阅读说明：本技术 用于从厚图像切片产生薄图像切片的系统和方法 (System and method for generating thin image slices from thick image slices ) 是由 方中南 A·S·乔杜里 李镇亨 B·A·哈格里夫斯 于 2019-03-12 设计创作，主要内容包括：本文中公开用于从厚切片图像产生薄切片图像的系统和方法。在一些实例中,深度学习系统可从厚切片图像计算残差并且将所述残差加入到所述厚切片图像以产生薄切片图像。在一些实例中,所述深度学习系统包含神经网络。在一些实例中,所述神经网络可包含一或多个层级,其中所述层级中的一或多个包含一或多个块。在一些实例中,每一层级包含卷积块和非线性激活函数块。在一些实例中,所述神经网络的所述层级可为级联式布置。(Systems and methods for generating thin-slice images from thick-slice images are disclosed herein. In some examples, the depth learning system may calculate a residual from a thick-slice image and add the residual to the thick-slice image to generate a thin-slice image. In some examples, the deep learning system includes a neural network. In some examples, the neural network may include one or more levels, where one or more of the levels includes one or more blocks. In some examples, each level includes a convolution block and a non-linear activation function block. In some examples, the hierarchy of the neural network may be a cascaded arrangement.)

1. A method of generating a thin-slice image from a thick-slice image, the method comprising:

Receiving, at a neural network, a first image having a first resolution;

performing a convolution on the first image through the neural network;

performing a non-linear activation function on the first image through the neural network;

repeating the convolution and nonlinear activation functions;

generating a residual based on the convolution; and

summing, by the neural network, the residual with the first image to produce a second image having a second resolution, wherein the second resolution is higher than the first resolution.

2. The method of claim 1, wherein the performing the convolution and performing the nonlinear activation function are performed and repeated in a plurality of layers of the neural network, wherein the first image is an input to a first layer of the plurality of layers, an output of the first layer of the plurality of layers is an input to a second layer of the plurality of layers, and an output of a last layer of the plurality of layers is the residual.

3. The method of claim 1, further comprising training the neural network on a training data set, wherein the training data set includes a plurality of first images and a plurality of second images.

4. The method of claim 1, further comprising testing the neural network on a test data set, wherein the test data set includes a plurality of first images.

5. The method of claim 1, wherein performing the convolution comprises performing a three-dimensional convolution.

6. The method of claim 1, wherein training the neural network includes dividing the first image into a plurality of patches of pixels.

7. The method of claim 1, wherein the nonlinear activation function comprises a form of r (x) max (0, x).

8. The method of claim 1, further comprising acquiring the first image from an imaging system.

9. The method of claim 8, wherein the imaging system is a magnetic resonance imaging system.

10. A system for generating a thin-slice image from a thick-slice image, the system comprising:

a non-transitory computer-readable medium including instructions for implementing a neural network, wherein the neural network includes a hierarchy including a rolling block and a modified linear cell nonlinear activation block, wherein the hierarchy is configured to generate a residual from a first image having a first resolution received by the neural network, wherein the neural network is configured to sum the first image and the residual to generate a second image having a second resolution, wherein the second resolution is higher than the first resolution; and

A processor configured to execute the instructions to implement the neural network.

11. The system of claim 10, further comprising a display configured to display the second image.

12. The system of claim 10, wherein the convolution block applies a three-dimensional convolution and thresholding using a modified linear unit function to the first image.

13. The system of claim 10, wherein the neural network includes a plurality of levels, wherein an output of a first level of the plurality of levels is provided as an input to a second level of the plurality of levels.

14. The system of claim 13, wherein a last level of the plurality of levels does not include the modified linear cell nonlinear activation block.

15. The system of claim 13, wherein a last level of the plurality of levels has a smaller size than other levels of the plurality of levels.

16. The system of claim 10, wherein the neural network divides the first image into a plurality of patches of pixels.

17. The system of claim 16, wherein the plurality of pixel tiles overlap.

18. The system of claim 17, wherein the plurality of pixel tiles overlap by 50%.

19. The system of claim 10, wherein the convolution block applies a zero-padding convolution and an output of the zero-padding convolution is cropped to an original size of the input image.

20. The system of claim 10, wherein the convolution block outputs a plurality of feature maps.

21. A system for generating a high resolution image from a low resolution image, the system comprising:

an image acquisition unit configured to acquire a first image of a feature of interest at a first resolution;

a computing system configured to implement a deep learning system, wherein the deep learning system is configured to receive the first image of the feature of interest and generate a second image of the feature of interest at a second resolution based at least in part on the first image, wherein the second resolution is higher than the first resolution; and

a display configured to display the second image of the feature of interest.

22. The system of claim 21, wherein the image acquisition unit is a magnetic resonance imaging system.

23. The system of claim 21, wherein the deep learning system generates the second image by supplementing the first image.

24. The system of claim 21, wherein the computing system is configured to implement a second deep learning system and the second image of the feature of interest is provided as an input to the second deep learning system.

Technical Field

Examples described herein generally relate to processing image slices from three-dimensional images. More specifically, examples described herein relate to processing thick image slices from a three-dimensional image to generate thin slices of the three-dimensional image. In some examples, thin slices can be analyzed, observed, or otherwise used to diagnose and/or treat a disease.

