CN110717910B - CT image target detection method based on convolutional neural network and CT scanner - Google Patents
CT image target detection method based on convolutional neural network and CT scanner Download PDFInfo
- Publication number
- CN110717910B CN110717910B CN201910984904.5A CN201910984904A CN110717910B CN 110717910 B CN110717910 B CN 110717910B CN 201910984904 A CN201910984904 A CN 201910984904A CN 110717910 B CN110717910 B CN 110717910B
- Authority
- CN
- China
- Prior art keywords
- current frame
- frame image
- target
- composite
- output data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B6/00—Apparatus for radiation diagnosis, e.g. combined with radiation therapy equipment
- A61B6/02—Devices for diagnosis sequentially in different planes; Stereoscopic radiation diagnosis
- A61B6/03—Computerised tomographs
- A61B6/032—Transmission computed tomography [CT]
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B6/00—Apparatus for radiation diagnosis, e.g. combined with radiation therapy equipment
- A61B6/52—Devices using data or image processing specially adapted for radiation diagnosis
- A61B6/5211—Devices using data or image processing specially adapted for radiation diagnosis involving processing of medical diagnostic data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/20—Image enhancement or restoration by the use of local operators
- G06T5/30—Erosion or dilatation, e.g. thinning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/50—Image enhancement or restoration by the use of more than one image, e.g. averaging, subtraction
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10081—Computed x-ray tomography [CT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20212—Image combination
- G06T2207/20221—Image fusion; Image merging
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Abstract
The embodiment of the invention provides a CT image target detection method and a CT scanner, wherein the CT image target detection method comprises the following steps: obtaining a residual error area based on a previous frame image and a next frame image of a current frame image in a video; fusing the residual error area with the current frame image to obtain a fused current frame image; carrying out high-pass filtering on the fused current frame image to obtain a high-frequency current frame image; carrying out low-pass filtering on the fused current frame image to obtain a low-frequency current frame image; fusing the high-frequency current frame image and the low-frequency frame image to obtain a composite current frame image; performing target detection on the current frame image based on the first convolution neural network to obtain a current frame target area; performing target detection on the composite current frame image based on a second convolutional neural network to obtain a composite target area; obtaining the distance between the current frame target area and the composite target area; and if the distance is smaller than the target value, acquiring a target image to be detected based on the current frame target area and the composite target area.
Description
Technical Field
The invention relates to the technical field of computers, in particular to a CT image target detection method based on a convolutional neural network and a CT scanner.
Background
Target detection is a relatively mature computer technology, and various methods such as a region-based target detection method and a texture-based target detection method exist, but the detection accuracy of these traditional methods is low. Currently, the target detection has a good effect of performing target detection based on a neural network. However, the target detection method based on the neural network depends on a large number of training samples, and the target detection method based on the neural network does not consider the difference of the training samples, which may result in an insufficient accuracy of the detection result for field application. In the scenes such as CT image target detection and the like which need to detect a target with high precision, the target detection result of a target detection method only depending on a neural network or a traditional target detection method is unsatisfactory.
Therefore, a method for detecting a target in a CT image, which can improve the accuracy of target detection, is needed.
Disclosure of Invention
The embodiment of the invention provides a CT image target detection method based on a convolutional neural network, which aims to solve the problems in the prior art.
In a first aspect, an embodiment of the present invention provides a method for detecting a CT image target based on a convolutional neural network, where the method includes:
acquiring a CT video of a to-be-detected region scanned by a CT scanner;
obtaining a residual error region based on a previous frame image and a next frame image of a current frame image in the CT video;
fusing the residual error area with the current frame image to obtain a fused current frame image;
carrying out high-pass filtering on the fused current frame image to obtain a high-frequency current frame image;
carrying out low-pass filtering on the fused current frame image to obtain a low-frequency current frame image;
fusing the high-frequency current frame image and the low-frequency frame image to obtain a composite current frame image;
performing target detection on the current frame image based on a first convolution neural network to obtain a current frame target area;
performing target detection on the composite current frame image based on a second convolutional neural network to obtain a composite target area;
obtaining the distance between the current frame target area and the composite target area;
and if the distance is smaller than the target value, acquiring a target image to be detected based on the current frame target area and the composite target area.
Optionally, the fusing the residual error region with the current frame image to obtain a fused current frame image includes:
obtaining the sum of the pixel value of a pixel point (i, j) in the residual region and the pixel value of a pixel point (i, j + k) in the current frame image, wherein i, j is a positive integer, and k is an integer greater than or equal to 0;
if the sum is larger than 255, the pixel value of the pixel point (i, j) of the fused current frame image is a first difference value, and the first difference value is a difference value between 255 and a remainder of a quotient of the sum and 255;
and if the sum is not more than 255, the pixel value of the pixel point (i, j) of the fused current frame image is the sum.
Optionally, the fusing the high-frequency current frame image and the low-frequency frame image to obtain a composite current frame image includes:
obtaining the pixel value of a pixel point (i, j) in the high-frequency current frame image and the average value of the pixel point (i, j) in the low-frequency current frame image;
and determining the average value as the pixel value of the pixel point (i, j) of the composite current frame image.
