CN114140332A - Deep learning-based breast ultrasound enhancement method - Google Patents
Deep learning-based breast ultrasound enhancement method Download PDFInfo
- Publication number
- CN114140332A CN114140332A CN202110967233.9A CN202110967233A CN114140332A CN 114140332 A CN114140332 A CN 114140332A CN 202110967233 A CN202110967233 A CN 202110967233A CN 114140332 A CN114140332 A CN 114140332A
- Authority
- CN
- China
- Prior art keywords
- layer
- image
- gland
- instrument
- mammary gland
- 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.)
- Pending
Links
- 210000000481 breast Anatomy 0.000 title claims abstract description 51
- 238000002604 ultrasonography Methods 0.000 title claims abstract description 39
- 238000000034 method Methods 0.000 title claims abstract description 32
- 238000013135 deep learning Methods 0.000 title claims abstract description 12
- 210000004907 gland Anatomy 0.000 claims abstract description 41
- 210000005075 mammary gland Anatomy 0.000 claims abstract description 39
- 230000000762 glandular Effects 0.000 claims abstract description 22
- 210000003205 muscle Anatomy 0.000 claims abstract description 14
- 230000002708 enhancing effect Effects 0.000 claims abstract description 7
- 230000011218 segmentation Effects 0.000 claims description 22
- 238000011176 pooling Methods 0.000 claims description 18
- 238000005070 sampling Methods 0.000 claims description 18
- 230000006870 function Effects 0.000 claims description 12
- 238000012549 training Methods 0.000 claims description 8
- 238000002372 labelling Methods 0.000 claims description 7
- 238000010276 construction Methods 0.000 claims description 6
- 238000013507 mapping Methods 0.000 claims description 6
- 238000013508 migration Methods 0.000 claims description 6
- 230000005012 migration Effects 0.000 claims description 6
- 239000013598 vector Substances 0.000 claims description 6
- 230000002787 reinforcement Effects 0.000 claims 1
- 238000001514 detection method Methods 0.000 abstract description 13
- 230000009286 beneficial effect Effects 0.000 abstract description 2
- 230000003902 lesion Effects 0.000 description 9
- 210000001519 tissue Anatomy 0.000 description 3
- 206010028980 Neoplasm Diseases 0.000 description 2
- 230000003416 augmentation Effects 0.000 description 2
- 125000004122 cyclic group Chemical group 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 206010006272 Breast mass Diseases 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 230000000670 limiting effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Images
Classifications
-
- G06T5/73—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- 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
- G06T2207/10016—Video; Image sequence
-
- 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
- G06T2207/10132—Ultrasound image
-
- 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
- G06T2207/20112—Image segmentation details
- G06T2207/20132—Image cropping
-
- 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
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30068—Mammography; Breast
Abstract
The invention discloses a mammary gland ultrasonic enhancement method based on deep learning, and belongs to the field of image enhancement. The method is realized by eliminating the influence of a fat layer and a muscle layer in the mammary gland ultrasonic video and improving the quality of the mammary gland ultrasonic video of a portable instrument. Firstly, a mammary gland ultrasonic video with low image quality acquired by a portable instrument is intercepted into a mammary gland ultrasonic image, and a gland region in the mammary gland ultrasonic video is segmented by UNet, so that the false positive rate caused by the interference of a fat layer and a muscle layer is reduced; the glandular region of the fuzzy breast ultrasound image is then enhanced using a cycle-generated confrontation network (CycleGan). The image quality of the gland region in the breast ultrasonic video of the portable equipment is improved, and the image quality and the accuracy of the detection result during target detection are further improved. The invention is suitable for the fields of medical treatment and the like, is used for enhancing the detected image, and is beneficial to the popularization of portable ultrasonic equipment.
Description
Technical Field
The invention relates to a mammary gland ultrasonic enhancement method based on deep learning, and belongs to the technical field of image enhancement.
Background
In the traditional mode, a patient in a first-aid diagnosis with abnormal breasts needs to go to a hospital again and is detected by a doctor with a high-end ultrasonic instrument to obtain a breast ultrasonic video, and the doctor judges the properties of the breast lumps according to the breast ultrasonic video. In order to improve the convenience of both doctors and patients, portable ultrasonic equipment is widely popularized in some hospitals, patients can detect and acquire own breast ultrasonic videos without going out, then the breast ultrasonic videos are uploaded to an artificial intelligent medical platform, and the breast ultrasonic videos of portable instruments are automatically detected by a target detection method. The portable instrument breast ultrasound video comprises three tissues of a fat layer, a gland layer and a muscle layer, and a tumor exists in the gland layer only, so that the interference of the fat layer and the muscle layer easily causes a high false positive rate; in addition, the gland region quality of the portable instrument mammary gland ultrasonic video is generally lower than that of the high-end instrument mammary gland ultrasonic video. The above problems restrict the popularization of portable ultrasound devices.
