CN114119635B - Fatty liver CT image segmentation method based on cavity convolution - Google Patents
Fatty liver CT image segmentation method based on cavity convolution Download PDFInfo
- Publication number
- CN114119635B CN114119635B CN202111391577.6A CN202111391577A CN114119635B CN 114119635 B CN114119635 B CN 114119635B CN 202111391577 A CN202111391577 A CN 202111391577A CN 114119635 B CN114119635 B CN 114119635B
- Authority
- CN
- China
- Prior art keywords
- convolution
- parameter
- cavity
- neural network
- pooling
- 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
- 238000000034 method Methods 0.000 title claims abstract description 25
- 208000004930 Fatty Liver Diseases 0.000 title claims abstract description 19
- 206010019708 Hepatic steatosis Diseases 0.000 title claims abstract description 19
- 208000010706 fatty liver disease Diseases 0.000 title claims abstract description 19
- 238000003709 image segmentation Methods 0.000 title claims abstract description 19
- 231100000240 steatosis hepatitis Toxicity 0.000 title claims abstract description 19
- 238000011176 pooling Methods 0.000 claims abstract description 37
- 230000004927 fusion Effects 0.000 claims abstract description 6
- 239000013598 vector Substances 0.000 claims description 33
- 238000012549 training Methods 0.000 claims description 27
- 238000013527 convolutional neural network Methods 0.000 claims description 26
- 238000010586 diagram Methods 0.000 claims description 14
- 238000013528 artificial neural network Methods 0.000 claims description 9
- 238000004364 calculation method Methods 0.000 claims description 6
- 239000011159 matrix material Substances 0.000 claims description 6
- 238000012360 testing method Methods 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 3
- 238000007500 overflow downdraw method Methods 0.000 claims description 3
- 238000012546 transfer Methods 0.000 claims description 3
- 210000004185 liver Anatomy 0.000 abstract description 16
- 239000010410 layer Substances 0.000 description 22
- 238000002591 computed tomography Methods 0.000 description 17
- 230000006870 function Effects 0.000 description 16
- 230000011218 segmentation Effects 0.000 description 12
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 239000011229 interlayer Substances 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000002059 diagnostic imaging Methods 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000035558 fertility Effects 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 210000000056 organ Anatomy 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
Classifications
-
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
- 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
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30056—Liver; Hepatic
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a fatty liver CT image segmentation method based on hole convolution, which comprises the steps of hole convolution, multi-language fusion, pooling and other operations of hole convolution, fully integrates liver multi-scale characteristics, adds initial attributes of different decoding layers in the decoding process, and performs corresponding image pooling treatment, thereby improving the accuracy of image segmentation.
Description
Technical Field
The invention relates to a fatty liver CT image segmentation method based on cavity convolution.
Background
The liver is a high-incidence area of human disease. Automatic liver segmentation may help doctors diagnose and treat patients early. In the field of medical imaging, computed Tomography (CT) is one of the most widely used medical examinations.
In recent years, with the rapid development of deep learning, and in particular, the application of revolutionary neural networks (CNNs) in computer vision tasks, such as image segmentation Long et al, have proposed a complete convolutional network (FCN) based on CNNs that can accept any size of image input. The recovered image is very rough. Through segmentation and fusion of single-channel pixels, thicker feature layers can be obtained, and additional semantic information is provided for subsequent segmentation.
For two-dimensional segmentation of liver, zhang et al propose three-dimensional inverse segmentation of a conditional network, and adopt a three-dimensional automatic segmentation framework from coarse to fine, introduce a spatial pyramid pool and a fertility mechanics learning technology into the u-net, and propose a liver CT automatic segmentation method based on sar-u-net. The 3du network can effectively utilize the segmentation features between adjacent liver segments to obtain better segmentation results.
Considering that 3D u-net utilizes information between disks in CT images, the 3D CNN of Lu et al simultaneously detects and segments the liver, lei et al designed a three-dimensional narrow back block and middle pool block suitable for LV network. By decoupling the transverse correction and the spatial correlation of the liver CT image and combining the extended convolution hybrid technology, the interlayer characteristics of the liver CT image are extracted, and the liver can be effectively segmented through the three-dimensional folding and the interlayer disk extracted by the background extraction.
Due to the characteristics of CT and u-net images, liver segmentation has the following problems:
(1) The low and high attributes are equally important for liver segmentation; the bottom-up function in u-net ignores the importance of the bottom-up feature, resulting in low network segmentation performance.