Optionally, the performing target detection on the current frame image based on the first convolutional neural network to obtain a current frame target region includes:
after performing convolution processing on the current frame image for at least one time, obtaining first output data;
performing pooling processing on the first output data twice to obtain second output data;
after performing convolution processing on the second output data for at least two times, obtaining third output data;
performing pooling processing on the third output data for at least three times to obtain fourth output data;
and classifying the fourth output data to obtain the current target area.
Optionally, the performing target detection on the composite current frame image based on the second convolutional neural network to obtain a composite target region includes:
performing pooling processing on the composite current frame image for at least three times to obtain fifth output data;
performing convolution processing on the fifth output data for at least two times to obtain sixth output data;
performing pooling processing and convolution processing on the sixth output data to obtain seventh output data;
fusing the sixth output data and the seventh output data to obtain eighth output data;
performing pooling processing and convolution processing at least twice on the eighth output data to obtain ninth output data;
and classifying the ninth output data to obtain the composite target area.
Optionally, the obtaining the distance between the current frame target region and the composite target region includes:
acquiring circumscribed circles of the current frame target area and the composite target area;
obtaining the Euclidean distance between the circle center of the circumscribed circle of the current frame target area and the circle center of the circumscribed circle of the composite target area;
and determining the Euclidean distance as the distance between the current frame target region and the composite target region.
Optionally, the obtaining the distance between the current frame target region and the composite target region includes:
respectively obtaining the gravity centers of the current frame target area and the composite target area;
obtaining the Euclidean distance between the gravity center of the current frame target area and the gravity center of the composite target area;
and determining the Euclidean distance as the distance between the current frame target region and the composite target region.
Optionally, the obtaining a target to be detected based on the current frame target region and the composite target region includes:
acquiring a cross region of the current frame target region and the composite target region;
and determining the cross region as the target to be detected.
Optionally, before determining that the intersection region is the target to be detected, the method further includes:
and rendering the intersection area.
In a second aspect, an embodiment of the present invention provides a CT scanner, which includes a scanning portion and a processing portion;
the scanning part is used for scanning a CT video of a region to be detected and sending the CT video to the processing part;
the processing portion is configured to perform any of the methods described above.
Compared with the prior art, the invention achieves the technical effects that:
the embodiment of the invention provides a CT image target detection method and a CT scanner based on a convolutional neural network, wherein the CT image target detection method comprises the following steps: acquiring a video of a detection area in a CT image to be scanned by a CT scanner; obtaining a residual error area based on a previous frame image and a next frame image of a current frame image in a video; fusing the residual error area with the current frame image to obtain a fused current frame image; carrying out high-pass filtering on the fused current frame image to obtain a high-frequency current frame image; carrying out low-pass filtering on the fused current frame image to obtain a low-frequency current frame image; fusing the high-frequency current frame image and the low-frequency frame image to obtain a composite current frame image; performing target detection on the current frame image based on a first convolution neural network to obtain a current frame target area; performing target detection on the composite current frame image based on a second convolutional neural network to obtain a composite target area; obtaining the distance between the current frame target area and the composite target area; and if the distance is smaller than the target value, acquiring a target image to be detected based on the current frame target area and the composite target area.
Obtaining a residual error region based on a previous frame image and a next frame image of a current frame image in a video, and fusing the residual error region and the current frame image to obtain a fused current frame image, so that the characteristic information of the current frame image is enhanced; the high-pass filtering is carried out on the fused current frame image to obtain a high-frequency current frame image, the high-frequency current frame image reserves the high-frequency characteristic information of the current frame image, the low-pass filtering is carried out on the fused current frame image to obtain a low-frequency current frame image, the low-frequency current frame image reserves the low-frequency characteristic information of the current frame image, the high-frequency current frame image and the low-frequency frame image are fused to obtain a composite current frame image, the characteristic information in the composite current frame image is enhanced, and meanwhile the fidelity of the characteristic information is improved. Performing target detection on the current frame image based on the first convolution neural network to obtain a current frame target area; performing target detection on the composite current frame image based on the second convolutional neural network to obtain a composite target area, further improving the probability of the target in the target area and improving the accuracy of target detection; the method comprises the steps of obtaining the distance between a current frame target area and a composite target area, obtaining a target image to be detected based on the current frame target area and the composite target area if the distance is smaller than a target value, combining a target detection result of a traditional neural network and a target detection result of a second neural network with high accuracy, improving the accuracy of target detection and improving the precision of the target to be detected. The technical problem of low target detection precision in the prior art is solved, and the technical effect of improving the target detection accuracy is achieved.
Drawings
Fig. 1 shows a flowchart of a CT image target detection method based on a convolutional neural network according to an embodiment of the present invention.
Fig. 2 is a schematic structural diagram of a first convolutional neural network according to an embodiment of the present invention.
Fig. 3 is a schematic structural diagram illustrating a second convolutional neural network according to an embodiment of the present invention.
Fig. 4 is a schematic diagram illustrating a schematic structure of a CT scanner 100 according to an embodiment of the present invention.