Disclosure of Invention
The method aims to provide a mammary gland ultrasonic enhancement method based on deep learning to solve the problems of false positive and low accuracy rate in automatic detection of a mammary gland ultrasonic video lesion area of a portable instrument. The method is suitable for the fields of medical treatment and the like, is used for enhancing the detected image, and is beneficial to the popularization of portable ultrasonic equipment.
The purpose of the invention is realized by the following technical means:
the invention discloses a mammary gland ultrasonic enhancement method based on deep learning, which is realized by eliminating the influence of a fat layer and a muscle layer in a mammary gland ultrasonic video and improving the quality of the mammary gland ultrasonic video of a portable instrument. Firstly, a mammary gland ultrasonic video with low image quality acquired by a portable instrument is intercepted into a mammary gland ultrasonic image, and a gland area in the mammary gland ultrasonic video is segmented by UNet. UNet can more effectively use limited marked samples through a data expansion technology, end-to-end training can be carried out by using very little data, and the network speed is very high; the glandular region of the fuzzy breast ultrasound image is then enhanced using a cycle-generated confrontation network (CycleGan). The most important characteristic of the CycleGan is that training data are not required to be paired, and the mapping of images among different domains can be successfully trained only by providing images of different domains, so that the image quality of a gland region in a mammary gland ultrasonic video of the portable equipment can be improved, the imaging level of high-end equipment is achieved, and the detection accuracy of the portable ultrasonic equipment is further improved.
A breast ultrasound video lesion detection method based on deep learning comprises the following steps:
the method comprises the following steps: and (4) segmenting the gland region in the portable instrument mammary gland ultrasonic video.
Step 1.1: a data set is constructed. Firstly, intercepting a mammary gland ultrasonic video collected from a portable instrument frame by frame to obtain a mammary gland ultrasonic image of the portable instrument; secondly, according to the labeling information of the breast ultrasound image of the portable instrument provided by the doctor (the labeling information gives coordinate data of three tissues of fat, gland and muscle on the whole breast ultrasound image), a segmentation map of the breast ultrasound image of the portable instrument is obtained by using MATLAB, and on the segmentation map, the pixel value of a background area is 0, the pixel value of a fat area is 1, the pixel value of a gland area is 2, and the pixel value of a muscle area is 3.
Step 1.2: UNet for segmentation was constructed. Firstly, constructing a coding structure, using a convolution kernel size of 3 multiplied by 3, wherein the number of features extracted by a first layer of convolution network is 16, the number of input image channels is 3, after two continuous convolutions, the number of the feature channels passes through a rectification linear unit and a maximum pooling layer for down sampling, the step length of the maximum pooling layer is 2, the number of the feature channels is doubled in each down sampling step, and the network structure of the convolutional layer, the convolutional layer and the pooling layer is repeated for 5 times; then, a decoding structure is constructed, an up-sampling method is adopted, the size of a convolution kernel is 2 x 2, and then the two 3x3 convolution layers are passed, and each layer is followed by a rectification linear unit; finally, the feature vectors are mapped to the corresponding classes using the 1 × 1 convolutional layer.
Step 1.3: and segmenting the portable instrument mammary gland ultrasonic image by using the trained UNet, and cutting a rectangular boundary frame of the gland region according to the segmentation result to obtain a portable instrument gland layer image.
Step two: and (4) segmenting a gland region in the mammary gland ultrasonic video of the high-end instrument.
Step 2.1: a data set is constructed. Firstly, intercepting a breast ultrasound video acquired from a high-end instrument frame by frame to obtain a breast ultrasound image of the high-end instrument; secondly, according to the labeling information of the breast ultrasound image of the high-end instrument provided by the doctor (the labeling information gives coordinate data of three tissues of fat, gland and muscle on the whole breast ultrasound image), a segmentation map of the breast ultrasound image of the high-end instrument is obtained by using MATLAB, wherein on the segmentation map, the pixel value of a background region is 0, the pixel value of a fat region is 1, the pixel value of a gland region is 2, and the pixel value of a muscle region is 3.
