Disclosure of Invention
The invention aims to provide a liver image reconstruction method and a liver image reconstruction system, which can automatically and accurately identify the relative positions of liver blood vessels and liver tumors, and facilitate doctors to formulate an operation scheme according to the relative positions of the liver blood vessels and the liver tumors of each patient.
In a first aspect, the present invention provides a liver image reconstruction method, including:
acquiring a CT image of a patient under a multi-scale receptive field, wherein the CT image is a liver tumor image or a hepatic vascular image of the patient;
extracting global semantic feature images and local visual feature images based on the CT images respectively;
performing feature fusion on the global semantic feature map and the local visual map by using a convolutional neural network to obtain a liver feature map;
and carrying out three-dimensional image reconstruction based on the liver feature map, and outputting a reconstruction result, wherein the reconstruction result is a three-dimensional liver tumor image or a three-dimensional liver blood vessel image.
Optionally, extracting the global semantic feature map and the local visual feature map based on the CT image respectively includes:
performing feature extraction and dimension reduction on the CT image to obtain a global semantic feature map;
cutting the CT image according to a preset standard to obtain a local feature map;
and extracting the visual features in the local feature map as a local visual feature map.
Optionally, the feature fusion of the global semantic feature map and the local visual map by using a convolutional neural network to obtain a liver feature map includes:
fusing and recovering the global semantic feature map and the local visual feature map in low dimension by using a residual network in combination with sampling;
and fusing the global semantic feature map and the local visual feature map with the same scale as channel dimensions to obtain a liver feature map.
Optionally, the reconstructing the three-dimensional image based on the liver feature map, outputting a reconstruction result, includes:
gray processing is carried out on the liver feature map, and a first liver feature map is determined;
constructing a three-dimensional data field based on the first liver feature map, and selecting data with highest gray values as an initial liver image;
processing the initial liver image using a morphological operation, the morphological operation including at least one of dilation, erosion, open operation, and closed operation, to determine a target liver image;
and reconstructing a three-dimensional image based on the target liver image, and outputting a reconstruction result.
Optionally, the method further comprises:
and storing the reconstruction result to a user database, wherein the file format of the reconstruction result in the user database is an nrrd format.
In a second aspect, the present invention provides a liver image reconstruction system comprising:
the image acquisition module is used for acquiring CT images of the patient under the multi-scale receptive field, wherein the CT images are liver tumor images or liver blood vessel images of the patient;
the feature map determining module is used for respectively extracting a global semantic feature map and a local visual feature map based on the CT image;
the feature fusion module is used for carrying out feature fusion on the global semantic feature map and the local visual map by utilizing a convolutional neural network so as to obtain a liver feature map;
the three-dimensional reconstruction module is used for carrying out three-dimensional image reconstruction based on the liver feature map and outputting a reconstruction result, wherein the reconstruction result is a three-dimensional liver tumor image or a three-dimensional liver blood vessel image.
Optionally, the feature map determining module includes:
the global feature acquisition unit is used for carrying out feature extraction and dimension reduction on the CT image to obtain a global semantic feature map;
the local feature acquisition unit is used for cutting the CT image according to a preset standard to obtain a local feature map; and extracting the visual features in the local feature map as a local visual feature map.
Optionally, the feature fusion module includes:
the recovery unit is used for fusing and recovering the global semantic feature map and the local visual feature map in low dimension by using a residual network in combination with sampling;
and the fusion unit is used for fusing the global semantic feature map and the local visual feature map with the same scale as channel dimensions to obtain a liver feature map.
Optionally, the three-dimensional reconstruction module includes:
the gray processing unit is used for carrying out gray processing on the liver feature map and determining a first liver feature map;
the image screening unit is used for constructing a three-dimensional data field based on the first liver feature map and initializing a liver image;
an arithmetic unit for processing the initial liver image using a morphological operation including at least one of an expansion, a corrosion, an open operation, and a closed operation, to determine a target liver image;
and the three-dimensional image reconstruction unit is used for carrying out three-dimensional image reconstruction based on the target liver image and outputting a reconstruction result.
Optionally, the method further comprises:
and the user database is used for storing the reconstruction result, and the file format of the reconstruction result in the user database is an nrrd format.
The embodiment of the invention provides a liver image reconstruction method and a reconstruction system, which can acquire CT images of a patient under a multi-scale receptive field, wherein the CT images are liver tumor images or liver blood vessel images of the patient; extracting global semantic feature images and local visual feature images based on the CT images respectively; performing feature fusion on the global semantic feature map and the local visual map by using a convolutional neural network to obtain a liver feature map; and carrying out three-dimensional image reconstruction based on the liver feature map, and outputting a reconstruction result, wherein the reconstruction result is a three-dimensional liver tumor image or a three-dimensional liver blood vessel image. The invention uses the deep learning technology (convolutional neural network) to finish the fusion of the global and local features of the CT image of the patient, and the fused liver feature map contains the liver tumor or liver blood vessel information of the patient, so that the three-dimensional reconstruction is carried out based on the liver feature map, the relative positions of the liver blood vessels and the liver tumor can be automatically and accurately identified, and the operation scheme is conveniently formulated according to the relative positions of the liver blood vessels and the liver tumor of each patient.
In addition, due to the adoption of the convolutional neural network in the scheme, after the convolutional neural network is trained, the information such as the hepatic blood vessels and the hepatic tumors can be automatically identified, the two hands are liberated, the situation that a doctor manually adds the hepatic blood vessels and the hepatic tumors is avoided, the whole process is rapidly and automatically completed, no professional is needed to intervene, meanwhile, the portability of the system is high, special installation is not needed, and the system is not limited by an operating system, computer configuration and regions.
Detailed Description
Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to like or similar elements or elements having like or similar functions throughout. The embodiments described below by referring to the drawings are illustrative and intended to explain the present invention and should not be construed as limiting the invention.
Referring to fig. 1, the present invention provides a liver image reconstruction method, which specifically includes:
step S11, acquiring a CT image of a patient under a multi-scale receptive field, wherein the CT image is a liver tumor image or a liver blood vessel image of the patient;
the receptive field is one of important concepts of the convolutional neural network, and the multiscale receptive field refers to that the receptive field ranges are different in size. If the receptive field is too small, the convolutional neural network only can observe local features of the image; if the receptive field is too large, the global information is more understood, but often contains much invalid information. Therefore, in the embodiment of the invention, CT images under the multiscale receptive field are acquired to provide sufficient image basis for the subsequent global semantic feature map and the local visual feature map.
Optionally, the CT image is a CT DICOM format picture.
And step S12, respectively extracting a global semantic feature map and a local visual feature map based on the CT images.
And step S13, performing feature fusion on the global semantic feature map and the local visual map by using a convolutional neural network so as to obtain a liver feature map.
And step S14, carrying out three-dimensional image reconstruction based on the liver feature map, and outputting a reconstruction result, wherein the reconstruction result is a three-dimensional liver tumor image or a three-dimensional liver blood vessel image.
The invention relates to a liver image reconstruction method and a reconstruction system, which are used for acquiring CT images of a patient under a multi-scale receptive field, wherein the CT images are liver tumor images or liver blood vessel images of the patient; extracting global semantic feature images and local visual feature images based on the CT images respectively; performing feature fusion on the global semantic feature map and the local visual map by using a convolutional neural network to obtain a liver feature map; and carrying out three-dimensional image reconstruction based on the liver feature map, and outputting a reconstruction result, wherein the reconstruction result is a three-dimensional liver tumor image or a three-dimensional liver blood vessel image. The invention uses the deep learning technology (convolutional neural network) to finish the fusion of the global and local features of the CT image of the patient, and the fused liver feature map contains the liver tumor or liver blood vessel information of the patient, so that the relative positions of the liver blood vessels and the liver tumor can be accurately identified by three-dimensional reconstruction based on the liver feature map, and the operation scheme can be conveniently formulated according to the relative positions of the liver blood vessels and the liver tumor of each patient.
In addition, due to the adoption of the convolutional neural network in the scheme, after the convolutional neural network is trained, the information such as the hepatic blood vessels and the hepatic tumors can be automatically identified, the two hands are liberated, the situation that a doctor manually adds the hepatic blood vessels and the hepatic tumors is avoided, the whole process is rapidly and automatically completed, no professional is needed to intervene, meanwhile, the portability of the system is high, special installation is not needed, and the system is not limited by an operating system, computer configuration and regions.
Specific implementations of embodiments of the present invention are described below.
In one embodiment, the extracting the global semantic feature map and the local visual feature map based on the CT image respectively may specifically include:
and carrying out feature extraction and dimension reduction on the CT image to obtain a global semantic feature map.
Optionally, a convolutional neural network multi-layer multi-head self-attention mechanism can be used for learning the CT image, and then the convolutional neural network is used for carrying out feature extraction and dimension reduction on the semantic related feature image, so that the dimensions are respectively obtained as follows: 128 x 128, 64 x 64, 32 x 32, 16 x 16, etc.
Further, the acquiring the local visual feature map based on the CT image may be:
cutting the CT image according to a preset standard to obtain a local feature map;
and extracting the visual features in the local feature map as a local visual feature map.
In this embodiment, based on a CT DICOM-format picture of the same patient, it is cut according to a preset standard, and visual features in the local feature map are extracted as the local visual feature map. Alternatively, the local visual feature map may be a feature map under a plurality of different scale receptive fields of 128×128, 64×64, 32×32, 16×16.
In a further embodiment of the invention, a liver feature map may also be determined based on the global semantic feature map and the local visual feature map.
The specific steps for determining the liver feature map include:
fusing and recovering the global semantic feature map and the local visual feature map in low dimension by using a residual network in combination with sampling; and fusing the global semantic feature map and the local visual feature map with the same scale as channel dimensions to obtain a liver feature map.
The residual network is one of convolutional neural networks, and specifically comprises the following components: and some layers of the neural network skip the connection of the next layer of neurons, and the interlayer is connected, so that the strong connection between each layer is weakened.
Optionally, in the embodiment of the present invention, a residual network is used to combine sampling to recover the low-dimensional global semantic feature map and the local visual feature map; and fusing the global semantic feature map and the local visual feature map with the same scale as channel dimensions to ensure that the module effectively learns the features with different scales and obtain a liver feature map.
Further, after obtaining the liver feature map, referring to fig. 2, the three-dimensional image reconstruction process may be:
step S21, gray processing is carried out on the liver feature map, and a first liver feature map is determined;
in addition, after the liver feature map is acquired, the Gaussian filter is used for setting a threshold value to filter irrelevant noise in the liver feature map before gray level processing.
S22, constructing a three-dimensional data field based on the first liver feature map, and determining an initial liver image;
optionally, after the three-dimensional data field, selecting the data with the highest gray value as the initial liver image.
Step S23, processing the initial liver image by using morphological operation to determine a target liver image.
Basic operations in morphological operations include expansion, erosion, open operation, closed operation, skeleton extraction, extreme erosion, hit-miss transformation, morphological gradients, top-hat transformation, particle analysis, drainage basin transformation, gray value morphological gradients, and the like.
And step S24, reconstructing a three-dimensional image based on the target liver image, and outputting a reconstruction result.
Alternatively, the reconstruction result may be obtained after reconstruction using a Visualization Tool Kit (VTK) and output in the form of a window.
In an alternative embodiment, the reconstruction result may also be saved to a user database, where the file format of the reconstruction result is nrrd format.
In a further embodiment of the present invention, a liver image reconstruction system is also provided. Referring to fig. 3, the liver image reconstruction system specifically includes:
the image acquisition module 300 is used for acquiring a CT image of a patient under a multi-scale receptive field, wherein the CT image is a liver tumor image or a hepatic vascular image of the patient;
a feature map determining module 400, configured to extract a global semantic feature map and a local visual feature map based on the CT images, respectively;
the feature fusion module 500 is configured to perform feature fusion on the global semantic feature map and the local visual map by using a convolutional neural network, so as to obtain a liver feature map;
the three-dimensional reconstruction module 600 is configured to perform three-dimensional image reconstruction based on the liver feature map, and output a reconstruction result, where the reconstruction result is a three-dimensional liver tumor image or a three-dimensional liver blood vessel image.
Optionally, the feature map determining module 400 includes:
a global feature obtaining unit 410, configured to perform feature extraction and dimension reduction on the CT image, so as to obtain a global semantic feature map;
and a local feature obtaining unit 420, configured to crop the CT image according to a preset standard, to obtain a local feature map; and extracting the visual features in the local feature map as a local visual feature map.
Optionally, the feature fusion module 500 includes:
a restoring unit 510, configured to use a residual network to perform sampling in conjunction with the global semantic feature map and the local visual feature map in a low dimension to perform fusion restoration;
and a fusion unit 520, configured to fuse the global semantic feature map and the local visual feature map with the same scale to obtain a liver feature map.
Optionally, the three-dimensional reconstruction module 600 includes:
a gray level processing unit 610, configured to perform gray level processing on the liver feature map, and determine a first liver feature map;
an image screening unit 620, configured to construct a three-dimensional data field based on the first liver feature map, and select data with the highest gray level value as an initial liver image;
an operation unit 630 for processing the initial liver image using a morphological operation including at least one of an expansion, a corrosion, an open operation, and a closed operation, to determine a target liver image;
and a three-dimensional image reconstruction unit 640, configured to perform three-dimensional image reconstruction based on the target liver image, and output a reconstruction result.
Optionally, the liver image reconstruction system further includes: and the user database 700 is used for storing the reconstruction result, and the file format of the reconstruction result in the user database is nrrd format.
The description of this embodiment may refer to the description of the foregoing liver image reconstruction method, and will not be repeated here.
The foregoing describes several embodiments of the present invention, and the various alternatives presented by the various embodiments may be combined, cross-referenced, with each other without conflict, extending beyond what is possible embodiments, all of which are considered to be embodiments of the present invention disclosed and disclosed.
Although the embodiments of the present invention are disclosed above, the present invention is not limited thereto. Various changes and modifications may be made by one skilled in the art without departing from the spirit and scope of the invention, and the scope of the invention should be assessed accordingly to that of the appended claims.