Background

In three-dimensional (3D) imaging modalities such as Magnetic Resonance Imaging (MRI), Computed Tomography (CT), Positron Emission Tomography (PET), and Ultrasound (US), a volume is imaged as a series of imaging planes (e.g., slices) having finite thicknesses (e.g., 0.5mm, 1mm, 0.5cm, 1 cm). The slices are combined to generate a 3D image of the volume, where each voxel of the 3D image corresponds to a portion of the imaged volume. The resolution of the 3D image may be based at least in part on the magnitude of the volume represented by each voxel. The larger the volume represented by the voxel, the lower the resolution of the 3D image. One technique for improving the resolution of 3D images is to acquire more slices within a volume. That is, the thickness of the slice is reduced and the volume represented by each voxel is reduced. For example, the volume imaging as a series of 0.5mm slices may produce higher resolution 3D images than when the volume imaging is a series of 1.0mm slices.

Higher resolution images may provide more information to the viewer than lower resolution images. For example, a clinician may be able to make more accurate diagnoses from medical images having higher resolutions than those having lower resolutions. However, in some applications, it may not be feasible to acquire thinner slices of the volume. For example, some imaging systems may be limited to a minimum slice thickness that cannot be reduced any more. In some applications, acquiring thinner slices may require additional imaging time. This may reduce patient comfort, for example, where the patient must stay in tight space for a long period of time and/or must hold his breath during the acquisition of slices. For some imaging modalities, acquiring thinner slices may require additional exposure to ionizing radiation, which may increase the health risk of the patient.

Therefore, there is a need to improve the resolution of 3D images without reducing the slice thickness acquired during imaging.

Disclosure of Invention

Examples of the techniques described herein may provide super-resolution techniques that use a deep learning system to generate thin-slice 3D images (e.g., MRI, fMRI, PET, US, and/or CT) from thicker input slices and compare this method to alternative trans-planar interpolation methods.

In accordance with some examples of the present disclosure, a deep learning system is described that includes a 3D convolutional neural network configured to learn a residual-based transform between a high-resolution thin-slice image and a lower-resolution thick-slice image. In some examples, the training neural network learns a residual-based transformation between the high-resolution thin-slice image and the lower-resolution thick-slice image at the same central position. Although a 3D musculoskeletal MR image is provided as an example, this method is generally applicable to any 3D image and imaging of any region of the human or animal body, including the brain, liver, cardiovascular system, etc.

A neural network (e.g., a 3D convolutional neural network) according to the principles of the present disclosure may be capable of resolving high resolution thin-slice MRI from lower resolution thicker slices. In some applications, the neural network can achieve superior quantitative and quality diagnostic performance while reducing the number of slices that need to be acquired compared to conventionally utilized and state-of-the-art methods for MRI, CT, and/or other 3D imaging modalities.

A method of generating a thin-slice image from a thick-slice image according to an example of the present disclosure may include receiving, at a neural network, a first image having a first resolution; performing a convolution on the first image through the neural network; performing a non-linear activation function on the first image through the neural network; repeating the convolution and nonlinear activation functions; generating, by the neural network, a residual based on another convolution; and summing, by the neural network, the residual with the first image to produce a second image having a second resolution, wherein the second resolution is higher than the first resolution.

A system of generating a thin-slice image from a thick-slice image according to an example of the present disclosure may include a non-transitory computer-readable medium including instructions for implementing a neural network, wherein the neural network includes a hierarchy including a volume block and a modified linear cell nonlinear activation block, wherein the hierarchy is configured to generate a residual from a first image having a first resolution received by the neural network, wherein the neural network is configured to sum the first image and the residual to generate a second image having a second resolution, wherein the second resolution is higher than the first resolution; and a processor configured to execute the instructions to implement the neural network.

A system for generating a high resolution image from a low resolution image according to examples of the present disclosure may include an image acquisition unit configured to acquire a first image of a feature of interest at a first resolution; a computing system configured to implement a deep learning system, wherein the deep learning system is configured to receive the first image of the feature of interest and generate a second image of the feature of interest at a second resolution based at least in part on the first image, wherein the second resolution is higher than the first resolution; and a display configured to display the second image of the feature of interest.

Drawings

Fig. 1 is a block diagram of a system arranged in accordance with examples described herein.

Fig. 2 is an illustration of a neural network arranged in accordance with an example described herein.

Fig. 3 is an illustration of a portion of the neural network shown in fig. 2.

Fig. 4 is an illustration of a portion of the neural network shown in fig. 2 and 3.

Fig. 5 illustrates an example workflow implementing a neural network according to examples described herein.

Fig. 6 is an exemplary coronal image with different downsampling factors and corresponding 2D structural similarity maps according to examples described herein.

Fig. 7 illustrates a comparison of an example embodiment of the present disclosure with other resolution enhancement techniques.

FIG. 8 is a set of simulated base-true (ground-truth) images and resulting generated images according to an example described herein.

Fig. 9 shows an example image of a horizontal tear in the lateral meniscus body according to examples described herein.

Fig. 10 shows an example image of a mild case of grade 2A chondromalacia of the lateral patellar cartilage according to the examples described herein.

Fig. 11 is an illustration of a system arranged in accordance with examples described herein.

Detailed Description

Certain details are set forth below to provide a sufficient understanding of the described embodiments. It will be apparent, however, to one skilled in the art that embodiments may be practiced without these specific details. In some instances, well-known brain imaging techniques and systems, circuits, control signals, timing protocols, and/or software operations have not been shown in detail in order to avoid unnecessarily obscuring the described embodiments.

