CN111192320B

CN111192320B - Position information determining method, device, equipment and storage medium

Info

Publication number: CN111192320B
Application number: CN201911390823.9A
Authority: CN
Inventors: 袁绍锋; 邹伟建; 毛玉妃
Original assignee: Shanghai United Imaging Healthcare Co Ltd
Current assignee: Shanghai United Imaging Healthcare Co Ltd
Priority date: 2019-12-30
Filing date: 2019-12-30
Publication date: 2023-07-25
Anticipated expiration: 2039-12-30
Also published as: CN111192320A

Abstract

The embodiment of the invention discloses a method, a device, equipment and a storage medium for determining position information, wherein the method for determining the position information comprises the following steps: inputting the first three-dimensional image into a candidate position determining neural network to obtain initial candidate position information corresponding to a first part in the first three-dimensional image; extracting a second three-dimensional image from the first three-dimensional image based on the initial candidate position information; inputting the second three-dimensional image into the accurate position determination neural network to obtain accurate position information of the second part, wherein the first part comprises the second part. The technical scheme of the embodiment of the invention can quickly and accurately realize the position information determination of the target part.

Description

Position information determining method, device, equipment and storage medium

Technical Field

The embodiment of the invention relates to the technical field of image processing, in particular to a method, a device, equipment and a storage medium for determining position information.

Background

In existing image processing, especially medical image processing, accurate location determination of key points or key areas is critical for subsequent operations.

Taking the determination of the location of the aortic valve anatomical keypoints in the medical image as an example, transcatheter aortic valve implantation is a minimally invasive valve replacement procedure, wherein the distance of the coronary opening to the aortic annulus, the diameter of the annulus, are important clinical measurement parameters for the selection of appropriate implantation equipment and valve size. In order to calculate the above important clinical measurement parameters, it is necessary to accurately locate important aortic valve anatomical key points, a right coronary artery opening point, a left coronary artery opening point, three sinus tube connections (junctions), and three aortic sinus bottom nadir.

The existing method for positioning the aortic valve anatomical key points mainly comprises a projection space learning algorithm, a partial strategy reinforcement learning algorithm, a detection method based on a multi-task full convolution network, a classic algorithm for Normalized Cut image analysis and the like, wherein the accuracy of positioning results of the projection space learning algorithm and the classic algorithm for Normalized Cut image analysis is poor, the partial strategy reinforcement learning algorithm and the detection method based on the multi-task full convolution network are used for independently detecting the positions of all the anatomical key points of the aortic valve, the method for independently detecting the positions of all the anatomical key points of the aortic valve ignores the relative positions among all the anatomical key points, and geometric drift or inaccurate detection may be generated in detection results.

Disclosure of Invention

The invention provides a method, a device, equipment and a storage medium for determining position information, which can quickly and accurately determine the position information of a target part.

In a first aspect, an embodiment of the present invention provides a location information determining method, where the location information determining method includes:

inputting a first three-dimensional image into a candidate position determining neural network to obtain initial candidate position information corresponding to a first part in the first three-dimensional image;

extracting a second three-dimensional image from the first three-dimensional image based on the initial candidate position information;

inputting the second three-dimensional image into a precise position determining neural network to obtain precise position information of a second part, wherein the first part comprises the second part.

In a second aspect, an embodiment of the present invention further provides a location information determining apparatus, where the location information determining apparatus includes:

the initial candidate position information determining module is used for inputting a first three-dimensional image into the candidate position determining neural network to obtain initial candidate position information corresponding to a first part in the first three-dimensional image;

a second three-dimensional image extraction module for extracting a second three-dimensional image from the first three-dimensional image based on the initial candidate position information;

The accurate position information determining module is used for inputting the second three-dimensional image into an accurate position determining neural network to obtain accurate position information of a second part, and the first part comprises the second part.

In a third aspect, an embodiment of the present invention further provides a computer apparatus, the apparatus including:

one or more processors;

storage means for storing one or more programs,

the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the location information determining method according to any of the embodiments of the present invention.

In a fourth aspect, embodiments of the present invention further provide a computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements a location information determining method according to any of the embodiments of the present invention.

According to the embodiment of the invention, the first three-dimensional image is input into a candidate position determining neural network to obtain initial candidate position information corresponding to a first part in the first three-dimensional image; extracting a second three-dimensional image from the first three-dimensional image based on the initial candidate position information; the second three-dimensional image is input into the accurate position determination neural network to obtain the accurate position information of the second part, and the first part comprises the second part, so that the position information of the target part can be rapidly and accurately determined.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, a brief description will be given below of the drawings required for the embodiments or the prior art descriptions, and it is obvious that the drawings in the following description are some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1a is a flow chart of a method for determining location information in accordance with a first embodiment of the present invention;

FIG. 1b is a schematic diagram of a network structure of a candidate position-determining neural network according to a first embodiment of the present invention;

FIG. 1c is a schematic diagram of a network architecture of a precise location determination neural network in accordance with a first embodiment of the present invention;

FIG. 1d is a schematic diagram of a U-Net network architecture of an accurate position determination neural network in accordance with a first embodiment of the present invention;

FIG. 1e is a schematic diagram of a network structure of a first level candidate position determination neural network according to a first embodiment of the invention;

FIG. 1f is a schematic diagram of a network structure of a second level candidate position determining neural network according to a first embodiment of the invention;

FIG. 1g is a schematic illustration of an aortic valve anatomical keypoint location determined based on a location information determination method according to a first embodiment of the invention;

FIG. 1h is a schematic illustration of an aortic valve anatomical keypoint location determined based on another method of determining location information in accordance with the first embodiment of the invention;

FIG. 1i is a schematic illustration of an aortic valve anatomical keypoint location determined based on another method of determining location information in accordance with the first embodiment of the invention;

fig. 1j is a schematic diagram of a method for determining a peripheral blood vessel segmentation at a hepatic vein junction based on position information according to a first embodiment of the present invention;

FIG. 1k is a schematic diagram of a method for determining a peripheral blood vessel segmentation at a hepatic vein junction based on position information according to another embodiment of the present invention;

fig. 2 is a flowchart of a position information determining method in the second embodiment of the present invention;

fig. 3 is a schematic diagram of a position information determining apparatus according to a third embodiment of the present invention;

fig. 4 is a schematic structural diagram of a computer device in a fourth embodiment of the present invention.