(2) The shape and size of the liver are variable, and its gray scale value is similar to that of the adjacent organs, so that information is easily lost.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide the novel solar cell.
The aim of the invention is realized by the following technical scheme:
the fatty liver CT image segmentation method based on the hole convolution comprises the step of training a hole convolution neural network, wherein the step of training the hole convolution neural network comprises the following steps of:
s1, building a cavity convolutional neural network, wherein the cavity convolutional neural network comprises a plurality of convolutional layers, a plurality of cavity convolutional layers and a plurality of deconvolution layers which are sequentially connected;
S2, initializing a weight value of the cavity convolutional neural network;
S3, inputting the fatty liver CT image in the pre-established image segmentation training camp into the cavity convolutional neural network initialized in the step S2, and obtaining the weight of the advanced semantic information through a Sigmoid function; then weighting the high-level features by using the detail information in the adjacent low-level features, and weighting the low-level features by using the semantic information in the adjacent high-level features; finally, the addition is performed to enhance the feature transfer, let F l∈RH×W×C denote F l that the encoder respectively represents the height, width and channel number of the feature in the H, W, C th, and the low-level feature channel attention vector V c and the spatial attention vector V s obtained by using this module are:
S4, inputting the fatty liver CT image in the pre-established picture test set into the cavity convolutional neural network obtained through training in the step S3, and outputting a corresponding picture depth map and a picture surface normal vector map; the image segmentation training camp is at least partially different from the fatty liver CT images in the picture test set; the image surface normal vector diagram is obtained by obtaining point cloud data of pixels through the depth diagram and adopting least square plane fitting;
S5, judging whether the prediction precision of the hole convolutional neural network, which is obtained through training in the step S3, on the picture depth and the picture surface normal vector accords with preset requirements according to the picture depth map and the picture surface normal vector map which are output in the step S4: if yes, finishing training; if not, returning to the step S3 to continue training until the training is performed to obtain the cavity convolutional neural network meeting the preset requirement.
Further, the step S3 further includes a fusion method of the neural network and the multi-scale semantic feature module: the high-level feature channel attention vector Z c and the feature space attention vector Z s obtained by using the multi-scale semantic feature module are:
Where l ε [1,4], fg (·) represents global average pooling, conv (·) represents convolution operations, Φ (·) represents Singmoid activation operations, vector is computed And/>Element multiplication is carried out on the low-level output characteristics and F l respectively, so that weighted low-level output characteristics L l are obtained; vector/>And/>Element multiplication is carried out on the weighted high-level output characteristics H l,Ll,Ll∈RH×W×C by F l respectively; when l=1, vector/>And/>The output result obtained by the input multi-scale semantic feature module can be expressed as:
Wherein Concat denotes a feature channel fusion operation.
Further, the step S4 further includes a method for encoding-decoding the hole convolution image, where the hole convolution image encodes the lost data in the hole and stores the encoded data together with the convolution kernel, and when the stored data needs to be output, the data is expanded by decoding.
Further, the step S5 further includes a pooling operation for the pictures, which specifically includes the following steps:
the method comprises the steps of respectively carrying out 1×1 convolution on features obtained through parallel cavity convolution to strengthen the extracted multi-scale features, fusing the multi-scale features with a feature map obtained after average pooling, carrying out 1×1 convolution on fused output to obtain output F' with fixed size, wherein the specific steps are as follows:
F″=σ(f1×1[f1×1(PA(F′));f1×1(f3×3,d=6(F′));f1×1(f3×3,d=12(F′));
f1×1(f3×3,d=18(F′));PA(F′)]
Wherein PA represents average pooling and d represents expansion rate;
let the feature map X of the afferent parameter pooling layer be W in width, H in height, and C in channel number, i.e. W H C matrix, then its output G can be according to the formula
X c is the C-th channel of the input feature map, and G c represents the input corresponding to the C-th channel; w is a weight parameter matrix, whose shape is p×q, W c represents the C-th channel of the parameter pooling layer weight, p, q is the length and width of the pooling core, β is a preactivation function, whose function is to convert the correlation operation into an interpretable pooling operation, any one value of the weight parameter is mapped to the real interval (0, 1) using the sigmoid function as the preactivation function, the parameter pooling layer assigns a parameter W c to each channel X c of the feature map, each parameter value being specified between 0 and 1;
After the weight parameters are added, the pooling calculation method is based on The method comprises the following steps:
Wherein, the subscript C represents the C-th channel of the calculation feature map, beta is the preactivation function, w is the weight parameter and the weight parameter needs to be updated, the gradient of the layer parameter to the loss function C needs to be calculated, the first layer is a parameter pooling layer, and the gradient to C is that
According toObtaining
Optimizing a parametric formula for the layer input using a falling gradient is/>
The beneficial effects of the invention are as follows: the multi-scale characteristics of the liver are fully integrated, the initial attributes of different decoding layers and corresponding image pooling processing are added in the decoding process, and the accuracy of image segmentation is improved.