Due to hardware limitations in modalities such as Magnetic Resonance Imaging (MRI), functional magnetic resonance imaging (fMRI), Positron Emission Tomography (PET), and Computed Tomography (CT), fast acquisition of high resolution 3D medical images with sufficient signal-to-noise ratio is challenging.

In MRI, clinical musculoskeletal imaging protocols typically include 2D Fast Spin Echo (FSE) sequences that are scanned in various scan planes, typically with the same contrast. While such sequences provide excellent in-plane resolution, there is always an associated risk of missing fine lesions due to partial volume effects in slices with high cross-sectional thickness (typically 2.5-4mm) and with slice gaps. Image acquisition of thick slices also prevents image reconstruction (reconstruction) into arbitrary scan planes, which makes it challenging to interrogate tissue (e.g., the anterior cruciate ligament of the femoral trochlear and articular cartilage) in an oblique orientation.

Several methods for low-profile musculoskeletal MRI have recently been proposed, with 3D fast spin echo (using the vendor products SPACE, CUBE or VISTA) being the popular choice. However, 3D FSE sequences typically compromise in-plane resolution for through-plane resolution and suffer from significant echo chain ambiguity problems, which are not conducive even with the variable flip angle approach. Thus, even if low profile images are acquired, the overall diagnostic quality of the 3D FSE may be limited. The 3D radial sequence is used for isotropic musculoskeletal imaging, but has not been widely adopted due to the different appearance of artifacts. In addition to sequence-based methods, Compressive Sensing (CS) and parallel imaging are promising methods for accelerating low-profile MRI, but are not ideal stand-alone methods. CS MRI is not common in 2D FSE clinical protocols because only one dimension is available to produce incoherent sampling, and long reconstruction times are also required. Parallel imaging acceleration is ultimately limited by SNR and g-factor. In addition, the geometry of common knee coils (e.g., rigid 8-channel) is not suitable for acceleration in all directions.

Single image super-resolution (the active field in image processing) may make it possible to increase MRI spatial resolution to produce low-profile MRI without compromising SNR or requiring additional MRI hardware or scan time. For example, interpolation is an original implementation of super-resolution. Due to the challenges of actually acquiring low profile imaging, MRI vendors provide retrospective slice interpolation using zero-filled Fourier Interpolation (FI) through options such as 'ZIP 2' and 'interpolation' for general electric medical group (GE Healthcare) and Siemens Healthcare (Siemens Healthcare) MRI scanners, respectively. Similarly, medical image viewing platforms such as OsiriX use tri-linear interpolation during image manipulation and multi-planar reconstruction of data. Low-order linear interpolation methods, such as FI and trilinear interpolation, are widely used in clinical and research protocols that attempt to achieve low-profile imaging, however, they do not produce images with high diagnostic quality.

In addition to interpolation, another major approach to MRI super-resolution imaging requires the exploitation of image sparsity, requiring only a single image as input. The most advanced MRI single image super resolution algorithm at present is based on sparse coding super resolution (ScSR), which was originally developed for natural images but later adapted for a few MRI applications. Although promising, this super-resolution approach alone has not been universally applied to medical imaging due to limited resolution improvement and slow execution speed of 3D data.

In accordance with the principles of the present disclosure, a deep learning system may be configured to apply super-resolution techniques for use with natural 2D images to 3D medical images. In some examples, the deep learning system may be implemented using an artificial intelligence system, a machine learning system, a neural network (e.g., a convolutional neural network), and/or other computing techniques. In some instances, a software-based post-processing neural network may overcome, at least in part, the limitations of 3D medical imaging just described. In some examples, the neural network may be a convolutional neural network. In some examples, the neural network may employ deep learning techniques. Because the images are used to train and deploy the neural network, the neural network is independent of the image acquisition hardware platform.

An example describing a musculoskeletal MRI will be used to explain the architecture of the deep learning system and its training and implementation methods. However, deep learning systems are generally applicable to different 3D medical images including MRI, fMRI, PET and CT. The application area is not limited to musculoskeletal imaging and encompasses images covering any area of, for example, the brain, liver, and cardiovascular system.

Described examples include systems that generate thin slice images using deep learning for MRI super-resolution and maintain high in-plane resolution to reduce total scan time. As used herein, "thick" and "thin" are used as relative terms. That is, the size of a slice referred to as "thick" in at least one dimension (e.g., elevation, azimuth, lateral) is larger than the size of a "thin" image in the corresponding dimension. The resolution of the image produced by the slice may be based on the magnitude of the volume represented by the voxels in the image. The larger the volume represented by the voxel, the lower the resolution of the image. An image generated from a thick slice may have a larger volume associated with each voxel than an image generated from a thin slice. The proposed deep learning system and method may be referred to as 'deep resolve' because it facilitates resolving high resolution features from low resolution inputs. The deepsolution may include a convolutional neural network for deep learning (e.g., machine learning). In particular, the examples described herein train neural networks using publicly available data sets to generate high resolution thin slice knee MR images from slices at the same orientation but with slice thicknesses 2-8 times higher. The deep learning system may not have to produce the same image as the underlying reality; rather, in some applications, the deep learning system may enhance the low resolution images to make them more similar to the underlying reality (e.g., high resolution images) than currently state-of-the-art methods that are typically utilized. This may enhance the diagnostic value of images acquired from thick slices in some applications.