Icon: 100-CT scanner; 110-a scanning section 110; 120-a processing section; 130-a display section; 140-storage section.
Detailed Description
As shown in fig. 1, an embodiment of the present invention provides a method for detecting a CT image target based on a convolutional neural network. The CT image target detection method comprises S101-S110 shown in FIG. 1. S101 to S110 are explained below with reference to fig. 1.
S101: and acquiring a CT video of the area to be detected scanned by the CT scanner. The CT video comprises a plurality of frames of images, and the area to be detected is a diseased area or an area to be detected. The images may be referred to as CT images.
S102: and obtaining a residual error region based on a previous frame image and a next frame image of a current frame image in the CT video.
S103: and fusing the residual error area with the current frame image to obtain a fused current frame image.
S104: and carrying out high-pass filtering on the fused current frame image to obtain a high-frequency current frame image.
S105: and carrying out low-pass filtering on the fused current frame image to obtain a low-frequency current frame image.
S106: and fusing the high-frequency current frame image and the low-frequency frame image to obtain a composite current frame image.
S107: and performing target detection on the current frame image based on the first convolution neural network to obtain a current frame target area.
S108: and performing target detection on the composite current frame image based on the second convolutional neural network to obtain a composite target area.
S109: and obtaining the distance between the current frame target area and the composite target area.
S110: and if the distance is smaller than the target value, acquiring a target image to be detected based on the current frame target area and the composite target area.
By adopting the scheme, the residual error region is obtained based on the previous frame image and the next frame image of the current frame image in the video, the residual error region and the current frame image are fused to obtain the fused current frame image, and the characteristic information of the current frame image is enhanced. The high-pass filtering is carried out on the fused current frame image to obtain a high-frequency current frame image, the high-frequency current frame image reserves the high-frequency characteristic information of the current frame image, the low-pass filtering is carried out on the fused current frame image to obtain a low-frequency current frame image, the low-frequency current frame image reserves the low-frequency characteristic information of the current frame image, the high-frequency current frame image and the low-frequency frame image are fused to obtain a composite current frame image, the characteristic information in the composite current frame image is enhanced, and meanwhile the fidelity of the characteristic information is improved. Performing target detection on the current frame image based on the first convolution neural network to obtain a current frame target area; and performing target detection on the composite current frame image based on the second convolutional neural network to obtain a composite target region, further improving the probability of the target in the target region, and improving the accuracy of target detection. The method comprises the steps of obtaining the distance between a current frame target area and a composite target area, obtaining a target to be detected based on the current frame target area and the composite target area if the distance is smaller than a target value, combining a target detection result of a traditional neural network and a target detection result of a second neural network with high accuracy, improving the accuracy of target detection and improving the precision of the target to be detected.
Wherein, S102 specifically is: and obtaining a front pixel point of the previous frame of image and a rear pixel point corresponding to the position of the front pixel point in the next frame of image. And obtaining first difference value information between the front pixel point and the rear pixel point through absolute difference value sum operation based on the front block and the rear block corresponding to the front pixel point. The front block is a block in the previous frame of image and the rear block is an area in the next frame of image. The front area block comprises a plurality of pixel points. When the previous block corresponding to the pixel point is a rectangular block with the previous pixel point as the center and the size is set, for example, the current block is a rectangular block of a block of 2 × 2. If the position of the previous pixel point is at the edge of the current frame image, the previous block comprises a plurality of pixel points which are adjacent to the current pixel point and are obtained by taking the current pixel point as the center, and a determined block formed by the pixel points is the current block. For example, if the front pixel is (0, 0), the front block is a block composed of (0, 0), (0, 1), (1, 0), and (1, 1). Aiming at each pixel point in the previous block, obtaining the difference value between the value of each pixel point and the value of the pixel point corresponding to the position of each pixel point in the later block; and carrying out summation operation on the absolute values of the differences to obtain first difference information, wherein a plurality of front pixel points correspond to a plurality of first difference information, and the plurality of first difference information form a residual block according to the corresponding relation with the front pixel points. Each pixel point in each previous block corresponds to a difference, and specifically, the absolute value of each difference is summed. In order to obtain a rear block corresponding to the position of the front block in the previous frame of image, a rear block corresponding to the position of the front block is obtained in the next frame of image, and each pixel point in the current block corresponds to each pixel point in the rear block in a one-to-one position. The position correspondence refers to position one-to-one correspondence, for example, the position of the front pixel point is the same as the position of the rear pixel point, which is specifically embodied that the value of the position of the front pixel point is the same as the value of the position of the rear pixel point, for example, if the value of the position of the front pixel point is (1, 2) and the value of the position of the rear pixel point is (1, 2), the front pixel point corresponds to the position of the rear pixel point. Thus, the sizes of the front and rear blocks are consistent. Specifically, the first difference information is obtained by the following formula (1).