Drawings
FIG. 1 is a schematic diagram of a flow architecture of the present invention;
FIG. 2 is a schematic diagram of a three-dimensional 3D visualization model of the hole convolution of the present invention;
FIG. 3 is a schematic diagram of a residual convolution block of the present invention;
Fig. 4 is a schematic diagram of a hole residual convolution block.
Detailed Description
The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments, and it is apparent that the described embodiments are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by a person skilled in the art without any inventive effort, are intended to be within the scope of the present invention, based on the embodiments of the present invention.
Referring to fig. 1-4, the present invention provides a technical solution: the fatty liver CT image segmentation method based on the cavity convolution comprises the step of training a cavity convolution neural network, wherein the step of training the cavity convolution neural network comprises the following steps of:
s1, constructing a cavity convolutional neural network, wherein the cavity convolutional neural network comprises a plurality of convolutional layers, a plurality of cavity convolutional layers and a plurality of deconvolution layers which are sequentially connected;
S2, initializing a weight value of the cavity convolutional neural network;
S3, inputting the fatty liver CT image in the pre-established image segmentation training camp into the cavity convolutional neural network initialized in the step S2, and obtaining the weight of the advanced semantic information through a Sigmoid function; then weighting the high-level features by using the detail information in the adjacent low-level features, and weighting the low-level features by using the semantic information in the adjacent high-level features; finally, the addition is performed to enhance the feature transfer, let F l∈RH×W×C denote F l that the encoder respectively represents the height, width and channel number of the feature in the H, W, C th, and the low-level feature channel attention vector V c and the spatial attention vector V s obtained by using this module are:
S4, inputting the fatty liver CT image in the pre-established picture test set into the cavity convolutional neural network obtained through training in the step S3, and outputting a corresponding picture depth map and a picture surface normal vector map; the fatty liver CT images in the image segmentation training set and the image test set are at least partially different; the normal vector diagram of the picture surface is obtained by obtaining point cloud data of pixels through a depth diagram and adopting least square plane fitting;
S5, judging whether the prediction precision of the hole convolutional neural network, which is obtained through training in the step S3, on the picture depth and the picture surface normal vector accords with preset requirements according to the picture depth map and the picture surface normal vector map which are output in the step S4: if yes, finishing training; if not, returning to the step S3 to continue training until the training is performed to obtain the cavity convolutional neural network meeting the preset requirement.
As shown in fig. 2, the present invention proposes a technique based on hole convolution, in which each pixel is enlarged while the image pixel is enlarged, but different liver shapes and sizes depend on the patient, so that the variation of the data amount is very important, especially in the distribution image. Top-down examples are typically used to obtain a better receptive field, but the spatial resolution is greatly reduced, while the spatial resolution is better controlled by the hole convolution method, so that the problem of local distortion is basically solved. Compared with the common convolution, the cavity convolution adds a cavity into the convolution kernel so as to achieve the purpose of expanding the receptive field on the premise of ensuring that the size of an output characteristic diagram is not changed, the receptive field is an indispensable ring in the convolution neural network, the image mapping function after convolution is firmly controlled, a point on the characteristic diagram corresponds to a region on an input diagram, further, the cavity residual convolution used by us consists of two 3d cavity convolutions with the cavity ratios of 1 and 2 and residual errors after Conv1 is 1*1, and the following diagram is adopted, so that the model can fully perceive the obtained liver multi-scale structure under the condition of not increasing the depth and the complexity, namely not increasing the workload, and the foundation is laid for the subsequent multi-scale characteristic fusion, and the steps are shown in fig. 3 and 4.