Fig. 1 is a schematic illustration of a system arranged in accordance with examples described herein. The system 100 includes an image slice (e.g., a thick slice) 102, a training data set 104, a computing system 106, a processor 108, executable instructions (also referred to herein simply as executable instructions) for generating a thin slice from an input image slice 110, a memory 112, a display 114, and a network interface 116. Additional, fewer, and/or other components may be used in other examples. In some examples, some or all of the components of the system 100 may be included with an imaging system. For example, the computing system 106 may be included in an imaging system. For example, the imaging system may be an MRI, CT, and/or ultrasound imaging system.

The image slices 102 and/or the training data set 104 may be provided by an imaging system. In some examples, the image slices 102 and/or the training data set 104 may be provided by multiple imaging systems. The imaging systems may have the same modality or different modalities. The image slices 102 and/or the training data set 104 may be stored in a memory that is accessible to the computing system 106 and/or transmitted to the computing system 106 (e.g., using wired or wireless communication). The computing system 106 may be configured to generate a deep learning system based on the training data set 104. The deep learning system may include one or more functions for generating thin-slice images from the image slices 102. The function may be generated and/or optimized based on the training data set 104.

The training dataset 104 may train the deep learning system to find the difference (e.g., residual) between the thick-slice image and the corresponding thin-slice image. After training, the deep learning system may generate a thin-slice image from the input image slice 102 by applying the function to the image slice 102. The function may augment the image slice 102 to produce a thin slice (e.g., high resolution) image. In some examples, the deep learning system may calculate the difference between the input image slice 102 and the desired high resolution image (e.g., a thin slice image) and add the difference to the input image slice 102 to generate the high resolution image. In some examples, the deep learning system may include one or more neural networks. In some examples, one or more of the neural networks may be convolutional neural networks. One or more of the neural networks may include a plurality of layers, with each layer including one or more functions or function portions for generating thin-slice images from the image slices 102.

The computing system 106 may be configured to generate thin-slice (e.g., higher resolution) images from the image slices 102 through a deep learning system generated from the training data set 104. In some examples, the memory 112 may be encoded with executable instructions 110 for a deep learning system configured to generate a thin-slice image based on a thick-slice 102 input. In some examples, the deep learning system may be partially or fully implemented in hardware (e.g., ASIC, FPGA). In some examples, executable instructions 110 may include instructions of a hardware description language (e.g., VHDL), which may be used to design hardware to implement some or all of the deep learning system.

Examples described herein may utilize computing systems that may generally include hardware and/or software for implementing a deep learning system for generating thin-slice images. For example, the computing system 106 may include one or more processors 108. For example, the processor 108 may be implemented using one or more Central Processing Units (CPUs), Graphics Processing Units (GPUs), Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs), or other processor circuits. In some examples, the processor 108 may execute some or all of the executable instructions 110. The processor 108 may be in communication with a memory 112. Memory 112 may generally be implemented by any computer-readable medium, such as Read Only Memory (ROM), Random Access Memory (RAM), flash, solid state drive, etc. Although a single memory 112 is shown, any number of memories 112 may be used, and may be integrated with the processor 108 in a single computing system 106 and/or located within another computing system and in communication with the processor 108.

In some examples, system 100 may include a display 114, which may be in communication with computing system 106 (e.g., using a wired and/or wireless connection), or display 114 may be integrated with computing system 106. The display 114 may display one or more image slices 102 and/or one or more thin slices generated by the computing system 106. Any number or variety of displays may be presented, including one or more LEDs, LCDs, plasma, or other display devices.

In some examples, system 100 may include network interface 116. The network interface 116 may provide a communication interface with any network (e.g., LAN, WAN, Internet). The network interface 116 may be implemented using a wired and/or wireless interface (e.g., Wi-Fi, bluetooth, HDMI, USB, etc.). The network interface 116 may communicate data, which may include the image slice 102 and/or thin slice images generated by a deep learning system implemented by the computing system 106.

Referring to fig. 2-4, example deep learning systems are described. During training, a set of i fundamental real high resolution thin-slice images with the same field of view and matrix sizeAnd corresponding low resolution thick-slice images interpolated to thin-slice orientationsMay be provided as a training set. The deep learning system may include one or more functions. In an example implementation, the deep learning system includes a function f ═ f (X)_i) Which computes a residual image r between the base real image and the interpolated low resolution imageⁱ This function can be trained by optimization (using L2-loss):

in the implementation examples described herein, the L2 loss function is used to optimize the deep learning system. In other examples, additional loss functions may be used, such as L1-norm, pSNR, SSIM, and/or any other metric that measures image difference. Other optimization functions may be used in other examples.

In some examples, the deep learning system may be an unsupervised learning algorithm. In some examples, the functions generated and/or optimized by training, such as the functions in equation (1), may be implemented in one or more neural networks, which may have multiple layers. In some examples, the function may be implemented on multiple layers of a neural network. Each layer of the neural network may include one or more blocks. Each block may implement at least a portion of a function. In some examples, each layer of the neural network includes the same blocks (e.g., blocks that repeat in each layer), but in other examples, the layers of the neural network may include different blocks.

In an example implementation, the deep learning system may be modeled as a cascaded convolution filter. As illustrated in fig. 2, each level (e.g., layer) 202 of the deep learning system 200 may include a convolution block 204 and a modified linear cell nonlinear activation function block (ReLu)206, which will be described in more detail later. The deep learning system 200 can compute a residual (e.g., residual image) 208 from an input Low Resolution (LR) image 210 (e.g., a thick slice image) in order to generate a corresponding High Resolution (HR) image 212 (e.g., a thin slice image).