Step 2.2: UNet for segmentation was constructed. Firstly, constructing a coding structure, using a convolution kernel size of 3 multiplied by 3, wherein the number of features extracted by a first layer of convolution network is 16, the number of input image channels is 3, after two continuous convolutions, the number of the feature channels passes through a rectification linear unit and a maximum pooling layer for down sampling, the step length of the maximum pooling layer is 2, the number of the feature channels is doubled in each down sampling step, and the network structure of the convolutional layer, the convolutional layer and the pooling layer is repeated for 5 times; then, a decoding structure is constructed, an up-sampling method is adopted, the size of a convolution kernel is 2 x 2, and then the two 3x3 convolution layers are passed, and each layer is followed by a rectification linear unit; finally, the feature vectors are mapped to the corresponding classes using the 1 × 1 convolutional layer.
Step 2.3: and segmenting the breast ultrasonic image of the high-end instrument by using the trained UNet, and cutting a rectangular boundary frame of a gland region according to a segmentation result to obtain a gland layer image of the high-end instrument.
Step three: enhancing the fuzzy gland layer.
Step 3.1: a data set is constructed. The data of the step is divided into two types: one is the portable instrument glandular layer image obtained in the step one; the other is the high-end instrument glandular layer image obtained in step two. The two types of glandular layer images are adjusted to be uniform in size. Because the gland layer definition of the high-end instrument is higher than that of the portable instrument, the gland layer image of the portable instrument is enhanced by using the gland layer image of the high-end instrument.
Step 3.2: a cycle generation countermeasure network (CycleGan) was constructed for augmentation. Firstly, constructing a generator G to realize the migration from distribution X to distribution Y; secondly, constructing a generator F to realize the migration from the distribution Y to the distribution X; then, a construction discriminator Dx judges whether the data is true X or the data which is generated by the generator F according to Y and has the same distribution with X; then, a construction discriminator Dy judges whether the data is true Y or the data which is generated by the generator G according to the X and is distributed with the Y; then, constructing a loss function of the network, wherein the loss function comprises a countermeasure loss function, a cycle consistency loss function and an Identity mapping loss; and finally, training the circularly generated countermeasure network, updating parameters and storing.
Step 3.3: and (5) enhancing the blurred portable instrument glandular layer by using the trained generator to obtain a final enhancement result.
Advantageous effects
1. The method trains UNet, segments the gland region in the mammary gland ultrasonic video, and reduces the false positive rate caused by the interference of a fat layer and a muscle layer.
2. The method trains a cyclic generation confrontation network model, and the blurred portable instrument glandular layer image can be directly generated into an enhanced image after being input, so that the image quality and the detection result accuracy rate during target detection are improved.
Drawings
FIG. 1 is a schematic diagram of the UNet structure of the present invention;
FIG. 2 is a schematic diagram of a loop generation countermeasure network structure according to the present invention;
FIG. 3 is a schematic flow chart of a deep learning-based breast ultrasound enhancement method and an embodiment of the invention;
FIG. 4 is a portable breast ultrasound image;
FIG. 5 is a portable instrument glandular layer image;
FIG. 6 is a breast ultrasound image of a high-end instrument;
FIG. 7 is an image of the glandular layer of a high-end instrument;
FIG. 8 is an enhanced portable instrument glandular layer image;
FIG. 9 is an enhanced portable instrument glandular layer image with detected lesion areas.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and examples. The technical problems and the advantages solved by the technical solutions of the present invention are also described, and it should be noted that the described embodiments are only intended to facilitate the understanding of the present invention, and do not have any limiting effect.
The portable instrument is convenient for people to obtain own breast ultrasonic images by hands, but due to the handheld instability, the quality of the breast ultrasonic video of the portable instrument is often inferior to that of the breast ultrasonic video of a high-end instrument, and the problem of high omission ratio can be caused.
In this embodiment, the practical application of the method in reducing the false positive rate and the false negative rate of the target detection result of the breast ultrasound video lesion area of the portable instrument is illustrated by automatically detecting tumors in 142 portable instrument breast ultrasound videos provided by a hospital.
Fig. 3 is a flowchart of a breast ultrasound enhancement method in an embodiment of the present invention, which specifically includes the following steps:
step 1: 142 portable instrument breast ultrasound videos provided by hospitals were selected and the gland regions in the videos were segmented.