Wherein a (i, j) represents the value of the pixel point (i, j) in the front block corresponding to the front pixel point (m, n), b (i, j) represents the value of the pixel point corresponding to the position of the pixel point (i, j) in the rear block, k represents the number of the pixel points of the front block in the horizontal axis direction, and s1(m, n) represents the first difference information. The method comprises the steps of obtaining the absolute value of an obtained difference value by adopting the value of each pixel point in a front block corresponding to a front pixel point, subtracting the value of each pixel point corresponding to each pixel point of the front block in a rear block corresponding to the front pixel point, summing the absolute values corresponding to the pixel points, obtaining first difference value information, and enabling the obtained residual block to accurately represent the difference of a front frame image relative to a rear frame image.
The method for fusing the residual error region with the current frame image to obtain a fused current frame image specifically comprises the following steps:
firstly, the sum of the pixel value of the pixel point (i, j) in the residual region and the pixel value of the pixel point (i, j + k) in the current frame image is obtained, wherein i, j is a positive integer, and k is an integer greater than or equal to 0. Then, if the sum is greater than 255, the pixel value of the pixel point (i, j) of the fused current frame image is a first difference value, and the first difference value is a difference value between 255 and a remainder of a quotient of the sum and 255. For example, the sum of the pixel value of the pixel point (i, j) in the residual region and the pixel value of the pixel point (i, j + k) in the current frame image is Y, Y >255, and the remainder of Y/255 is X, then the first difference value is equal to 255-X. If the sum is not more than 255, the pixel value of the pixel point (i, j) of the fused current frame image is the sum. Namely, the pixel value of the pixel point (i, j) of the fused current frame image is Y. By adopting the scheme, the characteristics that the pixel values of the current frame image are fused with the pixel values of the residual error region and the pixel points of the current frame image are fused, the characteristics of the fused current frame image are enhanced, meanwhile, the fidelity of the fused current frame image is ensured, and further, a foundation is laid for improving the precision of target detection.
The method comprises the following steps of fusing a high-frequency current frame image and a low-frequency frame image to obtain a composite current frame image, and specifically comprises the following steps: obtaining the pixel value of a pixel point (i, j) in the high-frequency current frame image and the average value of the pixel point (i, j) in the low-frequency current frame image; and determining the average value as the pixel value of the pixel point (i, j) of the composite current frame image. The accuracy of the characteristics of the composite current frame image is improved.
As an optional implementation manner, S106 is specifically: if the pixel value of the pixel point (i, j) in the high-frequency current frame image is equal to the pixel value of the pixel point (i, j) in the low-frequency current frame image, and the pixel value of the pixel point is a first value, the pixel value of the pixel point (i, j) of the composite current frame image is assigned to be a second value; if the pixel value of the pixel point (i, j) in the high-frequency current frame image is the same as that of the pixel point (i, j) in the low-frequency current frame image, and the pixel value of the pixel point is not a first value, the pixel point (i, j) of the composite current frame image is expanded, so that the pixel point (i, j) comprises a fusion channel; and assigning the fusion channel so as to enable the value of the fusion channel to be a second value, wherein the second value is different from the pixel value of the pixel point and the first value. By adopting the scheme, the obtained composite current frame image comprises the pixel information in the high-frequency current frame image and the pixel information in the low-frequency current frame image, the characteristics of the composite current frame image are enhanced, and the accuracy of the detected target is improved.
In the embodiment of the present invention, the step of performing target detection on the current frame image based on the first convolution neural network to obtain the current frame target area specifically includes: after performing convolution processing on the current frame image for at least one time, obtaining first output data; performing pooling processing on the first output data twice to obtain second output data; after carrying out convolution processing on the second output data for at least two times, obtaining third output data; performing pooling processing on the third output data for at least three times to obtain fourth output data; and classifying the fourth output data to obtain the current target area.
In order to clearly illustrate the step of performing target detection on the current frame image based on the first convolutional neural network to obtain the target region of the current frame, as shown in fig. 2, fig. 2 shows a schematic structural diagram of the first convolutional neural network, that is, the first convolutional neural network includes at least four convolutional layers, at least five pooling layers, and one classification layer. The first convolutional neural network may have a structure of a convolutional layer, a pooling layer, a convolutional layer, a pooling layer, and a classification layer. The first convolutional neural network may also have a structure of a convolutional layer, a pooling layer, a convolutional layer, a pooling layer, or a classification layer. Wherein the classification layer may be a support vector machine model. And the pooling layer connected with the classification layer outputs an image characteristic vector, and the classification layer classifies and identifies the image characteristic vector to obtain a current frame target area.
By adopting the scheme, the long convolution kernel pooling operation is carried out on the current frame image, and finally the classification layer carries out classification and identification on the image feature vectors to obtain the current frame target area, so that the accuracy of identifying and detecting the current frame target area is improved.
Optionally, the target detection is performed on the composite current frame image based on the second convolutional neural network to obtain a composite target region, specifically: performing pooling processing on the composite current frame image for at least three times to obtain fifth output data; performing convolution processing on the fifth output data for at least two times to obtain sixth output data; performing pooling processing and convolution processing on the sixth output data to obtain seventh output data; fusing the sixth output data and the seventh output data to obtain eighth output data; performing pooling processing and convolution processing at least twice on the eighth output data to obtain ninth output data; and classifying the ninth output data to obtain a composite target area.