After the deep learning system has been trained, the estimated residuals may be computed directly

And can produce high resolution super resolution images by

(e.g., generating thin slice images):

an example implementation of a deep learning system includes a neural network with 19 layers of pairs of convolution 204 and modified linear unit (ReLU) nonlinear activation function blocks 206 (e.g., activation functions), and an additional layer including convolution blocks 207. However, other numbers of layers may be used in other examples. The input LR image 210 can pass through the first layer and be processed by the convolution block 204 and ReLu block 206. The processed input image may then be passed as input to the convolution block 204 and the ReLu block 206 of the next layer. This process is repeated until the input LR image 210 has been processed by all layers of the neural network and the residual 208 is output.

As illustrated in fig. 3, the filter sizes of the first and last layers of the neural network of an implementation example of the deep learning system have dimensions of 32x1, while all other layers have dimensions of 32x 64, with the first 3 numbers representing the x, y, z spatial dimensions and the last number representing the number of filters (also referred to as the number of filter channels). The last layer may not contain an activation function to ensure that the resulting residual contains positive and negative values. However, in other examples, the activation function may be included in the last layer. Continuing with this example, as illustrated in fig. 4, during training of the neural network, each input LR image 210 of the 3D data set is divided into isotropic 32x32x32 pixel patches 402 and all patches from all training data set images are used during the training process. During the application phase (e.g., when a trained neural network is used to generate the thin-slice image), the input image may not have to be divided into tiles.

In some examples, one or more layers of the deep learning system may be trained to identify one or more features. In some examples, different layers may identify different features. Feature recognition may be used as part of optimizing the function generated for each layer. In some examples, the features may be in the form of a feature map. As mentioned previously, the example deep learning system is implemented using a cascaded convolution filter paired with a nonlinear activation function. In the convolution block 204, an input image (32 × 32 × 32 pixel patch) may be convolved by a plurality of (e.g., 64) 3-dimensional filters having a size of 3 × 3 × 3 and a span of 1 × 1 × 1. Thus, in the first neural network layer, the input patch is transformed into a feature map having dimensions of 32 × 32 × 32 × 64. Subsequent layers, except the last layer in which feature reduction is performed to produce a residual image of size 32 x 32, maintain the same size. Convolution is performed on the zero-filled input and the output is cropped to the size of the original input in order to maintain the same input and output dimensions. Other filter arrangements may be used in other examples. For example, different sized input tiles may also be used as different sized feature maps. Other feature reduction and/or padding methods may be performed based at least in part on the size of the input image and the desired size of the output image.

Continuing the example implementation, in the ReLU block 206, a nonlinear activation function of the form r (x) ═ max (0, x) is used to ensure that the feature mapping output is a nonlinear representation of the input. The input low resolution patches are transformed by the concatenated convolution and ReLU pair 202, except for the last layer 207, which produces a residual image of size 32 x 32. In this example, the last (20 th) convolutional layer contains no ReLU activation to ensure that the residual image may contain positive and negative values. However, in other examples, the last layer may include the ReLU activation block 206. The residual image 208 is added to the input image to produce a near high resolution image. In some examples, the mean square error L2-loss may be compared to the original high resolution image to evaluate training. If the mean square error L2-loss is unacceptably high, the deep learning system may require additional training and/or additional or fewer layers may be added to the deep learning system in addition to further training.

In an example implementation, using a small convolution filter size of 3 × 3 × 3 in a deep learning system may avoid averaging high frequency images and convolution details together. A filter with larger support can be effectively decomposed into several smaller filters, introducing unnecessary redundancy during the network training process. Thus, a small convolution kernel may effectively improve and/or maximize high resolution feature extraction. In some examples, pooling layers may be utilized, but they may average high resolution details together. However, this may be desirable in some applications for memory optimization. In some examples, one or more of the feature maps may be zero-filled prior to convolution and clipped to the original input size after convolution in order to maintain the equivalent patch input and convolution output sizes. This may ensure that the detail at the edges of the patch maintains fidelity. Further, in some examples, the patches may overlap. For example, a 50% patch size patch overlap may be used during the deep learning system training phase. However, in some examples, the test data may be divided into large patches and not overlap to reduce the net effect caused at the edges of the patches.

An example implementation of a deep learning system has a depth of 20. The depth may be selected based at least in part on the convolution filter and image patch size used. For example, after the first 3 × 3 × 3 convolution, the receptive field for each weight in the second layer is 3 × 3 × 3. After the second 3 × 3 × 3 convolution, the total receptive field for the next weight is 5 × 5 × 5, which can be generalized as (2D +1)³Where D is the current network depth. Thus, for a patch size of 32 × 32 × 32, the receptive field of the weights in layer 16 corresponds to the features encountered in the entire image patch. Additional deeper layers may provide even higher levels of abstraction. Since super-resolution is inherently a morbid problem, providing spatial cues at different length scales by adding additional layers may be carefully optimized, which may be useful given a high-dimensional deep learning system and a super-parameter space. Increasing the number of deep learning systems may improve super resolution quality. Further, the deep learning model may follow a power law loss relationship, which may help determine an ideal training data size.