Further, step S3 further includes a fusion method of the neural network and the multi-scale semantic feature module: the high-level feature channel attention vector Z c and the feature space attention vector Z s obtained by using the multi-scale semantic feature module are:
where l ε [1,4], fg (·) represents global average pooling, conv (·) represents convolution operations, Φ (·) represents Sigmoid activation operations, and vectors are computed And/>Element multiplication is carried out on the low-level output characteristics and F l respectively, so that weighted low-level output characteristics L l are obtained; vector/>And/>Element multiplication is carried out on the weighted high-level output characteristics H l,Ll,Ll∈RH×W×C by F l respectively; when l=1, vector/>And
The output result obtained by the input multi-scale semantic feature module can be expressed as:
Wherein Concat denotes a feature channel fusion operation.
Further, step S4 further includes a method for encoding-decoding the hole convolution image, where the hole convolution image encodes the lost data in the hole and stores the encoded data together with the convolution kernel, and when the stored data needs to be output, the data is expanded by decoding.
Further, step S5 further includes a pooling operation for the pictures, which specifically includes the following steps:
the method comprises the steps of respectively carrying out 1×1 convolution on features obtained through parallel cavity convolution to strengthen the extracted multi-scale features, fusing the multi-scale features with a feature map obtained after average pooling, carrying out 1×1 convolution on fused output to obtain output F' with fixed size, wherein the specific steps are as follows:
F″=σ(f1×1[f1×1(PA(F′));f1×1(f3×3,d=6(F′));f1×1(f3×3,d=12(F′));
f1×1(f3×3,d=18(F′))PA(F′)]
Wherein PA represents average pooling and d represents expansion rate;
let the feature map X of the afferent parameter pooling layer be W in width, H in height, and C in channel number, i.e. W H C matrix, then its output G can be according to the formula
X c is the C-th channel of the input feature map, and G c represents the input corresponding to the C-th channel; w is a weight parameter matrix, whose shape is p×q, W c represents the C-th channel of the parameter pooling layer weight, p, q is the length and width of the pooling core, β is a preactivation function, whose function is to convert the correlation operation into an interpretable pooling operation, any one value of the weight parameter is mapped to the real interval (0, 1) using the sigmoid function as the preactivation function, the parameter pooling layer assigns a parameter W c to each channel X c of the feature map, each parameter value being specified between 0 and 1;
After the weight parameters are added, the pooling calculation method is based on The method comprises the following steps:
Wherein, the subscript C represents the C-th channel of the calculation feature map, beta is the preactivation function, w is the weight parameter and the weight parameter needs to be updated, the gradient of the layer parameter to the loss function C needs to be calculated, the first layer is a parameter pooling layer, and the gradient to C is that
According toObtaining
Optimizing a parametric formula for the layer input using a falling gradient is/>
The foregoing is merely a preferred embodiment of the invention, and it is to be understood that the invention is not limited to the form disclosed herein and is not to be considered as excluding other embodiments, but is capable of numerous other combinations, modifications and environments and is capable of modifications within the scope of the inventive concept, either by the foregoing teachings or by the teaching of the relevant art. And that modifications and variations which do not depart from the spirit and scope of the invention are intended to be within the scope of the appended claims.
Claims (4)
1. A fatty liver CT image segmentation method based on cavity convolution is characterized by comprising the following steps of: the method comprises the step of training the hole convolutional neural network, wherein the step of training the hole convolutional neural network comprises the following steps of:
s1, building a cavity convolutional neural network, wherein the cavity convolutional neural network comprises a plurality of convolutional layers, a plurality of cavity convolutional layers and a plurality of deconvolution layers which are sequentially connected;
S2, initializing a weight value of the cavity convolutional neural network;
S3, inputting the fatty liver CT image in the pre-established image segmentation training camp into the cavity convolutional neural network initialized in the step S2, and obtaining the weight of the advanced semantic information through a Sigmoid function; then weighting the high-level features by using the detail information in the adjacent low-level features, and weighting the low-level features by using the semantic information in the high-level features; finally, adding is performed to enhance the transfer of the features, let F l∈RH×W×C, where F l represents the processed added features, where H, W, C represents the height, width and channel number of the features, respectively, and the low-level feature channel attention vector V c and the spatial attention vector V s obtained using the processed added features are:
S4, inputting the fatty liver CT image in the pre-established picture test set into the cavity convolutional neural network obtained through training in the step S3, and outputting a corresponding picture depth map and a picture surface normal vector map; the image segmentation training camp is at least partially different from the fatty liver CT images in the picture test set; the image surface normal vector diagram is obtained by obtaining point cloud data of pixels through the depth diagram and adopting least square plane fitting;
S5, judging whether the prediction precision of the hole convolutional neural network, which is obtained through training in the step S3, on the picture depth and the picture surface normal vector accords with preset requirements according to the picture depth map and the picture surface normal vector map which are output in the step S4: if yes, finishing training; if not, returning to the step S3 to continue training until the training is performed to obtain the cavity convolutional neural network meeting the preset requirement.