Detailed Description

The invention is described in further detail below with reference to the drawings and examples. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and are not limiting thereof. It should be further noted that, for convenience of description, only some, but not all of the structures related to the present invention are shown in the drawings.

Example 1

Fig. 1a is a flowchart of a method for determining location information according to an embodiment of the present invention, where the method may be performed by a location information determining device, the device may be implemented in software and/or hardware, and the device may be configured in a computer device. As shown in fig. 1a, the method of this embodiment specifically may include:

s110, inputting the first three-dimensional image into a candidate position determining neural network to obtain initial candidate position information corresponding to a first part in the first three-dimensional image.

The first three-dimensional image may be a medical three-dimensional image or a non-medical three-dimensional image (for example, any one of a medical two-dimensional image, a natural image, a remote sensing image, a multispectral image, etc.), and the medical three-dimensional image may be a reconstructed three-dimensional image. Taking the first three-dimensional image as an example of a medical three-dimensional image, it may include a CTA (CT angiography, CT full scale Computed Tomography, i.e. computed tomography) image with fewer slices, or may include a CT image with more slices, for example, a CT image including a chest and an abdomen.

Preferably, the candidate position-determining neural network may comprise at least one level of encoder-type depth neural network, wherein the encoder-type depth neural network may preferably be a neural network capable of encoding high-dimensional image data into low-dimensional feature vectors. For example, the candidate position-determining neural network may include a one-level depth (full) convolutional neural network, or may include a two-level and higher depth (full) convolutional neural network.

In this embodiment, the initial candidate position information may roughly reflect the approximate position of the first portion in the first three-dimensional image. Preferably, the initial candidate position information may include: the initial candidate frame positional shift information (the initial candidate frame positional shift information may be one or a plurality of), the classification information corresponding to the first portion, and the positional shift information of the target portion in the first portion. Preferably, the position information of the initial candidate frame in the first three-dimensional image may be determined based on the initial candidate frame position offset information; whether the corresponding part is the first part or not can be determined according to the classification information corresponding to the first part; the positional offset information of the target site may be utilized to provide a coarse positional reference for subsequent accurate positional information determination of the second site.

The initial candidate position information may further include initial candidate coordinate position information corresponding to the first portion (the coordinate position information may be coordinate position information of a key boundary point of the first portion or coordinate position information of a feature point of the first portion), and the position information of the first portion in the first three-dimensional image may be determined based on the initial candidate coordinate position information. It can be understood that the initial candidate coordinate position information output after the candidate position determining neural network processing may not be matched with the corresponding coordinates in the first three-dimensional image, and at this time, the conversion ratio of the image size corresponding to the candidate position determining neural network and the size of the first three-dimensional image may be determined based on the candidate position determining neural network, the conversion relationship between the initial candidate coordinate position information and the corresponding coordinates in the first three-dimensional image may be determined, and then the position information of the first part in the first three-dimensional image may be determined.

Before the first three-dimensional image is input into the candidate position determination neural network to obtain initial candidate position information corresponding to the first part in the first three-dimensional image, the candidate position determination neural network may preferably be trained in advance by using training set data corresponding to the first part. The training process and the established candidate position determination neural network structure to be trained can be designed according to actual tasks. Taking medical image anatomical key point detection as an example, the data volume of the detection data set is not large, so that a multi-task learning method can be adopted, and the detection accuracy of the position where the anatomical key point is located is improved. Preferably, the parameter optimization of the candidate position determining neural network may employ a variant algorithm Adam of the random gradient descent method, whose coefficients for calculating the gradient moving average and the square value of the average may be set to 0.9 and 0.999, respectively. In the multi-task learning, the cross entropy loss function may be used for the two classification tasks corresponding to the first portion, and the euclidean distance loss function may be used for both the initial candidate frame position offset information prediction task corresponding to the first portion and the prediction task of the position offset information of the target portion.

The candidate location determination neural network may preferably also be tested using the test set data after the candidate location determination neural network training is completed. In the test phase, the first-stage deep convolutional network outputs only the classification information corresponding to the first part (which is represented in the form of probability, for example), the initial candidate frame position offset information corresponding to the first part, and the position offset information of the target part in the first part. If the multi-level candidate position determination neural network is trained, then in a test phase, the next multi-level candidate position determination neural network may receive a number of second three-dimensional images including the first location and a number of second three-dimensional images not including the first location, further exclude the second three-dimensional images not including the first location, and may optimize the boundary position corresponding to the second three-dimensional images until only the second three-dimensional image of one first location remains.

And S120, extracting a second three-dimensional image from the first three-dimensional image based on the initial candidate position information.

Based on the initial candidate location information, each possible location of the first portion in the first three-dimensional image may be determined, and preferably, each screened possible location may be extracted from the first three-dimensional image so that each possible location may be purposefully verified later. The second three-dimensional image may be one or more.

S130, inputting the second three-dimensional image into the accurate position determination neural network to obtain accurate position information of the second part, wherein the first part comprises the second part.

Preferably, the accurate position determination neural network includes an encoder-decoder type symmetric deep neural network and an encoder-decoder type asymmetric deep neural network; wherein the encoder-decoder type symmetric deep neural network comprises any one of a U-Net deep convolutional neural network, a V-Net deep convolutional neural network and a HourgassNet deep convolutional neural network. Among them, the encoder-decoder type network may be a network that encodes high-dimensional image data into low-dimensional feature vectors using an encoder and then decodes the low-dimensional feature vectors into high-dimensional output images of the same size as the original input images using a decoder.

Before the second three-dimensional image is input into the accurate position determination neural network to obtain the accurate position information of the second part, preferably, the accurate position determination neural network may be trained in advance by using training set data corresponding to the second part. The training process and the established accurate position determination neural network structure to be trained can be designed according to actual tasks. Taking medical image anatomical key point detection as an example, a variant algorithm Adam of a random gradient descent method can be adopted for optimizing parameters, and coefficients of a gradient moving average value and a square value of the gradient moving average value can be set to be 0.9 and 0.999 respectively. The accurate location information prediction task for the second location may preferably use a euclidean distance loss function.

In the test stage, the accurate position determining neural network receives the output result of the previous candidate position determining neural network, outputs a multi-channel output image with the same size as the input second three-dimensional image, and each channel outputs the accurate position information of the second part. If the candidate position determining neural network comprises multiple stages, the accurate position determining neural network only receives a unique output result of the previous candidate position determining neural network, outputs a multi-channel output image with the same size as the input second three-dimensional image, and each channel outputs accurate position information of the second part.