Referring back to fig. 1, the training described with reference to fig. 2-4 may be implemented by the processor 108 of the computing system 106 based on the training data set 104. The processor 108 may generate executable instructions 110 to implement a deep learning system as described with reference to fig. 2-4 and store the executable instructions 110 on the memory 112 of the computing system 106. The processor 108 may then execute the stored executable instructions 110 to implement a deep learning system. Once the executable instructions 110 are stored, the executable instructions 110 may be transferred to another computing system. That is, once the deep learning system is trained, it can be provided to other computing systems without retraining. In some examples, executable instructions 110 may be used to generate hardware to implement some or all of a deep learning system (e.g., ASIC), which may be provided to one or more computing systems for implementing the deep learning system.

Training, validation, and testing data for example embodiments of the deep learning system described herein were obtained from the Osteoarthritis Initiative (OAI). A total of 176 3D sagittal plane steady state Dual Echo (DESS) datasets from OAI were used for training. All images had double knee weight bearing radiographs to analyze kellogen-lorens (KL) osteoarthritis grade.

For training, thick-slice images are simulated from a training set of thin-slice images. The ratio of the base true slice thickness and the downsampled low resolution slice thickness is referred to as the downsampling factor (DSF). High resolution thin-slice image representation thick-slice representations are produced by anti-aliasing filtering the thin-slice images in the left-right direction, followed by downsampling. A1D finite impulse response low pass 48 th order Hamming windowing filter with a normalized passband of 1/DSF is generated for retrospective downsampling. This simulated acquisition of thicker slices has magnetization from surrounding slices. The training data set of the example implementation includes an underlying real high resolution image and a simulated low resolution image. In some examples, the low resolution image may be upscaled using 1D Triple Cubic Interpolation (TCI) in the slice direction (left and right) at the base true slice orientation. However, other methods of obtaining thick-slice images may be used. For example, a dataset that includes a thin-slice image and a thick-slice image taken from the same volume may be used to train a deep learning system.

Other data sets of different resolutions may be used in other examples. The training data set used to process a particular feature may contain previous patient images having that feature. For example, in gastrointestinal imaging, volumetric images of the liver and/or the liver including a lesion morphology of interest may be used to train a deep learning system to produce an improved resolution image of the liver.

In some examples, the training data may be pre-processed prior to training the deep learning system. In some applications, preprocessing the training data may improve the training process and result in a better trained deep learning system. Preprocessing may include formatting the data to ensure uniform size, cropping the image to include only relevant anatomical structures, uniformly scaling the intensity values, and/or subdividing into patches.

Thus, the deep learning system described herein may be trained using high resolution images (e.g., "base real images") and low resolution images that are simulated based on those high resolution images. The simulated low resolution image may be a downsampled version of the high resolution image. In some examples, after downsampling, the image may be interpolated such that the simulated low resolution image and the high resolution image (e.g., the "base real image") have the same or similar size. In this way, a simulated low resolution image set is provided that has a known association with a high resolution image set. The deep learning system described herein may be trained on a simulated low resolution image set (e.g., coefficients and other parameters for operating the deep learning system may be selected). The coefficients and other parameter sets may be selected to provide an accurate output of a "base real image" from the simulated low resolution image in general.

Fig. 5 illustrates an example workflow 500 in accordance with an example of the present disclosure. In an example implementation, the high resolution base real slice (a) is used to simulate the acquisition of slices (b) with higher cross-sectional thicknesses for different downsampling factors (DSFs). These slices are then interpolated triply to the base true slice orientation (c), and the residual function (d) identified by the training deep learning system may be added to the low-resolution thick slice to generate the high-resolution thin slice (e). During inference, a residual image for the test input is generated using the learned residual model. Thus, the residual may be added to the low resolution input in order to output a super resolution image. In other words, the deep learning system may receive as input a thick-slice image and generate as output a simulated thin-slice image. In some applications, the deep learning system may allow fewer image slices to be acquired while still maintaining sufficient resolution for diagnostic purposes. This may allow for reduced patient imaging time in some applications, which may in turn improve patient comfort and/or reduce patient exposure to ionizing radiation.

FIG. 6 illustrates exemplary coronal plane deep learning images with different downsampling factors (DSFs) (panes a-f) and corresponding 2D Structural Similarity (SSIM) mappings (panes g-l), according to an example of the present disclosure. The deep learning image and the SSIM mapping may be compared to the underlying real slice (e.g., the actually acquired thin slice image). Although images with higher DSF may be smoothed, the directionality of the blur is more pronounced. In the femoral bone marrow, pixels with incomplete fat saturation appear more blurred in the left-right direction than in the superior-inferior direction (dashed arrow). The medial collateral ligament can be a tissue to illustrate image fidelity in coronal reconstruction, which is that it is very thin in the left-right direction (solid arrow). SSIM mapping states that MCL is not reproduced with high diagnostic quality for DSFs 6 and 8. Coronal plane images with different DSFs and their SSIM mappings indicate that as DSFs increase, fine details of the Medial Collateral Ligament (MCL) and the Anterior Cruciate Ligament (ACL) become blurred (arrows on the SSIM mapping). In addition, residual femoral bone marrow signaling is due to incomplete fat suppression, manifested as excessive smoothing, reducing local SSIM. In this example, the sagittal image is least affected by blurring as DSF increases, and the axial and coronal reconstructions exhibit higher blurring as DSF increases.