2. The fatty liver CT image segmentation method based on hole convolution as set forth in claim 1, wherein: the step S3 further comprises a fusion method of the neural network and the multi-scale semantic feature module: the high-level feature channel attention vector Z c and the feature space attention vector Z s obtained by using the multi-scale semantic feature module are:
Where l ε [1,4], f g (·) represents global average pooling, conv (·) represents convolution operations, Φ (·) represents Sigmoid activation operations, and the vector is computed And/>Element multiplication is carried out on the low-level output characteristics and F l respectively, so that weighted low-level output characteristics L l are obtained; vector/>And/>Element multiplication is carried out on the weighted high-level output characteristics H l,Ll,Ll∈RH×W×C by F l respectively; when l=1, vector/>And/>The output result obtained by the input multi-scale semantic feature module can be expressed as:
Wherein Concat denotes a feature channel fusion operation.
3. The fatty liver CT image segmentation method based on hole convolution as set forth in claim 1, wherein: the step S4 also comprises a coding-decoding method of the cavity convolution image, wherein the lost data in the cavity is coded by the cavity convolution image and then stored together with the convolution kernel, and when the stored data needs to be output, the data is unfolded through decoding.
4. The fatty liver CT image segmentation method based on hole convolution as set forth in claim 1, wherein: the step S5 further comprises the pooling operation of the pictures, and the specific steps are as follows:
The method comprises the steps of respectively carrying out 1X 1 convolution on features obtained through parallel cavity convolution to strengthen the extracted multi-scale features, fusing the multi-scale features with a feature map obtained after average pooling, carrying out 1X 1 convolution on fused output to obtain output F' with fixed size, wherein the specific steps are as follows:
F″=σ(f1×1[f1×1(PA(F′));f1×1(f3×3,d=6(F′));f1×1(f3×3,d=12(F′));
f1×1(f3×3,d=18(F′));PA(F′)]
Wherein P A represents average pooling and d represents expansion rate;
let the feature map X of the afferent parameter pooling layer be W in width, H in height, and C in channel number, i.e. W H C matrix, then its output G can be according to the formula
X i is the C-th channel of the input feature map, and G c represents the input corresponding to the C-th channel; w is a weight parameter matrix, whose shape is p×q, W c represents the C-th channel of the parameter pooling layer weight, p, q is the length and width of the pooling core, β is a preactivation function, whose function is to convert the correlation operation into an interpretable pooling operation, any one value of the weight parameter is mapped to the real interval (0, 1) using the sigmoid function as the preactivation function, the parameter pooling layer assigns a parameter W c to each channel X c of the feature map, each parameter value being specified between 0 and 1;
After the weight parameters are added, the pooling calculation method is based on The method comprises the following steps:
Wherein, the subscript C represents the C-th channel of the calculation feature map, beta is the preactivation function, w is the weight parameter and the weight parameter needs to be updated, the gradient of the parameter pooling layer parameter to the loss function C needs to be calculated, the 1 st layer is the parameter pooling layer, and the gradient to C is that
According toObtaining
Optimizing a parametric formula for the layer input using a falling gradient is/>
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111391577.6A CN114119635B (en) | 2021-11-23 | 2021-11-23 | Fatty liver CT image segmentation method based on cavity convolution |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111391577.6A CN114119635B (en) | 2021-11-23 | 2021-11-23 | Fatty liver CT image segmentation method based on cavity convolution |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114119635A CN114119635A (en) | 2022-03-01 |
CN114119635B true CN114119635B (en) | 2024-05-24 |
Family
ID=80439689
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111391577.6A Active CN114119635B (en) | 2021-11-23 | 2021-11-23 | Fatty liver CT image segmentation method based on cavity convolution |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114119635B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114758137B (en) * | 2022-06-15 | 2022-11-01 | 深圳瀚维智能医疗科技有限公司 | Ultrasonic image segmentation method and device and computer readable storage medium |