Step 1.1: adopting 142 portable instrument mammary gland ultrasonic video screenshots provided by a hospital, generating a corresponding segmentation graph according to a label provided by a doctor by using the portable instrument mammary gland ultrasonic video screenshots and the corresponding segmentation graph as a training set of UNet.
Step 1.2: constructing a coding structure, using a convolution kernel size of 3 multiplied by 3, wherein the number of the features extracted by a first layer of convolution network is 16, the number of input image channels is 3, after two continuous convolutions, the coding structure passes through a rectification linear unit and a maximum pooling layer for down sampling, the step length of the maximum pooling layer is 2, the number of the feature channels is doubled in each down sampling step, and the network structure of the convolution layer, the convolution layer and the pooling layer is repeated for 5 times; then, a decoding structure is constructed, an up-sampling method is adopted, the size of a convolution kernel is 2 x 2, and then the two 3x3 convolution layers are passed, and each layer is followed by a rectification linear unit; finally, the feature vectors are mapped to the corresponding classes using the 1 × 1 convolutional layer.
Step 1.3: inputting the breast ultrasound image of the portable instrument to be segmented into the trained UNet, outputting the segmentation result, and cutting to obtain the glandular layer image of the portable instrument, wherein fig. 5 is the glandular layer image of the portable instrument.
Step 2: 142 high-end instrument breast ultrasound videos provided by hospitals are selected, and gland areas in the videos are segmented.
Step 2.1: adopting 142 high-end instrument mammary gland ultrasonic video screenshots provided by a hospital, generating a corresponding segmentation graph according to doctor labeling by using a portable instrument mammary gland ultrasonic image shown in figure 6, and taking the mammary gland ultrasonic video screenshots and the corresponding segmentation graph as a training set of UNet.
Step 2.2: constructing a coding structure, using a convolution kernel size of 3 multiplied by 3, wherein the number of the features extracted by a first layer of convolution network is 16, the number of input image channels is 3, after two continuous convolutions, the coding structure passes through a rectification linear unit and a maximum pooling layer for down sampling, the step length of the maximum pooling layer is 2, the number of the feature channels is doubled in each down sampling step, and the network structure of the convolution layer, the convolution layer and the pooling layer is repeated for 5 times; then, a decoding structure is constructed, an up-sampling method is adopted, the size of a convolution kernel is 2 x 2, and then the two 3x3 convolution layers are passed, and each layer is followed by a rectification linear unit; finally, the feature vectors are mapped to the corresponding classes using the 1 × 1 convolutional layer.
Step 2.3: inputting the breast ultrasound image of the high-end instrument to be segmented into the trained UNet, outputting the segmentation result, and cutting to obtain the glandular layer image of the high-end instrument, wherein fig. 7 is the glandular layer image of the high-end instrument.
And step 3: enhancing the blurred breast ultrasound image.
Step 3.1: and (3) generating a training set of the countermeasure network by taking the 6906 portable instrument glandular layer images obtained in the step 1 and the 6906 high-end instrument glandular layer images obtained in the step 2 as a cycle.
Step 3.2: constructing a cyclic generation countermeasure network (CycleGan) for augmentation by first constructing a generator G to effect a migration from a distribution X, which is a high-end instrument breast ultrasound image, to a distribution Y, which is a portable instrument breast ultrasound image; secondly, constructing a generator F to realize the migration from the distribution Y to the distribution X; then, a construction discriminator Dx judges whether the data is true X or the data which is generated by the generator F according to Y and has the same distribution with X; then, a construction discriminator Dy judges whether the data is true Y or the data which is generated by the generator G according to the X and is distributed with the Y; then, constructing a loss function of the network, wherein the loss function comprises a countermeasure loss function, a cycle consistency loss function and an Identity mapping loss; and finally, training the circularly generated countermeasure network, updating parameters and storing.
Step 3.3: inputting the portable instrument gland layer image to be enhanced into a trained cycle generation countermeasure network (CycleGan), and outputting an enhancement result close to the quality of the high-end instrument gland layer, wherein the enhancement result excludes the interference of the fat layer and the muscle layer in the original breast ultrasound image, and the quality of the portable instrument gland layer is improved to the quality of the high-end instrument gland layer, and fig. 8 is the enhanced portable instrument gland layer image. It should be noted that the lesion automatic detection accuracy of the breast ultrasound image of the high-end instrument is generally better than that of the breast ultrasound image of the portable instrument.