A comparison of an example implementation of the deep learning system with other super resolution methods including TCI, FI, and ScSR methods for DSF 3x and the base real image is shown in fig. 7. An example coronal base real image (pane a) and a resolution enhanced image with a reduced sampling factor (DSF) of 3x at the same slice orientation for (pane b) deepsolution, (pane c) Fourier Interpolation (FI), (pane d) Triple Cubic Interpolation (TCI), and (pane e) sparse coding super resolution (ScSR) are shown. Between the comparison methods, the residual map (pane f) generated by the deep learning system and the difference image (e.g., difference map) are scaled by 5 times to emphasize subtle differences and show the underlying real image (pane g-j). Also shown (pane k-n) is a corresponding 2D Structural Similarity (SSIM) mapping for resolution enhanced images compared to the base real. The deep learning system of the present disclosure maintains the fine features in the coronal plane reconstruction and is most similar to the original underlying real image in terms of quality row visual quality and quantitative structural similarity measures. High resolution features such as MCL (solid arrow), small spurs on the lateral tibial plateau (dashed arrow), and inflammation with sharp features (dashed arrow) are easily visualized on the deep learning system image, however, visualization is far more challenging than other methods (a-e). This is further indicated by difference mapping (g-j) and pixel-by-pixel structural similarity mapping (k-n), where unlike the deep learning system of the present disclosure, TCI, FI, and ScSR all have lower similarity around the medial collateral ligament.

The deep learning system described herein may provide the ability to diagnose subtle lesions from thick slices. For example, three pairs of artificial meniscal and cartilage lesions with different signal intensities were simulated, and lesions subtle in space and signal intensity were generated manually.

Fig. 8 illustrates a simulated base real image and a resulting generated image according to an example of the present disclosure. The base real image (pane a) is modified to contain artificial meniscal lesions (solid arrows) and cartilage lesions (dashed arrows) of different signal intensities (panes b-d). The simulated meniscal tear is only two pixels thick in the horizontal and vertical directions. With the underlying real simulation as input, the 3x DSF show deep learning system (pane e-h) reproduces the lesion reasonably well and with lower blur compared to the TCI image (pane i-l). Compared to TCI, the deep learning system of the present disclosure has always better similarity measures (higher SSIM and pSNR, lower MSE) in the ground truth.

The example image of fig. 9 shows an example of a horizontal tear in a lateral meniscal body that can be identified by a high intensity DESS signal. Tears (arrows) can be found relatively similarly in sagittal plane basis real, DeepResolve, and TCI images (panes a-c). However, some blurring is evident in the retrofemoral cartilage where cartilage thickness is overestimated. Coronal reconstruction also indicated the same meniscal tear (pane d-f). The scaled slice (yellow bracket) of the coronal image shows that the deep learning system image appears smoother and less noisy than the base real image (pane g-i). Comparing the TCI image with the underlying real and deep learning system images shows that there is significantly more blur in the TCI image. The borders of inflammation in the TCI image (green arrows) do not have the same contour as the basal real, while the central cartilage has a stepped jagged appearance (dashed arrows) rather than smooth edges.

The example image of fig. 10 shows an example of identifying subtle cases of grade 2A chondromalacia of the lateral patellar cartilage through a low intensity DESS signal. Cartilage (indicated by arrows) shows similarity in all three sagittal plane images. However, the axial reconstruction exhibits slightly higher fidelity of the deep learning system image compared to the TCI image (panes a-c). Joint effusion (dashed arrow) on axial images around the lateral femoral condyle exhibits a significantly jagged appearance on TCI images compared to the underlying real and deep learning system (pane d-f). The scaled sections of the axial reconstruction show that the contour and inhomogeneous signal of the articular cartilage is better maintained in the deep learning system image than in the TCI image, but are not as good as in the underlying real image (pane g-i) at all.

The example of meniscal tears of the flanks of all three sets of images demonstrates that the sagittal plane TCI image appears more blurred compared to the deep learning system image, however, the coronal plane TCI reconstruction is significantly more blurred than the deep learning system of the present disclosure. The contours of fine structures such as cartilage, meniscus and joint inflammation in the deep learning system image appear more similar to the underlying real image. In contrast to significant meniscal tears, it was very subtle to find that grade 2A lateral patellochondromalacia (according to the modified Noyes scale) behaves similarly on all three sagittal plane image sets, but axial deep learning system reconstruction has better image quality than TCI. The fine contour of cartilage and cartilage signal inhomogeneity are well delineated on the underlying real image but appear more blurred on the deep learning system and TCI images. However, the deep learning system of the present disclosure does maintain a higher image quality compared to TCI. As illustrated in fig. 8-10, the deep learning system of the present disclosure has the potential for diagnostic use. For example, healthy MCLs typically have sub-millimeter thicknesses in the coronal plane and the deep learning system is able to maintain fine MCL details in an acquisition that simulates a slice thickness of 2.1 mm. In another example, meniscal tears and delicate cartilage are visible in images produced by the deep learning system of the present disclosure And (6) pathological changes. In some applications, a deep learning system may be used to acquire slices of 2-3mm slice thickness and then transform them into thinner slices (e.g., higher resolution) for clinical use in multi-planar reconstruction. Such methods may also be particularly applicable to newer embodiments of DESS, which achieve synchronization T₂Relaxation time, morphometry and semi-quantitative radiology evaluation. For example, in some applications, thicker slices can be used to generate data for quantifying T₂High SNR of the measurement, while thin slices can be used for accurate morphometry and semi-quantitative whole joint assessment. Bilateral knee imaging methods that acquire hundreds of slices may also benefit from the deep learning system of the present disclosure.