The second convolutional neural network at least comprises six pooling layers, at least five convolutional layers, a fusion layer and a classification layer. A schematic diagram of a second convolutional neural network shown in fig. 3. The structure of the second convolutional neural network may be a pooling layer, a convolutional layer, a pooling layer, a convolutional layer, a fusion layer, a pooling layer, a convolutional layer, a classification layer. The structure of the second convolutional neural network may be a pooling layer, a convolutional layer, a pooling layer, a convolutional layer, a fusion layer, a pooling layer, a convolutional layer, a sorting layer. Wherein the classification layer may be a support vector machine model as described above. The fusion layer is used for respectively carrying out merging processing on the output of the neurons of the convolution layer connected with the fusion layer, the merging mode can be weighted accumulation or OR operation and the like to obtain a merging processing result, and then the merging processing result is output to the pooling layer for pooling processing. And finally, outputting the region vector by the convolution layer, and classifying the region vector through a classification layer to obtain a composite target region.
By adopting the scheme, the composite current frame image is subjected to multi-length convolution kernel pooling operation, the final convolution layer outputs the region vector, the region vector is classified through the classification layer to obtain the target region, and the precision of identifying and detecting the composite target region is improved.
Optionally, a specific implementation manner of obtaining the distance between the current frame target region and the composite target region is as follows:
acquiring circumscribed circles of a current frame target area and a composite target area; obtaining the Euclidean distance between the circle center of the circumscribed circle of the current frame target area and the circle center of the circumscribed circle of the composite target area; and determining the Euclidean distance as the distance between the current frame target area and the composite target area. The method for obtaining the distance between the current frame target area and the composite target area specifically comprises the following steps: respectively obtaining the gravity centers of a current frame target area and a composite target area; obtaining the Euclidean distance between the gravity center of the current frame target area and the gravity center of the composite target area; and determining the Euclidean distance as the distance between the current frame target area and the composite target area.
The specific implementation manner of obtaining the target to be detected based on the current frame target region and the composite target region in S110 may be: acquiring a cross region of a current frame target region and a composite target region; and determining the cross region as a target to be detected. Thus, the accuracy of the target to be detected is improved.
In order to further improve the accuracy of the target to be detected, before determining the intersection region as the target to be detected, the CT image target detection method further includes: and rendering the intersection area. The rendering method may be color rendering according to pathological features and color features of the examined region, for example, if the image is a CT image of a lung, the rendered object to be detected is a color image of the lung. If the image is a CT image of the brain, the rendered target to be detected is a gray-scale brain image or a head skeleton image and the like.
In order to improve the rendering accuracy and further improve the target detection accuracy, before rendering the intersection region, the CT image target detection method further includes: and performing expansion processing on the intersection area so as to enable the size of the intersection area to be close to that of the target to be measured. And expanding the intersection area to enable the size of the intersection area to be close to that of the target to be detected, so that the definition of the rendered intersection area is improved, and the accuracy of target detection is improved.
In the embodiment of the present invention, the target value may be a preset value according to experience. In order to further improve the accuracy of target detection, the target value is obtained by: obtaining a first target vector based on a current frame target area; obtaining a second target vector based on the composite target region; classifying the first target vector based on the trained support vector machine model to obtain a first target value; classifying the second target vector based on the trained support vector machine model to obtain a second target value; and obtaining the difference between the first target value and the second target value to obtain a target difference value, and taking the target difference value as a target value.
As an alternative embodiment, the CT image target detection method includes: and acquiring a CT video of the area to be detected scanned by the CT scanner. And obtaining a residual error region based on a previous frame image and a next frame image of a current frame image in the CT video. And fusing the residual error area with the current frame image to obtain a fused current frame image. And carrying out high-pass filtering on the fused current frame image to obtain a high-frequency current frame image. And carrying out low-pass filtering on the fused current frame image to obtain a low-frequency current frame image. And fusing the high-frequency current frame image and the low-frequency frame image to obtain a composite current frame image. And performing target detection on the current frame image based on the first convolution neural network to obtain a current frame target area. A first target vector is obtained based on a current frame target region. And performing target detection on the composite current frame image based on the second convolutional neural network to obtain a composite target area. A second target vector is obtained based on the composite target region. Classifying the first target vector based on the trained support vector machine model to obtain a first target value; classifying the second target vector based on the trained support vector machine model to obtain a second target value; and obtaining the difference between the first target value and the second target value to obtain a target difference value, and taking the target difference value as the target value. And obtaining the distance between the current frame target area and the composite target area. And if the distance is smaller than the target value, obtaining a target image to be detected based on the current frame target area and the composite target area.