Preferably, the accurate position information of the second portion may be accurate coordinate position information, or may be an image capable of determining the position information, for example, may be a gaussian thermonuclear image.

Preferably, the present embodiment can be applied to any scene of aortic valve anatomy key point detection, aortic root positioning and segmentation, aortic valve leaflet calcification point detection, hepatic vein afflux position peripheral blood vessel positioning and segmentation, hepatic portal vein left and right branch bifurcation position peripheral blood vessel positioning and segmentation, hepatic right inferior vein and inferior vena cava intersection position and segmentation, splenic portal peripheral blood vessel positioning and segmentation, renal portal peripheral blood vessel positioning and segmentation, and common iliac artery left and right branch bifurcation position peripheral blood vessel positioning and segmentation. For example, the application scenario is aortic valve anatomical key point detection, the first part is an aortic valve root, the second part is 8 aortic anatomical key points, and the first part is a right coronary artery opening point, a left coronary artery opening point, three sinus tube connecting parts (boundary points) and three aortic sinus bottom lowest points.

The above procedure is specifically described by taking the application scenario as an aortic valve anatomical key point detection as an example:

when the candidate position-determining neural network includes only one stage, the network structure of the network may be configured to include 8 convolutional layers, 5 batch normalization layers (Batch Normalization, BN), 5 parameterized ReLU function (Parametric Rectified Linear Units, prime) activation layers, and 2 Max Pooling (MP) layers. The candidate position determining neural network is a full convolution network and is characterized in that a segmented image is used as a training sample during training, and the whole three-dimensional image is used as test data during testing, so that the target detection speed and efficiency are improved. The candidate position-determining neural network has an input of 24 x 1, where 24 x 24 represents a three-dimensional medical data image patch, 1 indicates that the number of color channels of the input image is 1, i.e., the image is segmented into grayscale images.

Fig. 1b is a schematic diagram of a network structure of a candidate location determining neural network according to a first embodiment of the present invention, where, as shown in fig. 1b, the network structure of the candidate location determining neural network sequentially includes, from left to right:

the first layer 1101, the convolutional layer Conv, whose convolutional kernel size k is 3 x 3, the input channel size 1, the output channel size f is 16, the moving step s is 1, and the batch normalization layer and the PReLU function activation layer are connected.

A second layer 1102, a convolutional layer Conv, having a convolutional kernel size k of 3 x 3, an input channel size of 16, the output channel size f is 16, the moving step s is 1, and the batch normalization layer and the PReLU function activation layer are connected.

The third layer 1103, the maximum pooling layer MP, the pooling interval k is 3 multiplied by 3, the movement step s is 2.

A fourth layer 1104, a convolutional layer Conv, having a convolutional kernel size k of 3 x 3, an input channel size of 16, the output channel size f is 32, the moving step s is 1, and the batch normalization layer and the PReLU function activation layer are connected.

A fifth layer 1105, a convolutional layer Conv, having a convolutional kernel size k of 3 x 3, an input channel size of 32, the output channel size f is 32, the moving step s is 1, and the batch normalization layer and the PReLU function activation layer are connected.

A sixth layer 1106, a max-pooling layer MP, the pooling interval k is 3 multiplied by 3, the movement step s is 2.

Seventh layer 1107, convolutional layer Conv, has a convolutional kernel size k of 3 x 3, an input channel size of 32, the output channel size f is 64, the movement step s is 1, and the batch normalization layer and the PReLU function activation layer are connected.

Eighth layer 1108,3 convolution layers Conv1-Conv3, each sub-layer being concatenated with the seventh layer, having a convolution kernel size k of 1 x 1, the input channel size is 64, the output channel size f is 1, 6 and 24 from top to bottom respectively, and the 1 st convolution layer Conv1 is connected with the Sigmoid function activation layer. The used supervision information is the classification information of the aortic root and the non-aortic root, the position offset information of the aortic root candidate frame and the position information of the aortic valve anatomical key points, wherein the aortic anatomical key points comprise a right coronary artery opening point, a left coronary artery opening point, three sinus tube connecting parts (junction points) and three aortic sinus bottom lowest points.

Inputting the first aortic valve three-dimensional images into a first-stage candidate position determination neural network, obtaining classification information of aortic root and non-aortic root, position deviation information of an aortic root candidate frame and position deviation information of aortic valve anatomical key points, determining a plurality of second aortic valve root three-dimensional images according to the information, and respectively inputting the plurality of second aortic valve root three-dimensional images into a second-stage accurate position determination neural network.

Preferably, the network structure of the precise location determining neural network may be configured to include 42 convolutional layers (residual block occupies 13×3=39 convolutional layers), 40 batch normalization (Batch Normalization, BN) layers, 42 ReLU function (Rectified Linear Units, reLU) activation layers, 4 Max Pooling (MP) layers, 4 nearest neighbor upsampling layers, and 4 skip layers with residual block transforms. The input to the accurate position determination neural network is 64 x 1, where 64 x 64 represents a three-dimensional medical data image patch, 1 indicates that the number of color channels of the input image is 1, i.e., the image is segmented into grayscale images.

Fig. 1c is a schematic diagram of a network structure of an accurate position determining neural network according to a first embodiment of the present invention, where, as shown in fig. 1c, the network structure of the accurate position determining neural network sequentially includes, from left to right:

The first layer 1201, the convolutional layer Conv, has a convolutional kernel size k of 3 x 3, an input channel size of 1, the output channel size f is 32, the moving step s is 1, the filling size p is 1, and the ReLU function activation layer is connected.

A second layer 1202, a convolutional layer Conv, having a convolutional kernel size k of 3 x 3, an input channel size of 32, the output channel size f is 64, the movement step s is 1, the filling size p is 1, and the ReLU function activation layer is connected.

A third layer 1203, a residual block ResBlock, comprising 3 convolutions layers Conv, a first layer sub-layer convolution kernel size k of 1 x 1, a second layer sub-layer convolution kernel size k of 3 x 3, the convolution kernel size k of the third sub-layer is 1 multiplied by 1, and all sub-layers are preceded by a batch normalization layer and a ReLU function activation layer. In addition, the first layer sub-layer input channel size is 64, the output channel size is 32, the second layer sub-layer input channel size is 32, the output channel size is 32, the third layer sub-layer input channel size is 32, and the output channel size is 64. The residual block ResBlock described later is identical to the above description.