The test result shows that the method can effectively realize the detection of the lesion area in the breast ultrasound video, and fig. 9 is an enhanced portable instrument glandular layer image in which the lesion area is detected. On 142 portable instrument mammary gland ultrasonic videos provided by doctors, the video missing rate obtained by the traditional lesion detection method is 25%, and the false positive rate is 7.02%; the video missing rate obtained by the method is 14.2%, and the false positive rate is 1.75%. The method really overcomes the defects of high false positive rate and high omission factor of the traditional lesion detection method.
The above detailed description is intended to illustrate the objects, aspects and advantages of the present invention, and it should be understood that the above detailed description is only exemplary of the present invention and is not intended to limit the scope of the present invention, and any modifications, equivalents, improvements and the like within the spirit and principle of the present invention should be included in the scope of the present invention.
Claims (4)
1. A mammary gland ultrasonic enhancement method based on deep learning is characterized in that: the method is realized by eliminating the influence of a fat layer and a muscle layer in the mammary gland ultrasonic video and improving the quality of the mammary gland ultrasonic video of a portable instrument, and specifically comprises the following steps:
the method comprises the following steps: dividing a gland region in the breast ultrasonic video of the portable instrument;
step two: dividing a gland region in a mammary gland ultrasonic video of a high-end instrument;
step three: enhancing the fuzzy gland layer.
2. The deep learning-based breast ultrasound enhancement method of claim 1, wherein: the implementation method of the first step comprises the following steps:
step 1.1: constructing a data set; firstly, intercepting a mammary gland ultrasonic video collected from a portable instrument frame by frame to obtain a mammary gland ultrasonic image of the portable instrument; secondly, obtaining a segmentation map of the portable instrument mammary gland ultrasonic image by using MATLAB according to labeling information of the portable instrument mammary gland ultrasonic image provided by a doctor, wherein on the segmentation map, the pixel value of a background region is 0, the pixel value of a fat region is 1, the pixel value of a gland region is 2, and the pixel value of a muscle region is 3;
step 1.2: constructing UNet for segmentation; firstly, constructing a coding structure, using a convolution kernel size of 3 multiplied by 3, wherein the number of features extracted by a first layer of convolution network is 16, the number of input image channels is 3, after two continuous convolutions, the number of the feature channels passes through a rectification linear unit and a maximum pooling layer for down sampling, the step length of the maximum pooling layer is 2, the number of the feature channels is doubled in each down sampling step, and the network structure of the convolutional layer, the convolutional layer and the pooling layer is repeated for 5 times; then, a decoding structure is constructed, an up-sampling method is adopted, the size of a convolution kernel is 2 x 2, and then the two 3x3 convolution layers are passed, and each layer is followed by a rectification linear unit; finally, mapping the feature vectors to corresponding classes using a 1 × 1 convolutional layer;
step 1.3: and segmenting the portable instrument mammary gland ultrasonic image by using the trained UNet, and cutting a rectangular boundary frame of the gland region according to the segmentation result to obtain a portable instrument gland layer image.
3. The deep learning-based breast ultrasound enhancement method of claim 1, wherein: the implementation method of the second step is as follows:
step 2.1: constructing a data set; firstly, intercepting a breast ultrasound video acquired from a high-end instrument frame by frame to obtain a breast ultrasound image of the high-end instrument; secondly, according to the labeling information of the breast ultrasound image of the high-end instrument provided by the doctor, obtaining a segmentation map of the breast ultrasound image of the high-end instrument by using MATLAB, wherein on the segmentation map, the pixel value of a background region is 0, the pixel value of a fat region is 1, the pixel value of a gland region is 2, and the pixel value of a muscle region is 3;
step 2.2: constructing UNet for segmentation; firstly, constructing a coding structure, using a convolution kernel size of 3 multiplied by 3, wherein the number of features extracted by a first layer of convolution network is 16, the number of input image channels is 3, after two continuous convolutions, the number of the feature channels passes through a rectification linear unit and a maximum pooling layer for down sampling, the step length of the maximum pooling layer is 2, the number of the feature channels is doubled in each down sampling step, and the network structure of the convolutional layer, the convolutional layer and the pooling layer is repeated for 5 times; then, a decoding structure is constructed, an up-sampling method is adopted, the size of a convolution kernel is 2 x 2, and then the two 3x3 convolution layers are passed, and each layer is followed by a rectification linear unit; finally, mapping the feature vectors to corresponding classes using a 1 × 1 convolutional layer;
step 2.3: and segmenting the breast ultrasonic image of the high-end instrument by using the trained UNet, and cutting a rectangular boundary frame of a gland region according to a segmentation result to obtain a gland layer image of the high-end instrument.
