settings, filter settings, and / or image generation settings.
[0096] The image processor 82 is a control processor, a general purpose processor, a digital signal processor, a 3D data processor, a graphics processing unit, an application specific integrated circuit, a field programmable gate array, a digital circuit, an analog circuit, an artificial intelligence processor, a combination thereof, or other devices now known or later developed for processing medical image data such as localizers. The image processor 82 is a single device, a plurality of devices, or a network of devices. For more than one device, parallel or sequential processing partitioning can be used. The different devices that make up the image processor 82 can perform different functions, such as applying a machine learning model to an X-ray image by one device and rendering a view of a 3D shape output from the application by another device. In one embodiment, the image processor 82 is a control processor or other processor of a medical diagnostic imaging system such as a medical imager 86. The image processor 82 operates according to stored instructions, hardware, and / or firmware to perform the various actions described herein.
[0097] The image processor 82 is configured to train a machine learning architecture. Based on a user-provided or other source of network architecture and training data, the image processor 82 learns features of encoders, decoders, discriminators, or other network components to train the model. The result of the machine training is one or more machine learning models for 3D shape prediction with or without mask parameterization.
[0098] Alternatively or additionally, the image processor 82 is configured to apply one or more machine learning models. The machine learning models are applied as standalone applications on a workstation or local device or as a service deployed on a computer network (cloud) architecture.
[0099] The machine learning model generates an estimate of the 3D shape of the patient's organ or anatomical structure in response to the input of the X-ray image or the X-ray image and the mask. For example, the machine learning generation network is applied to the X-ray image or the X-ray image and the mask of the patient 88. The machine learning model includes input channels and corresponding modules (e.g., encoders) for the X-ray image and the mask (if provided). When trained, the model includes input channels and modules (e.g., encoders) for the 3D shape and / or for the mask for regularization. Then, in the absence of one or more encoders used in the training, the resulting X-ray image encoder with or without the mask encoder is used in the application.
[0100] A single X-ray image can be used to predict a 3D shape. A single X-ray image or an X-ray image and a mask image derived from the X-ray image are input to generate an estimate of the 3D shape of an organ or other object. The input is all views of the patient 88 from the same direction, such as using the same projection lines that are parallel or as diverge from a common source. Views from a single direction are input to estimate the 3D shape. In an alternative embodiment, without tomography, views from different directions are input to estimate the 3D shape.
[0101] Based on previous training, the machine learning model generates a 3D shape in response to application of an X-ray image with or without a mask image.The image processor 82 may be configured to compute one or more quantities, such as volume, from the 3D shape.
[0102] The image processor 82 is configured to generate an image. The 3D shape output from the machine learning model can be volume rendered or surface rendered to create an image. Alternatively, annotations or graphics such as quantities determined from the 3D shape can be generated images (e.g., displaying a report with quantities) or can be added to the X-ray image. In another alternative, the image processor 82 projects from the predicted 3D shape to create a segmentation or mask. The projection is displayed or used to segment the X-ray image, where the segmentation is displayed as an image.
[0103] The display 80 is a CRT, LCD, projector, plasma, printer, tablet device, smart phone or other display device known now or developed later for displaying the generated image. The display 80 displays the medical image generated based on the information from the predicted 3D shape, such as the information of the estimated 3D organ.
[0104] X-ray images, masks, network definitions, features, machine learning models, training data, output 3D shapes, information derived from 3D shapes, images, and / or other information are stored in a computer-readable memory such as memory 84. Memory 84 is an external storage device, RAM, ROM, database, and / or local memory (e.g., a solid-state drive or hard drive). The same or different non-transitory computer-readable media can be used for instructions and other data. Memory 84 can be implemented using a database management system (DBMS) and resides on a memory such as a hard disk, RAM, or removable media. Alternatively, memory 84 is internal to processor 82 (e.g., cache).
[0105] Instructions for implementing the training or application processes, methods, and / or techniques discussed herein are provided on a non-transitory computer-readable storage medium or memory, such as a cache, buffer, RAM, removable media, hard drive, or other computer-readable storage medium (e.g., memory 84). Computer-readable storage media include various types of volatile and non-volatile storage media. In response to one or more instruction sets stored in or on a computer-readable storage medium, the functions, actions, or tasks illustrated in the figures or described herein are performed. The functions, actions, or tasks are independent of a particular type of instruction set, storage medium, processor, or processing strategy, and may be performed by software, hardware, integrated circuits, firmware, microcode, etc., operating alone or in combination.
[0106] In one embodiment, the instructions are stored on a removable media device for reading by a local or remote system. In other embodiments, the instructions are stored in a remote location for transmission over a computer network. In yet other embodiments, the instructions are stored within a given computer, CPU, GPU, or system. Because some of the constituent system components and method steps depicted in the accompanying drawings can be implemented in software, the actual connections between the system components (or processing steps) may vary depending on the manner in which the present embodiment is programmed.
[0107] The various improvements described herein may be used together or separately. Although illustrative embodiments of the present invention have been described herein with reference to the accompanying drawings, it should be understood that the present invention is not limited to those precise embodiments, and that various other changes and modifications may be made therein by those skilled in the art without departing from the scope or spirit of the present invention.
Claims
1. A method for reconstructing a three-dimensional shape from a patient positioning film (21) in a medical imaging system, the method include: Acquiring (40) a patient positioning film representing a projection through the patient in two-dimensional form; Reconstructing (42) a three-dimensional shape of an object represented in a patient positioning film (21), the reconstruction (42) being performed by a machine learning generative network in response to inputting the patient positioning film (21) into the machine learning generative network, wherein the machine learning generative network comprises: a mask encoder configured to predict a first latent coordinate of an organ shape given a 2D mask; a 2D image encoder configured to predict a second latent coordinate given the 2D topogram image; a combiner configured to map outputs of the mask encoder and the 2D image encoder to common latent coordinates; a 3D shape encoder configured to input a 3D shape, wherein the 3D shape encoder is used as a regularizer in training the machine learning generative network and is not used in application of the machine learning generative network; and a shape decoder configured to input the common latent coordinates and output a three-dimensional shape; and Display (44) information from the three-dimensional shape.
2. The method of claim 1, wherein acquiring (40) comprises acquiring (40) using an X-ray imager (86).
3. The method of claim 1, wherein reconstructing (42) comprises reconstructing (42) the three-dimensional shape as an outer surface of the object.
4. The method of claim 1, wherein acquiring (40) comprises acquiring (40) an X-ray image as a patient positioning film (21), the X-ray image comprising a projection of an object, the object comprising an organ, and wherein reconstructing (42) comprises reconstructing (42) the organ of the patient.
5. The method of claim 1, wherein reconstructing (42) comprises reconstructing (42) by a machine learning generative network that has been trained as a generative adversarial network.
6. The method of claim 1, wherein reconstructing (42) comprises reconstructing (42) via a machine learning generative network based on variational autoencoding, the machine learning generative network comprising an encoder (24) and a decoder (28).
7. The method of claim 1, wherein displaying (44) comprises displaying (44) a three-dimensional rendering of a three-dimensional shape of the object.
8. The method of claim 1, wherein displaying (44) includes displaying (44) quantities determined from the three-dimensional shape.
9. A method for machine training for reconstructing a three-dimensional representation of an anatomical structure from a first topogram in a medical imaging system, the method include: defining (10) a machine learning (16) architecture comprising a mask encoder configured to predict a first latent coordinate of an organ shape given a 2D mask; a 2D image encoder configured to predict a second latent coordinate given the 2D topogram image; a combiner configured to map outputs of the mask encoder and the 2D image encoder to common latent coordinates; A 3D shape encoder configured to input a 3D shape; as well as a shape decoder configured to input the common latent coordinates and output a three-dimensional shape; performing machine learning (16) using a machine learning (16) architecture to output a three-dimensional representation (29) in response to an input of a first topogram, wherein the 3D shape encoder is used as a regularizer in training the machine learning generative network and is not used in applying the machine learning generative network; and The machine learning model generated by the machine learning (16) is stored (18).
10. The method of claim 9, wherein defining (10) comprises defining (10) the machine learning (16) architecture as a generative model based on variational autoencoding.
11. A medical imaging system for three-dimensional organ estimation, the medical imaging system include: an X-ray imager (86) for generating an X-ray image (21) of the patient; An image processor (82) configured to generate an estimate of a three-dimensional organ based on an input of an X-ray image (21) to a machine learning model, the machine learning model being configured to receive as input a view from a single direction, wherein the machine learning model comprises: a mask encoder configured to predict a first latent coordinate of an organ shape given a 2D mask; a 2D image encoder configured to predict a second latent coordinate given the 2D topogram image; a combiner configured to map outputs of the mask encoder and the 2D image encoder to common latent coordinates; a 3D shape encoder configured to input a 3D shape, wherein the 3D shape encoder is used as a regularizer in training the machine learning generative network and is not used in applying the machine learning generative network; and a shape decoder configured to input the common latent coordinates and output a three-dimensional shape, the X-ray image (21) being a view from a single direction; and A display (80) is configured to display information derived from the estimate of the three-dimensional organ.