A fourth layer 1204, a sixth layer 1206, an eighth layer 1208 and a tenth layer 1210, the maximum value is pooled into a layer MP, the pooling interval k is 2 multiplied by 2, the movement step s is 2.

Fifth layer 1205, seventh layer 1207, ninth layer 1209, eleventh layer 1211, thirteenth layer 1213, fifteenth layer 1215, seventeenth layer 1217, and nineteenth layer 1219, and a residual block ResBlock.

The twelfth, fourteenth, sixteenth, and eighteenth layers 1212, 1214, 1216, 1218, nearest neighbor upsampling layers may magnify the feature map size by a factor of 2.

Twentieth layer 1220, convolutional layer Conv, has a convolutional kernel size k of 1 x 1, an input channel size of 64, and an output channel size of the number of detected anatomical keypoints, 8 of which need to be located. Illustratively, the supervision information used is position information of aortic valve anatomical keypoints represented in the form of gaussian nuclear heat maps.

The accurate position determination neural network in this embodiment may be a U-Net network in addition to the encoder-decoder type depth network in the above example. The U-Net network structure of the precise position determination neural network may be configured to include 15 convolutional layers, 14 batch normalization (Batch Normalization, BN) layers, 14 ReLU function (Rectified Linear Units, reLU) activation layers, 3 max pooling layers, 3 transposed convolutional layers, and 3 skip layers without any transforms. The input to the U-Net network of the accurate position determination neural network is 64 x 1, wherein, 64 x 64 represents three-dimensional medical data image segmentation, 1 indicates that the number of color channels of the input image is 1, i.e., the image is segmented into grayscale images.

Fig. 1d is a schematic diagram of a U-Net network structure of a precise location determining neural network according to a first embodiment of the present invention, where, as shown in fig. 1d, the U-Net network structure of the precise location determining neural network sequentially includes, from left to right:

a first layer 1301, a convolutional layer Conv, having a convolutional kernel size k of 3 x 3, an input channel size of 1, an output channel size f of 8, the movement step s is 1, the filling size p is 1, and the batch normalization layer and the ReLU function activation layer are connected.

The second layer 1302, the convolutional layer Conv, has a convolutional kernel size k of 3 x 3, an input channel size of 8, an output channel size f of 16, a move step s of 1, a fill size p of 1, followed by a batch normalization layer and a ReLU function activation layer.

The third layer 1303, the maximum pooling layer MP, the pooling interval k is 2 multiplied by 2, the movement step s is 2.

A fourth layer 1304, a convolutional layer Conv, having a convolutional kernel size k of 3 x 3, the input channel size is 16, the output channel size f is 16, the movement step s is 1, the filling size p is 1, and the batch normalization layer and the ReLU function activation layer are connected.

Fifth layer 1305, convolutional layer Conv, with a convolutional kernel size k of 3 x 3, an input channel size of 16, an output channel size f of 32, the movement step s is 1, the filling size p is 1, and the batch normalization layer and the ReLU function activation layer are connected.

A sixth layer 1306, a maximum pooling layer MP, the pooling interval k is 2 multiplied by 2, the movement step s is 2.

A seventh layer 1307, a convolutional layer Conv, whose convolutional kernel size k is 3 x 3, input channel size 32, output channel size f 32, the movement step s is 1, the filling size p is 1, and the batch normalization layer and the ReLU function activation layer are connected.

An eighth layer 1308, a convolutional layer Conv, having a convolutional kernel size k of 3 x 3, an input channel size of 32, an output channel size f of 64, the movement step s is 1, the filling size p is 1, and the batch normalization layer and the ReLU function activation layer are connected.

A ninth layer 1309, a maximum pooling layer MP, the pooling interval k is 2 multiplied by 2, the movement step s is 2.

A tenth layer 1310, a convolutional layer Conv, a convolutional kernel size k of 3 x 3, an input channel size of 64, an output channel size f of 64, the movement step s is 1, the filling size p is 1, and the batch normalization layer and the ReLU function activation layer are connected.

Eleventh layer 1311, convolutional layer Conv, has a convolutional kernel size k of 3 x 3, an input channel size of 64, the output channel size f is 128, the movement step s is 1, the filling size p is 1, and the batch normalization layer and the ReLU function activation layer follow.

The twelfth layer 1312 transposes the convolutional layer TConv with a convolutional kernel size k of 2 x 2 and a shift step s of 2.

Thirteenth layer 1313 and fourteenth layer 1314, convolutional layer Conv, having a convolutional kernel size k of 3 x 3, an input channel size of 64, the output channel size f is 64, the movement step s is 1, the filling size p is 1, and the batch normalization layer and the ReLU function activation layer follow.

Fifteenth layer 1315 transposes the convolutional layer TConv with a convolutional kernel size k of 2 x 2 and a shift step s of 2.

Sixteenth layer 1316 and seventeenth layer 1317, convolutional layer Conv, having a convolutional kernel size k of 3 x 3, an input channel size of 32, the output channel size f is 32, the movement step s is 1, the filling size p is 1, and the batch normalization layer and the ReLU function activation layer are connected.

Eighteenth layer 1318, transposed convolutional layer TConv, the convolution kernel size k is 2 x 2, the movement step s is 2.

Nineteenth layer 1319 and twentieth layer 1320, convolutional layer Conv, with a convolutional kernel size k of 3 x 3, an input channel size of 16, the output channel size f is 16, the movement step s is 1, the filling size p is 1, and the batch normalization layer and the ReLU function activation layer are connected.

The twenty-first layer 1321, the convolutional layer Conv, has a convolutional kernel size k of 3 x 3, an input channel size of 16, an output channel size f of the type of segmented anatomy, for example, the peripheral blood vessels at the junction of hepatic veins need to be segmented into categories 3, namely the background, hepatic veins and inferior vena cava. The supervision information used is a manually annotated vessel mask image.

If the candidate position-determining neural network comprises two levels, the network structure of the first level candidate position-determining neural network may be configured to include 6 convolutional layers, 3 batch normalization layers, 3 parameterized ReLU function (Parametric Rectified Linear Units, PReLU) activation layers, and 2 max pooling layers. The input to the first level candidate position-determining neural network is 24 x 1, where 24 x 24 represents a three-dimensional medical data image patch, 1 indicates that the number of color channels of the input image is 1, i.e., the image is segmented into grayscale images.

Fig. 1e is a schematic network structure diagram of a first-stage candidate position determining neural network according to an embodiment of the present invention, where, as shown in fig. 1e, the network structure of the first-stage candidate position determining neural network is sequentially from left to right:

a first layer 1401, a convolutional layer Conv, having a convolutional kernel size k of 3 x 3, an input channel size of 1, the output channel size f is 16, the moving step s is 1, and the batch normalization layer and the PReLU function activation layer are connected.

The second layer 1402, the max-pooling layer MP, the pooling interval k is 3 multiplied by 3, the movement step s is 2.

A third layer 1403, a convolutional layer Conv, having a convolutional kernel size k of 3 x 3, an input channel size of 16, the output channel size f is 32, the moving step s is 1, and the batch normalization layer and the PReLU function activation layer are connected.

A fourth layer 1404, a maximum pooling layer MP, the pooling interval k is 3 multiplied by 3, the movement step s is 2.

Fifth layer 1405, convolutional layer Conv, with a convolutional kernel size k of 3 x 3, an input channel size of 32, the output channel size f is 64, the movement step s is 1, and the batch normalization layer and the PReLU function activation layer are connected.

A sixth layer 1406,3 of the convolution layers Conv1-Conv3, each of which is concatenated with the fifth layer, having a convolution kernel size k of 1 x 1, the input channel size is 64, the output channel size f is 1, 6 and 24 from top to bottom respectively, the 1 st convolution layer Conv1 is connected with a Sigmoid function activation layer, and the used supervision information is respectively classified information of an aortic root and a non-aortic root, position offset information of an aortic root candidate frame and position offset information of an aortic valve anatomical key point.

The second level candidate position-determining neural network may be configured to include 4 convolutional layers, 4 batch normalization layers, 5 parameterized ReLU function (Parametric Rectified Linear Units, prilu) activation layers, and 3 max pooling layers and 4 Full Connected (FC) layers. The input to the second level candidate position-determining neural network is 48 x 1, where 48 x 48 represents a three-dimensional medical data image patch, 1 indicates that the number of color channels of the input image is 1, i.e., the image is segmented into grayscale images.

Fig. 1f is a schematic diagram of a network structure of a second-stage candidate position determining neural network according to a first embodiment of the present invention, where, as shown in fig. 1f, the network structure of the second-stage candidate position determining neural network sequentially includes, from left to right:

a first layer 1501, a convolution layer Conv, having a convolution kernel size k of 3 x 3, an input channel size of 1, the output channel size f is 32, the moving step s is 1, and the batch normalization layer and the PReLU function activation layer are connected.

The second layer 1502, max-pooling layer MP, the pooling interval k is 3 multiplied by 3, the movement step s is 2.

A third layer 1503, a convolutional layer Conv, having a convolutional kernel size k of 3 x 3, an input channel size of 32, the output channel size f is 64, the movement step s is 1, and the batch normalization layer and the PReLU function activation layer are connected.

A fourth layer 1504, a maximum pooling layer MP, the pooling interval k is 3 multiplied by 3, the movement step s is 2.

A fifth layer 1505, a convolutional layer Conv, having a convolutional kernel size k of 3 x 3, an input channel size of 64, the output channel size f is 64, the movement step s is 1, and the batch normalization layer and the PReLU function activation layer are connected.

A sixth layer 1506, a maximum pooling layer MP, the pooling interval k is 2 multiplied by 2, the movement step s is 2.

Seventh layer 1507, convolutional layer Conv, with a convolutional kernel size k of 2 x 2, an input channel size of 64, the output channel size f is 128, the moving step s is 1, and the batch normalization layer and the PReLU function activation layer are connected.

An eighth layer 1508, a fully connected layer FC, having an input neuron number of 128 x 2, the number f of output neurons is 256, followed by the PReLU function activation layer.

The ninth layer 1509,3 full-connection layers FC1-FC3 are connected with the eighth layer in series, the number of input neurons is 256, the number of output neurons is 1, 6 and 24 from top to bottom respectively, the 1 st full-connection layer FC1 is connected with a Sigmoid function activation layer, and used monitoring information is classified information of an aortic root and a non-aortic root, position offset information of an aortic root candidate frame and position offset information of an aortic valve anatomical key point respectively.

Fig. 1g, 1h and 1i are schematic diagrams of an aortic valve anatomical key location determined based on a location information determining method according to a first embodiment of the invention. As shown in fig. 1g, fig. 1g-1 is a cross-sectional view of a first three-dimensional image related to an aortic valve, fig. 1g-2 to fig. 1g-7 are position offset information of a frame selection region of an aortic root candidate frame and aortic valve anatomical keypoints in the frame selection region obtained after the first three-dimensional image is input into a candidate position determination neural network, and fig. 1g-8 are specific positions of eight anatomical keypoints of the aortic valve output from the accurate position determination neural network. As shown in fig. 1h, fig. 1h-1 is a cross-sectional view of another first three-dimensional image related to an aortic valve, fig. 1h-2 to fig. 1h-7 are position offset information of a frame selection region of an aortic root candidate frame and aortic valve anatomical keypoints in the frame selection region obtained after the first three-dimensional image is input into a candidate position determination neural network, and fig. 1h-8 are specific positions of eight anatomical keypoints of the aortic valve output from the accurate position determination neural network. As shown in fig. 1i, fig. 1i-1 to 1i-3 are cross-sectional views of another first three-dimensional image related to an aortic valve, fig. 1i-4 to 1i-9 are frame selection areas of aortic root candidate frames obtained after inputting the first three-dimensional image into a candidate position determination neural network and position offset information of aortic valve anatomical keypoints in the frame selection areas, and fig. 1i-10 are specific positions of eight anatomical keypoints of the aortic valve output by the accurate position determination neural network.

Table 1 shows the comparison results of the position information determination method in the present embodiment with respect to the accuracy of determining the position of the aortic valve anatomical key point. As shown in table 1:

TABLE 1

The method of the present embodiment can be used to perform peripheral vessel segmentation at the hepatic vein junction, in addition to determining aortic valve anatomical keypoint detection. Fig. 1j and 1k are schematic diagrams of a method for determining a peripheral blood vessel of a hepatic vein junction based on position information according to a first embodiment of the present invention. As shown in fig. 1j, fig. 1j-1 to 1j-3 are cross-sectional views of a first three-dimensional image related to a hepatic vein, fig. 1j-4 to 1j-9 are frame selection areas of a hepatic vein afflux-site peripheral blood vessel candidate frame obtained after inputting the first three-dimensional image into a candidate position determination neural network, and position shift information of a hepatic vein afflux-site peripheral blood vessel within the frame selection areas, and fig. 1j-10 are specific positions of a hepatic vein afflux-site peripheral blood vessel output by the accurate position determination neural network. As shown in fig. 1k, fig. 1k-1 to 1k-3 are cross-sectional views of a first three-dimensional image related to a hepatic vein, fig. 1k-4 to 1k-9 are frame-selected areas of a hepatic vein afflux-site peripheral blood vessel candidate frame obtained after inputting the first three-dimensional image into a candidate position determination neural network, and positional shift information of a hepatic vein afflux-site peripheral blood vessel within the frame-selected areas, respectively, and fig. 1k-10 are specific positions of a hepatic vein afflux-site peripheral blood vessel output by the accurate position determination neural network. According to the position information determining method provided by the embodiment, a first three-dimensional image is input into a candidate position determining neural network to obtain initial candidate position information corresponding to a first part in the first three-dimensional image; extracting a second three-dimensional image from the first three-dimensional image based on the initial candidate position information; the second three-dimensional image is input into the accurate position determination neural network to obtain the accurate position information of the second part, the first part comprises the second part, the positioning area of the image is reduced by determining the initial candidate position information, and the accurate position information of the target part is determined on the basis of the initial candidate position information, so that the position information of the target part can be rapidly and accurately determined.

On the basis of the above embodiments, further, before inputting the first three-dimensional image into the candidate position-determining neural network, the method further includes:

isotropically transforming the first three-dimensional image into a first three-dimensional image having a first voxel, the first voxel having a first volume;

after extracting the second three-dimensional image from the first three-dimensional image according to the candidate position information, further comprising:

isotropically transforming the second three-dimensional image into a second three-dimensional image having a second voxel, the volume of the second voxel being a second volume, wherein the first volume is larger than the second volume.

In fact, in the first three-dimensional image, each of its original voxels is not equal in length in the three directions X, Y and Z, in order to ensure a uniform distribution of the voxels in the first three-dimensional image, it is preferable that the original voxels in the first three-dimensional image are isotropic to the first three-dimensional image having the first voxels, wherein the lengths of the first voxels in the three directions X, Y and Z are equal.

Illustratively, if the lengths of the original voxels in the directions X, Y and Z are 1mm×1.5mm×2mm, respectively, the original voxels may be isotropically formed into the first voxels of 2.3mm×2.3mm, and the number of total prime numbers may be reduced while ensuring uniform distribution of the voxels in the first three-dimensional image, so as to reduce the calculation amount of the first-order network.

Accordingly, since the second three-dimensional image is an image extracted from the first three-dimensional image, the volume thereof is relatively small, and in order to be able to display more detailed information thereof, it is preferable that the first voxel in the second three-dimensional image is also isotropic to be a second voxel, wherein the volume of the second voxel is smaller than the volume of the first voxel. Illustratively, the first voxel is 2.3mm and the second voxel is 1.0mm.

On the basis of the above embodiments, further, inputting the second three-dimensional image into the accurate position determining neural network to obtain accurate position information of the second part, including:

inputting the second three-dimensional image into a precise position determining neural network to obtain a Gaussian heat nuclear map corresponding to the second part;

and determining the accurate position information of the second part according to the Gaussian thermonuclear diagram.

Wherein the size of the gaussian thermal kernel map is the same as the size of the second three-dimensional image. One accurate position corresponds to one gaussian thermal kernel map, taking 8 aortic dissection keypoints as an example, and each aortic dissection keypoint corresponds to one gaussian thermal kernel map.

Each Gaussian thermonuclear map comprises a Gaussian kernel, and each Gaussian kernel is the approximate position of the second part. Preferably, the position with the largest median value in each gaussian heat kernel graph can be used as the accurate position information of the second part.

Example two

Fig. 2 is a flowchart of a method for determining location information according to a second embodiment of the present invention. In this embodiment, on the basis of the foregoing embodiments, the selecting the initial candidate position information to include initial candidate frame position information corresponding to the first portion, and extracting, based on the initial candidate position information, a second three-dimensional image from the first three-dimensional image includes: determining a conversion ratio of the neural network to the image size by using the image size of the first three-dimensional image and the candidate position, and converting initial candidate frame position information into final candidate frame position information; and extracting a second three-dimensional image from the first three-dimensional image according to the final candidate frame position information. And extracting a second three-dimensional image from the first three-dimensional image according to the final candidate frame position information, including: expanding a frame selection range of a candidate frame in the first three-dimensional image based on the final candidate frame position information; and extracting a second three-dimensional image from the first three-dimensional image according to the expanded frame selection range. As shown in fig. 2, the method of this embodiment specifically may include:

S210, inputting the first three-dimensional image into a candidate position determining neural network to obtain initial candidate frame position information corresponding to a first part in the first three-dimensional image.

S220, determining the conversion ratio of the neural network to the image size by using the image size of the first three-dimensional image and the candidate position, and converting the initial candidate frame position information into final candidate frame position information.

Preferably, 1-3 initial candidate frame position information may be output. The initial candidate frame position information may preferably be initial candidate frame position offset information. It should be noted that, when the first three-dimensional image passes through the candidate position determining neural network, the coordinates of the output initial candidate frame cannot be matched with the corresponding coordinates in the first three-dimensional image, and at this time, the initial candidate frame position offset information needs to be processed to obtain the coordinates of the initial candidate frame in the first three-dimensional image.

Preferably, the conversion ratio of the neural network to the image size may be determined based on the initial candidate frame position offset information in combination with the image size of the first three-dimensional image and the candidate position, and the initial candidate frame position information may be converted into final candidate frame position information to determine the position information of the initial candidate frame in the first three-dimensional image.

S230, expanding a frame selection range of the candidate frame in the first three-dimensional image based on the final candidate frame position information;

s240, extracting a second three-dimensional image from the first three-dimensional image according to the expanded frame selection range.

In practical applications, the candidate frame corresponding to the final candidate frame position information may only include a part of the first portion, and in order to ensure that the first portion can maximally fall within the candidate frame, preferably, a frame selection range of the candidate frame in the first three-dimensional image may be enlarged, to obtain an enlarged frame selection range, and based on the enlarged frame selection range, the second three-dimensional image is extracted from the first three-dimensional image.