Fig. 11 is an illustration of a system 1100 arranged in accordance with examples described herein. The system 1100 may include an image acquisition unit 1102, a computing system 1108, and a workstation 1118. In some embodiments, the image acquisition unit 1102, the computing system 1108, and the workstation 1118 may be integrated into a single unit.

In some examples, the image acquisition unit 1102 may be a Magnetic Resonance Imaging (MRI) imaging system. In some examples, the image acquisition unit 1102 may be a Computed Tomography (CT) imaging system. In some examples, the image acquisition unit 1102 may be an Ultrasound (US) imaging system. In some examples, the image acquisition unit 1102 may be another imaging modality capable of acquiring image slices of limited thickness. The image acquisition unit 1102 may acquire one or more image slices 1106 from the individual 1104. The subject 1104 may be a human or animal subject. In some examples, the image slice 1106 may include one or more features of interest. In some examples, the feature of interest may be an anatomical feature. In the example shown in fig. 11, the image slice 1106 includes a knee joint. In some examples, the image slice 1106 may be provided to the computing system 1108.

In some examples, computing system 1108 may be implemented by computing system 106 shown in fig. 1. Computing system 1108 may include one or more memories (not shown) capable of storing executable instructions and one or more processors (not shown) capable of executing the executable instructions. The one or more memories may be configured to store one or more image slices 1106 provided by the image acquisition unit 1102. Computing system 1108 may be configured to implement deep learning system 1110. The deep learning system 1110 can be configured to generate a high resolution (e.g., thin slice) image 1114 from the image slice 1106. In some examples, the deep learning system 1110 can be implemented by the deep learning system 200 shown in fig. 2-4. The high resolution image 1114 may be provided to the workstation 1118. In some examples, computing system 1108 may be included in workstation 1118. In some examples, the computing system 1108 may be remote from the workstation 1118 and may communicate with the workstation 1118 via a network interface (e.g., the network interface 116 shown in fig. 1).

Optionally, in some examples, computing system 1108 may be configured to implement second deep learning system 1112. The second deep learning system 1112 may be configured to receive the high resolution image 1114 and output diagnostic information 1116. The trainable deep learning system 1112 identifies anatomical features associated with one or more pathologies (e.g., torn meniscus) in the high resolution image 1114 and outputs diagnostic information 1116 associated with the identified pathologies. However, in other examples, other deep learning systems may be used to implement deep learning system 1112. Diagnostic information 1116 may be provided to the workstation 1118.

The workstation 1118 may include a display 1120 and a user interface 1122. An example of data that may be provided on display 1120 is shown in block 1124. In some examples, display 1120 may provide individual information 1126. The individual information may include a name, date of birth, identification number, and/or other information. In some examples, the display 1120 may provide the image slice 1106 and/or the high resolution image 1114. In some examples, the display 1120 may provide multiple image slices 1106 and/or high resolution images 1114 simultaneously. In some examples, the display 1120 may provide a variety of other information 1130, such as a menu of options (e.g., edit images, save images, acquire measurements from images), information about the displayed images (e.g., image acquisition unit 1102 settings, anatomical views of the slices 1106, 1114) and/or other information (e.g., diagnostic information 1116, physician notes).

The workstation 1118 may include a user interface 1122. The user interface 1122 may include any type of hard user control and/or soft user control. Examples of hard user controls include, but are not limited to, a keyboard, a trackball, a mouse, and switches. Examples of soft user controls include, but are not limited to, a touch screen with software implemented graphical buttons, sliders, and/or toggles. User interface 1122 may be configured to receive user input from a user 1132. For example, the user input may determine when and/or under what conditions (e.g., thickness) to acquire the image slice 1106. In another example, the user input may determine what image is displayed on display 1120.

In some applications, a deep learning system described herein (e.g., deep learning system 1110) may permit an image acquisition unit (e.g., image acquisition unit 1102) to acquire fewer slices of a volume of subject 1104 (e.g., slices 1106 while still providing images of diagnostic quality, such as high resolution images 1114.

A user of a deep learning system described herein (e.g., user 1132) may make a diagnosis of a subject (e.g., individual 1104) using the output of the deep learning system (e.g., high resolution images 1114). For example, a physician may be able to diagnose chondromalacia of the lateral patellar cartilage and/or stage osteoarthritis based on the features of interest (e.g., the knee) visible in the high resolution image 1114. The user 1132 may further make treatment determinations based on the diagnosis made from the high-resolution images 1114. Examples of treatment options include, but are not limited to, surgery, physiotherapy, and supportive stents. In some instances, it may not be possible to make a diagnostic and/or therapeutic determination based on the initially acquired image slices 1106.

Collectively, a deep learning system as described herein may be able to resolve high resolution thin slice features from an initially significantly thicker slice. In other words, the deep learning system generates an image in which each voxel represents less volume than the voxels in the originally acquired image. It should be understood that while the example embodiment uses musculoskeletal MRI, other examples of deep learning systems and images of other modalities (e.g., brain MRI or CT) may also be used. Additionally, a deep learning system may be used to enhance resolution in all three dimensions. The deep learning system may be implemented on one or more computer systems, such as a system having one or more processing units (e.g., processors, such as Central Processing Units (CPUs) and/or Graphics Processing Units (GPUs)) and a computer-readable medium (e.g., memory, storage) encoded with executable instructions for performing the deep learning techniques and/or training described herein.

From the foregoing, it will be appreciated that, although specific embodiments have been described herein for purposes of illustration, various modifications may be made while remaining within the scope of the claimed technology.