By adopting the scheme, the residual error region is obtained by adopting the scheme based on the previous frame image and the next frame image of the current frame image in the video, the residual error region and the current frame image are fused to obtain the fused current frame image, and the characteristic information of the current frame image is enhanced. The high-pass filtering is carried out on the fused current frame image to obtain a high-frequency current frame image, the high-frequency current frame image reserves the high-frequency characteristic information of the current frame image, the low-pass filtering is carried out on the fused current frame image to obtain a low-frequency current frame image, the low-frequency current frame image reserves the low-frequency characteristic information of the current frame image, the high-frequency current frame image and the low-frequency frame image are fused to obtain a composite current frame image, the characteristic information in the composite current frame image is enhanced, and meanwhile the fidelity of the characteristic information is improved. Performing target detection on the current frame image based on the first convolution neural network to obtain a current frame target area; and performing target detection on the composite current frame image based on the second convolutional neural network to obtain a composite target region, further improving the probability of the target in the target region, and improving the accuracy of target detection. Obtaining a first target vector based on a current frame target area, obtaining a second target vector based on a composite target area, and classifying the first target vector based on a trained support vector machine model to obtain a first target value; classifying the second target vector based on the trained support vector machine model to obtain a second target value; and obtaining the difference between the first target value and the second target value to obtain a target difference value, and taking the target difference value as the target value. The method comprises the steps of obtaining the distance between a current frame target area and a composite target area, obtaining a target to be detected based on the current frame target area and the composite target area if the distance is smaller than a target value, combining a target detection result of a traditional neural network and a target detection result of a second neural network with high accuracy, improving the accuracy of target detection and improving the precision of the target to be detected. Meanwhile, the target value is obtained based on the target area of the previous frame and the composite target area, and the target value is used as a judgment reference, so that the target detection accuracy can be improved.
The CT image target detection method can be used for detecting infection positions, heart positions, lung positions and the like in the CT image. But is not limited thereto.
On the basis of the CT image target detection method, the embodiment of the invention also provides a CT scanner. The CT scanner 100 includes a scanning unit 110, a processing unit 120, and a display unit 130. As shown in fig. 4, the scanning unit 110 is configured to scan a CT image of a region to be detected and transmit the CT image to the processing unit 120.
The processing unit 120 is configured to execute any of the above CT image target detection methods to perform target detection on a CT image in a CT video to obtain a target image to be detected, and send the obtained target image to be detected to the display unit 130.
The display unit 130 is used to display an object image to be detected.
When the processing unit 120 executes any of the above CT image target detection methods, it executes the following functions: and acquiring a CT video of the area to be detected scanned by the CT scanner. And obtaining a residual error region based on a previous frame image and a next frame image of a current frame image in the CT video. And fusing the residual error area with the current frame image to obtain a fused current frame image. And carrying out high-pass filtering on the fused current frame image to obtain a high-frequency current frame image. And carrying out low-pass filtering on the fused current frame image to obtain a low-frequency current frame image. And fusing the high-frequency current frame image and the low-frequency frame image to obtain a composite current frame image. And performing target detection on the current frame image based on the first convolution neural network to obtain a current frame target area. And performing target detection on the composite current frame image based on the second convolutional neural network to obtain a composite target area. And obtaining the distance between the current frame target area and the composite target area. And if the distance is smaller than the target value, acquiring a target image to be detected based on the current frame target area and the composite target area.
By adopting the scheme, the residual error region is obtained by adopting the scheme based on the previous frame image and the next frame image of the current frame image in the video, the residual error region and the current frame image are fused to obtain the fused current frame image, and the characteristic information of the current frame image is enhanced. The high-pass filtering is carried out on the fused current frame image to obtain a high-frequency current frame image, the high-frequency current frame image reserves the high-frequency characteristic information of the current frame image, the low-pass filtering is carried out on the fused current frame image to obtain a low-frequency current frame image, the low-frequency current frame image reserves the low-frequency characteristic information of the current frame image, the high-frequency current frame image and the low-frequency frame image are fused to obtain a composite current frame image, the characteristic information in the composite current frame image is enhanced, and meanwhile the fidelity of the characteristic information is improved.
Performing target detection on the current frame image based on the first convolution neural network to obtain a current frame target area; and performing target detection on the composite current frame image based on the second convolutional neural network to obtain a composite target region, further improving the probability of the target in the target region, and improving the accuracy of target detection.
Obtaining a first target vector based on a current frame target area, obtaining a second target vector based on a composite target area, and classifying the first target vector based on a trained support vector machine model to obtain a first target value; classifying the second target vector based on the trained support vector machine model to obtain a second target value; and obtaining the difference between the first target value and the second target value to obtain a target difference value, and taking the target difference value as the target value. The method comprises the steps of obtaining the distance between a current frame target area and a composite target area, obtaining a target to be detected based on the current frame target area and the composite target area if the distance is smaller than a target value, combining a target detection result of a traditional neural network and a target detection result of a second neural network with high accuracy, improving the accuracy of target detection and improving the precision of the target to be detected. Meanwhile, the target value is obtained based on the target area of the previous frame and the composite target area, and the target value is used as a judgment reference, so that the target detection accuracy can be improved.
Optionally, the CT scanner 100 further includes a storage unit 140, where the storage unit is configured to store the detected target image and any data in the CT image target detection method, and specifically, when the composite current frame image is stored, for each pixel point, the storage unit is divided into two storage sections, where a first storage section is used to store a value of the fusion channel, and a second storage section is used to store position information and RGB information of the pixel point. Therefore, the accuracy and the effectiveness of storage are improved, and the fidelity of the stored composite current frame image is ensured.