4. The deep learning-based breast ultrasound enhancement method of claim 1, wherein: the third step is realized by the following steps:
step 3.1: constructing a data set; the data of the step is divided into two types: one is the portable instrument glandular layer image obtained in the step one; the other type is the high-end instrument glandular layer image obtained in the step two; adjusting the two types of glandular layer images to be in a uniform size; because the gland layer definition of the high-end instrument is higher than that of the portable instrument, the gland layer image of the portable instrument is enhanced by using the gland layer image of the high-end instrument;
step 3.2: constructing a cycle generation countermeasure network for reinforcement; firstly, constructing a generator G to realize the migration from distribution X to distribution Y; secondly, constructing a generator F to realize the migration from the distribution Y to the distribution X; then, a construction discriminator Dx judges whether the data is true X or the data which is generated by the generator F according to Y and has the same distribution with X; then, a construction discriminator Dy judges whether the data is true Y or the data which is generated by the generator G according to the X and is distributed with the Y; then, constructing a loss function of the network, wherein the loss function comprises a countermeasure loss function, a cycle consistency loss function and an Identity mapping loss; finally, training the circularly generated countermeasure network, updating and storing the parameters;
step 3.3: and (5) enhancing the blurred portable instrument glandular layer by using the trained generator to obtain a final enhancement result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110967233.9A CN114140332A (en) | 2021-08-23 | 2021-08-23 | Deep learning-based breast ultrasound enhancement method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110967233.9A CN114140332A (en) | 2021-08-23 | 2021-08-23 | Deep learning-based breast ultrasound enhancement method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114140332A true CN114140332A (en) | 2022-03-04 |
Family
ID=80393613
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110967233.9A Pending CN114140332A (en) | 2021-08-23 | 2021-08-23 | Deep learning-based breast ultrasound enhancement method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114140332A (en) |
-
2021
- 2021-08-23 CN CN202110967233.9A patent/CN114140332A/en active Pending
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109493328B (en) | Medical image display method, viewing device and computer device | |
CN107833231B (en) | Medical image display method, apparatus and computer storage medium | |
WO2020164468A1 (en) | Medical image segmentation method, image segmentation method, related device and system | |
CN112086197B (en) | Breast nodule detection method and system based on ultrasonic medicine | |
JP2004033749A (en) | Semiautomatic segmental algorithm of pet (positron emission tomograph) tumor image | |
US10867375B2 (en) | Forecasting images for image processing | |
CN111127430A (en) | Method and device for determining medical image display parameters | |
CN111091127A (en) | Image detection method, network model training method and related device | |
US11823384B2 (en) | CT image generation method for attenuation correction of pet images | |
CN106530236B (en) | Medical image processing method and system | |
EP3971830B1 (en) | Pneumonia sign segmentation method and apparatus, medium and electronic device | |
CN110211143A (en) | A kind of medical image analysis method based on computer vision | |
CN109146993A (en) | A kind of Method of Medical Image Fusion and system | |
CN108269272B (en) | Liver's CT method for registering and system | |
JP2004174241A (en) | Image forming method | |
CN111210431B (en) | Blood vessel segmentation method, device, equipment and storage medium | |
CN115471470A (en) | Esophageal cancer CT image segmentation method | |
Senthil et al. | Enhancement Sushisen algorithms in images analysis technologies to increase computerized tomography images | |
CN112200780B (en) | Bone tissue positioning method, device, computer equipment and storage medium | |
CN106557767A (en) | A kind of method of ROI region in determination interventional imaging | |
CN112967254A (en) | Lung disease identification and detection method based on chest CT image | |
CN114140332A (en) | Deep learning-based breast ultrasound enhancement method | |
JP7265805B2 (en) | Image analysis method, image analysis device, image analysis system, control program, recording medium | |
CN110648333B (en) | Real-time segmentation system of mammary gland ultrasonic video image based on middle-intelligence theory | |
KR20220095401A (en) | Apparatus and method for diagnosing facial fracture based on artificial intelligence |
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 |