For example, on the basis of the position information of the final candidate frame, each face corresponding to each four points in the candidate frame is moved vertically by 8.0mm in a direction away from the center of the candidate frame, and each face is extended to be vertically tangent to obtain a closed cuboid or cube, and the cuboid or cube is the expanded frame selection range.

S250, inputting the second three-dimensional image into the accurate position determining neural network to obtain accurate position information of the second part, wherein the first part comprises the second part.

According to the position information determining method provided by the embodiment, a first three-dimensional image is input into a candidate position determining neural network to obtain initial candidate position information corresponding to a first part in the first three-dimensional image; extracting a second three-dimensional image from the first three-dimensional image based on the initial candidate position information; the second three-dimensional image is input into the accurate position determination neural network to obtain the accurate position information of the second part, the first part comprises the second part, the positioning area of the image is reduced by determining the initial candidate position information, and the accurate position information of the target part is determined on the basis of the initial candidate position information, so that the position information of the target part can be rapidly and accurately determined.

Example III

Fig. 3 is a schematic diagram of a position information determining apparatus according to a third embodiment of the present invention. As shown in fig. 3, the apparatus of this embodiment includes:

an initial candidate position information determining module 310, configured to input the first three-dimensional image into a candidate position determining neural network, to obtain initial candidate position information corresponding to a first location in the first three-dimensional image;

a second three-dimensional image extraction module 320, configured to extract a second three-dimensional image from the first three-dimensional image based on the initial candidate position information;

the accurate position information determining module 330 is configured to input the second three-dimensional image into the accurate position determining neural network to obtain accurate position information of the second location, where the first location includes the second location.

The position information determining device provided by the embodiment inputs the first three-dimensional image into the candidate position determining neural network by using the initial candidate position information determining module to obtain initial candidate position information corresponding to the first part in the first three-dimensional image; extracting a second three-dimensional image from the first three-dimensional image based on the initial candidate position information by using a second three-dimensional image extraction module; the accurate position information determining module is used for inputting the second three-dimensional image into the accurate position determining neural network to obtain the accurate position information of the second part, the first part comprises the second part, the positioning area of the image is reduced by determining the initial candidate position information, the accurate position information of the target part is determined on the basis of the initial candidate position information, and the position information of the target part can be rapidly and accurately determined.

Based on the above technical solutions, optionally, the initial candidate position information includes initial candidate frame position information corresponding to the first portion, and the second three-dimensional image extraction module 320 may specifically include:

a final candidate frame position information determining unit for determining a conversion ratio of the neural network to the image size by using the image size of the first three-dimensional image and the candidate position, and converting the initial candidate frame position information into final candidate frame position information;

and the second three-dimensional image extraction unit is used for extracting a second three-dimensional image from the first three-dimensional image according to the final candidate frame position information.

Based on the above technical solutions, optionally, the second three-dimensional image extraction unit may specifically include:

a frame selection range expanding subunit, configured to expand a frame selection range of the candidate frame in the first three-dimensional image based on the final candidate frame position information;

and the second three-dimensional image extraction subunit is used for extracting the second three-dimensional image from the first three-dimensional image according to the expanded frame selection range.

On the basis of the above aspects, optionally, the position information determining apparatus may further include a first isotropy module configured to isotropy the first three-dimensional image into a first three-dimensional image having a first voxel, a volume of the first voxel being the first volume, before inputting the first three-dimensional image into the candidate position determining neural network;

The position information determination apparatus may further include a second isotropy module for isotropy the second three-dimensional image into a second three-dimensional image having a second voxel having a second volume after extracting the second three-dimensional image from the first three-dimensional image based on the candidate position information, wherein the first volume is larger than the second volume.

Based on the above technical solutions, optionally, the accurate location information determining module 330 may specifically include:

the Gaussian heat nuclear map determining unit is used for inputting the second three-dimensional image into the accurate position determining neural network to obtain a Gaussian heat nuclear map corresponding to the second part;

and the accurate position information determining unit is used for determining the accurate position information of the second part according to the Gaussian heat nuclear diagram.

Based on the above technical solutions, optionally, the candidate position determining neural network includes at least one level of encoder-type deep neural network, and the accurate position determining neural network includes an encoder-decoder type symmetric deep neural network and an encoder-decoder type asymmetric deep neural network; wherein the encoder-decoder type symmetric deep neural network comprises any one of a U-Net deep convolutional neural network, a V-Net deep convolutional neural network and a HourgassNet deep convolutional neural network.

Based on the above technical solutions, the position information determining device may be specifically applied to any scene of aortic valve anatomy key point detection, aortic root positioning and segmentation, aortic valve leaflet calcification point detection, hepatic vein junction peripheral blood vessel positioning and segmentation, hepatic portal left and right branch bifurcation peripheral blood vessel positioning and segmentation, hepatic right inferior vein and inferior vena cava junction positioning and segmentation, splenic portal peripheral blood vessel positioning and segmentation, renal portal peripheral blood vessel positioning and segmentation, and common iliac artery left and right bifurcation peripheral blood vessel positioning and segmentation.

The position information determining device provided by the embodiment of the invention can execute the position information determining method provided by any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the executing method.

Example IV

Fig. 4 is a schematic structural diagram of a computer device according to a fourth embodiment of the present invention. Fig. 4 illustrates a block diagram of an exemplary computer device 412 suitable for use in implementing embodiments of the invention. The computer device 412 shown in fig. 4 is only an example and should not be construed as limiting the functionality and scope of use of embodiments of the invention.

As shown in FIG. 4, computer device 412 is in the form of a general purpose computing device. Components of computer device 412 may include, but are not limited to: one or more processors 416, a memory 428, a bus 418 that connects the various system components (including the memory 428 and the processor 416).

Bus 418 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processor, or a local bus using any of a variety of bus architectures. By way of example, and not limitation, such architectures include Industry Standard Architecture (ISA) bus, micro channel architecture (MAC) bus, enhanced ISA bus, video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus.

Computer device 412 typically includes a variety of computer system readable media. Such media can be any available media that is accessible by computer device 412 and includes both volatile and nonvolatile media, removable and non-removable media.

Memory 428 may include computer system readable media in the form of volatile memory, such as Random Access Memory (RAM) 430 and/or cache memory 432. The computer device 412 may further include other removable/non-removable, volatile/nonvolatile computer system storage media. By way of example only, storage 434 may be used to read from or write to non-removable, non-volatile magnetic media (not shown in FIG. 4, commonly referred to as a "hard disk drive"). Although not shown in fig. 4, a magnetic disk drive for reading from and writing to a removable non-volatile magnetic disk (e.g., a "floppy disk"), and an optical disk drive for reading from or writing to a removable non-volatile optical disk (e.g., a CD-ROM, DVD-ROM, or other optical media) may be provided. In such cases, each drive may be coupled to bus 418 via one or more data medium interfaces. Memory 428 may include at least one program product having a set (e.g., at least one) of program modules configured to carry out the functions of embodiments of the invention.

A program/utility 440 having a set (at least one) of program modules 442 may be stored in, for example, memory 428, such program modules 442 including, but not limited to, an operating system, one or more application programs, other program modules, and program data, each or some combination of which may include an implementation of a network environment. Program modules 442 generally perform the functions and/or methodologies in the described embodiments of the invention.

The computer device 412 may also communicate with one or more external devices 414 (e.g., keyboard, pointing device, display 424, etc., wherein the display 424 may be configured as desired), with one or more devices that enable a user to interact with the computer device 412, and/or with any device (e.g., network card, modem, etc.) that enables the computer device 412 to communicate with one or more other computing devices. Such communication may occur through an input/output (I/O) interface 422. Moreover, computer device 412 may also communicate with one or more networks such as a Local Area Network (LAN), a Wide Area Network (WAN) and/or a public network, such as the Internet, through network adapter 420. As shown, network adapter 420 communicates with other modules of computer device 412 over bus 418. It should be appreciated that although not shown in fig. 4, other hardware and/or software modules may be used in connection with computer device 412, including, but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, data backup storage, and the like.

The processor 416 performs various functional applications and data processing by running a program stored in the memory 428, for example, implementing the location information determining method provided by the embodiment of the present invention.

Example five

A fifth embodiment of the present invention provides a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements a location information determining method as provided by the embodiment of the present invention, including:

inputting the first three-dimensional image into a candidate position determining neural network to obtain initial candidate position information corresponding to a first part in the first three-dimensional image;

inputting the second three-dimensional image into the accurate position determination neural network to obtain accurate position information of the second part, wherein the first part comprises the second part.

Of course, the computer-readable storage medium provided by the embodiments of the present invention, on which the computer program is stored, is not limited to performing the method operations described above, but may also perform related operations in the computer-device-based location information determination method provided by any of the embodiments of the present invention.

The computer storage media of embodiments of the invention may take the form of any combination of one or more computer-readable media. The computer readable medium may be a computer readable signal medium or a computer readable storage medium. The computer readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or a combination of any of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium would include the following: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In this document, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.

The computer readable signal medium may include a propagated data signal with computer readable program code embodied therein, either in baseband or as part of a carrier wave. Such a propagated data signal may take any of a variety of forms, including, but not limited to, electro-magnetic, optical, or any suitable combination of the foregoing. A computer readable signal medium may also be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

Computer program code for carrying out operations of the present invention may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, smalltalk, C ++ and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any kind of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or may be connected to an external computer (for example, through the Internet using an Internet service provider).

Note that the above is only a preferred embodiment of the present invention and the technical principle applied. It will be understood by those skilled in the art that the present invention is not limited to the particular embodiments described herein, but is capable of various obvious changes, rearrangements and substitutions as will now become apparent to those skilled in the art without departing from the scope of the invention. Therefore, while the invention has been described in connection with the above embodiments, the invention is not limited to the embodiments, but may be embodied in many other equivalent forms without departing from the spirit or scope of the invention, which is set forth in the following claims.

Claims

1. A position information determination method, comprising:

inputting the second three-dimensional image into a precise position determining neural network to obtain precise position information of a second part, wherein the first part comprises the second part;

the initial candidate position information includes initial candidate frame position information corresponding to the first portion, and based on the initial candidate position information, extracting a second three-dimensional image from the first three-dimensional image includes:

determining a conversion ratio of the neural network to the image size by using the image size of the first three-dimensional image and the candidate position, and converting initial candidate frame position information into final candidate frame position information;

and extracting a second three-dimensional image from the first three-dimensional image according to the final candidate frame position information.

2. The method of claim 1, wherein extracting a second three-dimensional image from the first three-dimensional image based on the final candidate frame position information, comprises:

Expanding a frame selection range of a candidate frame in the first three-dimensional image based on the final candidate frame position information;

and extracting a second three-dimensional image from the first three-dimensional image according to the expanded frame selection range.

3. The method of claim 1, further comprising, prior to inputting the first three-dimensional image into the candidate position-determining neural network:

after extracting a second three-dimensional image from the first three-dimensional image according to the candidate position information, further comprising:

isotropically transforming the second three-dimensional image into a second three-dimensional image having a second voxel whose volume is a second volume, wherein the first volume is larger than the second volume.

4. A method according to any one of claims 1-3, wherein inputting the second three-dimensional image into a precise location determining neural network to obtain precise location information of the second location comprises:

And determining the accurate position information of the second part according to the Gaussian thermal nuclear diagram.

5. A method according to any of claims 1-3, wherein the candidate position-determining neural network comprises at least one level of encoder-type depth neural network, and the precise position-determining neural network comprises an encoder-decoder type symmetric depth neural network and an encoder-decoder type asymmetric depth neural network; wherein the encoder-decoder type symmetric deep neural network comprises any one of a U-Net deep convolutional neural network, a V-Net deep convolutional neural network and a HourgassNet deep convolutional neural network.

6. A method according to any one of claims 1-3, wherein the method is applied to any one of aortic valve anatomical keypoint detection, aortic root positioning and segmentation, aortic valve leaflet calcification detection, perihepatic vein junction vessel positioning and segmentation, perihepatic left and right branch bifurcation vessel positioning and segmentation, subhepatic right vein and inferior vena cava junction positioning and segmentation, perisplenic portal vessel positioning and segmentation, perirenal portal vessel positioning and segmentation, and common iliac artery left and right branch vessel positioning and segmentation.

7. A position information determining apparatus, comprising:

the accurate position information determining module is used for inputting the second three-dimensional image into an accurate position determining neural network to obtain accurate position information of a second part, and the first part comprises the second part;

the second three-dimensional image extraction module includes:

8. A computer device, comprising:

One or more processors;

storage means for storing one or more programs,

the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the location information determination method of any of claims 1-6.

9. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the position information determination method according to any one of claims 1-6.