Claims (10)
1. A CT image target detection method based on a convolutional neural network is characterized by comprising the following steps:
acquiring a CT video of a to-be-detected region scanned by a CT scanner;
obtaining a residual error region based on a previous frame image and a next frame image of a current frame image in the CT video;
fusing the residual error area with the current frame image to obtain a fused current frame image;
carrying out high-pass filtering on the fused current frame image to obtain a high-frequency current frame image;
carrying out low-pass filtering on the fused current frame image to obtain a low-frequency current frame image;
fusing the high-frequency current frame image and the low-frequency current frame image to obtain a composite current frame image;
performing target detection on the current frame image based on a first convolution neural network to obtain a current frame target area;
performing target detection on the composite current frame image based on a second convolutional neural network to obtain a composite target area;
obtaining the distance between the current frame target area and the composite target area;
and if the distance is smaller than the target value, acquiring a target image to be detected based on the current frame target area and the composite target area.
2. The method according to claim 1, wherein said fusing the residual region with the current frame image to obtain a fused current frame image comprises:
obtaining the sum of the pixel value of a pixel point (i, j) in the residual region and the pixel value of a pixel point (i, j + k) in the current frame image, wherein i, j is a positive integer, and k is an integer greater than or equal to 0;
if the sum is larger than 255, the pixel value of the pixel point (i, j) of the fused current frame image is a first difference value, and the first difference value is a difference value between 255 and a remainder of a quotient of the sum and 255;
and if the sum is not more than 255, the pixel value of the pixel point (i, j) of the fused current frame image is the sum.
3. The method according to claim 1, wherein said fusing the high-frequency current frame image and the low-frequency current frame image to obtain a composite current frame image comprises:
obtaining the pixel value of a pixel point (i, j) in the high-frequency current frame image and the average value of the pixel point (i, j) in the low-frequency current frame image;
and determining the average value as the pixel value of the pixel point (i, j) of the composite current frame image.
4. The method of claim 1, wherein the performing target detection on the current frame image based on the first convolutional neural network to obtain a current frame target region comprises:
after performing convolution processing on the current frame image for at least one time, obtaining first output data;
performing pooling processing on the first output data twice to obtain second output data;
after performing convolution processing on the second output data for at least two times, obtaining third output data;
performing pooling processing on the third output data for at least three times to obtain fourth output data;
and classifying the fourth output data to obtain the current target area.
5. The method of claim 1, wherein the performing target detection on the composite current frame image based on the second convolutional neural network to obtain a composite target region comprises:
performing pooling processing on the composite current frame image for at least three times to obtain fifth output data;
performing convolution processing on the fifth output data for at least two times to obtain sixth output data;
performing pooling processing and convolution processing on the sixth output data to obtain seventh output data;
fusing the sixth output data and the seventh output data to obtain eighth output data;
performing pooling processing and convolution processing at least twice on the eighth output data to obtain ninth output data;
and classifying the ninth output data to obtain the composite target area.
6. The method of claim 1, wherein obtaining the distance between the current frame target region and the composite target region comprises:
acquiring circumscribed circles of the current frame target area and the composite target area;
obtaining the Euclidean distance between the circle center of the circumscribed circle of the current frame target area and the circle center of the circumscribed circle of the composite target area;
and determining the Euclidean distance as the distance between the current frame target region and the composite target region.
7. The method of claim 1, wherein obtaining the distance between the current frame target region and the composite target region comprises:
respectively obtaining the gravity centers of the current frame target area and the composite target area;
obtaining the Euclidean distance between the gravity center of the current frame target area and the gravity center of the composite target area;
and determining the Euclidean distance as the distance between the current frame target region and the composite target region.
8. The method according to claim 1, wherein the obtaining the target image to be detected based on the current frame target region and the composite target region comprises:
acquiring a cross region of the current frame target region and the composite target region;
and determining the cross region as the target to be detected.
9. The method of claim 8, wherein prior to said determining said intersection region as said object to be detected, said method further comprises:
and rendering the intersection area.