CN117058380B (en) * | 2023-08-15 | 2024-03-26 | 北京学图灵教育科技有限公司 | Multi-scale lightweight three-dimensional point cloud segmentation method and device based on self-attention |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110910405A (en) * | 2019-11-20 | 2020-03-24 | 湖南师范大学 | Brain tumor segmentation method and system based on multi-scale cavity convolutional neural network |
CN112330681A (en) * | 2020-11-06 | 2021-02-05 | 北京工业大学 | Attention mechanism-based lightweight network real-time semantic segmentation method |
CN112927255A (en) * | 2021-02-22 | 2021-06-08 | 武汉科技大学 | Three-dimensional liver image semantic segmentation method based on context attention strategy |
WO2021155230A1 (en) * | 2020-01-31 | 2021-08-05 | James R. Glidewell Dental Ceramics, Inc. | Teeth segmentation using neural networks |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11775836B2 (en) * | 2019-05-21 | 2023-10-03 | Magic Leap, Inc. | Hand pose estimation |
-
2021
- 2021-11-23 CN CN202111391577.6A patent/CN114119635B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110910405A (en) * | 2019-11-20 | 2020-03-24 | 湖南师范大学 | Brain tumor segmentation method and system based on multi-scale cavity convolutional neural network |
WO2021155230A1 (en) * | 2020-01-31 | 2021-08-05 | James R. Glidewell Dental Ceramics, Inc. | Teeth segmentation using neural networks |
CN112330681A (en) * | 2020-11-06 | 2021-02-05 | 北京工业大学 | Attention mechanism-based lightweight network real-time semantic segmentation method |
CN112927255A (en) * | 2021-02-22 | 2021-06-08 | 武汉科技大学 | Three-dimensional liver image semantic segmentation method based on context attention strategy |
Non-Patent Citations (3)
Title |
---|
Study on MRI Medical Image Segmentation Technology Based on CNN-CRF Model;Naiqin Feng et al.;《IEEE Access》;20200323;第08卷;第60505 - 60514页 * |
一种改进DeeplabV3网络的烟雾分割算法;汪梓艺;苏育挺;刘艳艳;张为;;西安电子科技大学学报;20190904;第46卷(第06期);第52-59页 * |
基于全卷积网络的CT图像肝脏肿瘤分割方法研究;马瑞;《中国优秀硕士学位论文全文数据库(电子期刊)医药卫生科技辑》;20210515;全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN114119635A (en) | 2022-03-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110443842B (en) | Depth map prediction method based on visual angle fusion | |
CN107492071A (en) | Medical image processing method and equipment | |
Zhang et al. | Deep hierarchical guidance and regularization learning for end-to-end depth estimation | |
CN113674253B (en) | Automatic segmentation method for rectal cancer CT image based on U-transducer | |
CN114119635B (en) | Fatty liver CT image segmentation method based on cavity convolution | |
CN111091589A (en) | Ultrasonic and nuclear magnetic image registration method and device based on multi-scale supervised learning | |
CN111932550B (en) | 3D ventricle nuclear magnetic resonance video segmentation system based on deep learning | |
CN115482241A (en) | Cross-modal double-branch complementary fusion image segmentation method and device | |
CN109377520A (en) | Cardiac image registration arrangement and method based on semi-supervised circulation GAN | |
CN110363068B (en) | High-resolution pedestrian image generation method based on multiscale circulation generation type countermeasure network | |
CN110363802B (en) | Prostate image registration system and method based on automatic segmentation and pelvis alignment | |
CN112990077B (en) | Face action unit identification method and device based on joint learning and optical flow estimation | |
CN111369565A (en) | Digital pathological image segmentation and classification method based on graph convolution network | |
CN113222124B (en) | SAUNet + + network for image semantic segmentation and image semantic segmentation method | |
CN111080591A (en) | Medical image segmentation method based on combination of coding and decoding structure and residual error module | |
CN113570658A (en) | Monocular video depth estimation method based on depth convolutional network | |
CN116664397B (en) | TransSR-Net structured image super-resolution reconstruction method | |
CN113177592B (en) | Image segmentation method and device, computer equipment and storage medium | |
CN116486074A (en) | Medical image segmentation method based on local and global context information coding | |
CN117710760B (en) | Method for detecting chest X-ray focus by using residual noted neural network | |
CN113269774B (en) | Parkinson disease classification and lesion region labeling method of MRI (magnetic resonance imaging) image | |
Lu et al. | Parallel region-based deep residual networks for face hallucination | |
CN112489048B (en) | Automatic optic nerve segmentation method based on depth network | |
CN111401209B (en) | Action recognition method based on deep learning | |
CN111209946B (en) | Three-dimensional image processing method, image processing model training method and medium |
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 | ||
GR01 | Patent grant |