10. A CT scanner is characterized in that the CT scanner comprises a scanning part and a processing part;
the scanning part is used for scanning a CT video of a region to be detected and sending the CT video to the processing part;
the processing portion is configured to perform the method of any one of claims 1-9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910984904.5A CN110717910B (en) | 2019-10-16 | 2019-10-16 | CT image target detection method based on convolutional neural network and CT scanner |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910984904.5A CN110717910B (en) | 2019-10-16 | 2019-10-16 | CT image target detection method based on convolutional neural network and CT scanner |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110717910A CN110717910A (en) | 2020-01-21 |
| CN110717910B true CN110717910B (en) | 2020-09-08 |
Family
ID=69211787
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910984904.5A Active CN110717910B (en) | 2019-10-16 | 2019-10-16 | CT image target detection method based on convolutional neural network and CT scanner |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110717910B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111079546B (en) * | 2019-11-22 | 2022-06-07 | 重庆师范大学 | Unmanned aerial vehicle pest detection method |
| CN112132480A (en) * | 2020-09-29 | 2020-12-25 | 姜锡忠 | Master and resource matching method and system for big data online education platform |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1717058A (en) * | 2004-06-29 | 2006-01-04 | 三洋电机株式会社 | Method and apparatus for coding images and method and apparatus for decoding the images |
| CN103400360A (en) * | 2013-08-03 | 2013-11-20 | 浙江农林大学 | Multi-source image fusing method based on Wedgelet and NSCT (Non Subsampled Contourlet Transform) |
| CN107909556A (en) * | 2017-11-27 | 2018-04-13 | 天津大学 | Video image rain removing method based on convolutional neural networks |
| CN109035260A (en) * | 2018-07-27 | 2018-12-18 | 京东方科技集团股份有限公司 | A kind of sky areas dividing method, device and convolutional neural networks |
| CN110060225A (en) * | 2019-03-28 | 2019-07-26 | 南京信息工程大学 | A kind of Medical image fusion method based on rapid finite shearing wave conversion and rarefaction representation |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109461495B (en) * | 2018-11-01 | 2023-04-14 | 腾讯科技（深圳）有限公司 | Medical image recognition method, model training method and server |
-
2019
- 2019-10-16 CN CN201910984904.5A patent/CN110717910B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1717058A (en) * | 2004-06-29 | 2006-01-04 | 三洋电机株式会社 | Method and apparatus for coding images and method and apparatus for decoding the images |
| CN103400360A (en) * | 2013-08-03 | 2013-11-20 | 浙江农林大学 | Multi-source image fusing method based on Wedgelet and NSCT (Non Subsampled Contourlet Transform) |
| CN107909556A (en) * | 2017-11-27 | 2018-04-13 | 天津大学 | Video image rain removing method based on convolutional neural networks |
| CN109035260A (en) * | 2018-07-27 | 2018-12-18 | 京东方科技集团股份有限公司 | A kind of sky areas dividing method, device and convolutional neural networks |
| CN110060225A (en) * | 2019-03-28 | 2019-07-26 | 南京信息工程大学 | A kind of Medical image fusion method based on rapid finite shearing wave conversion and rarefaction representation |
Non-Patent Citations (2)
| Title |
|---|
| Medical Image Super Resolution Using Improved Generative Adversarial Networks;XINYANG BING 等;《IEEE Access》;20191001;第1-9页 * |
| 医学影像配准融合分割技术及其在肺癌精准医疗的应用;史益新;《中国优秀硕士学位论文全文数据库医药卫生科技辑》;20170315;正文第1-50页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110717910A (en) | 2020-01-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113658132B (en) | Computer vision-based structural part weld joint detection method | |
| JP5045371B2 (en) | Foreground / background classification apparatus, method, and program for each pixel of moving image | |
| CN109859226B (en) | Detection method of checkerboard corner sub-pixels for graph segmentation | |
| CN109211198B (en) | Intelligent target detection and measurement system and method based on trinocular vision | |
| CN109961506A (en) | A kind of fusion improves the local scene three-dimensional reconstruction method of Census figure | |
| CN105809716B (en) | Foreground extraction method integrating superpixel and three-dimensional self-organizing background subtraction method | |
| JPH0935061A (en) | Image processing method | |
| CN104217459B (en) | A kind of spheroid character extracting method | |
| CN111445389A (en) | Wide-view-angle rapid splicing method for high-resolution images | |
| CN105335952B (en) | Matching power flow computational methods and device and parallax value calculating method and equipment | |
| CN106530407A (en) | Three-dimensional panoramic splicing method, device and system for virtual reality | |
| CN110310305B (en) | Target tracking method and device based on BSSD detection and Kalman filtering | |
| CN110717910B (en) | CT image target detection method based on convolutional neural network and CT scanner | |
| CN112288758B (en) | Infrared and visible light image registration method for power equipment | |
| CN117036641A (en) | Road scene three-dimensional reconstruction and defect detection method based on binocular vision | |
| CN113095371A (en) | Feature point matching method and system for three-dimensional reconstruction | |
| CN110120012B (en) | Video stitching method for synchronous key frame extraction based on binocular camera | |
| CN108182700B (en) | Image registration method based on two-time feature detection | |
| CN112364881B (en) | Advanced sampling consistency image matching method | |
| CN110717471B (en) | B-ultrasonic image target detection method based on support vector machine model and B-ultrasonic scanner | |
| JPH05215547A (en) | Method for determining corresponding points between stereo images | |
| CN109978916B (en) | Vibe moving target detection method based on gray level image feature matching | |
| CN116342519A (en) | Image processing method based on machine learning | |
| JP2981382B2 (en) | Pattern matching method | |
| CN111695552B (en) | Multi-feature fusion underwater target modeling and optimizing method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| TA01 | Transfer of patent application right | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20200818 Address after: No.2505, unit 1, building 1, Changtong Century Commercial Plaza, No.100 Yingze West Street, Wanbailin District, Taiyuan City, Shanxi Province Applicant after: Taiyuan Huarui Xingchen Pharmaceutical Technology Co., Ltd Address before: 610000 15 / F, block a, building 1, Shangding international, Chengdu hi tech Zone, Chengdu, Sichuan Province Applicant before: Jiang Tongyuan |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant |