CN109949276B - Lymph node detection method for improving SegNet segmentation network - Google Patents
Lymph node detection method for improving SegNet segmentation network Download PDFInfo
- Publication number
- CN109949276B CN109949276B CN201910152279.8A CN201910152279A CN109949276B CN 109949276 B CN109949276 B CN 109949276B CN 201910152279 A CN201910152279 A CN 201910152279A CN 109949276 B CN109949276 B CN 109949276B
- Authority
- CN
- China
- Prior art keywords
- lymph node
- layer
- image
- segmentation
- segnet
- 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
Abstract
The invention discloses a lymph node detection method based on an improved SegNet segmentation network, which comprises the following steps: dividing a lymph node image data set into a training set and a testing set; constructing a SegNet segmentation network based on cavity convolution operation; training the SegNet segmentation network by using a training set, minimizing a sine-cosine cross entropy loss function into a network optimization objective function, and optimizing the SegNet segmentation network; and identifying and segmenting lymph nodes in the lymph image to be identified by using the trained SegNet segmentation network. The invention uses the hole convolution to extract the characteristics, increases the receptive field area of the hollow convolution under the condition of not increasing extra calculation amount, avoids the loss of down-sampling information and solves the problem of the reduction of the resolution of the sampled image. The sample with small prediction error is weighted less than the cross entropy loss function through the sine and cosine cross entropy loss function, so that the problem of imbalance of the training positive and negative samples is solved. And performing post-processing on the segmentation result through a Markov random field to further refine the edge part of the segmentation object.
Description
Technical Field
The invention belongs to the technical field of image segmentation, and particularly relates to a lymph node detection method based on an improved SegNet segmentation network.
Background
Traditional medical image segmentation techniques can be divided into three major categories: (1) the method is based on the segmentation of the image completely, and all information required by the segmentation comes from the image; (2) object model based methods that incorporate a priori information about the object to be segmented (e.g., shape information of the object). Common are graph (Atlas) -based segmentation methods; (3) the hybrid method firstly uses information based on images to carry out preliminary segmentation, and then further segments the target based on prior constraint information. A semantic segmentation network based on a deep learning technology belongs to a first type of segmentation technology, namely, information required by segmentation is extracted from an image. Because the features do not need to be extracted from the image manually and the segmentation result is good, the method becomes a research hotspot in the field of current medical image segmentation.
Currently, there are many different segmentation networks based on deep learning, such as FCN, SegNet, deep Lab-V3, etc., which have been studied extensively in natural image segmentation. SegNet is a semantic segmentation network based on convolution operation, and consists of two parts, namely an encoding unit and a decoding unit. However, SegNet-based lymph node segmentation methods have the following drawbacks: the problem of the number imbalance of the positive and negative samples in the training network is not fully considered; the segmentation of the object from a multi-scale, multi-resolution perspective is not possible.
Disclosure of Invention
Aiming at the defects of the prior art, the invention aims to solve the technical problems that the prior art does not fully consider the unbalanced number of positive and negative samples in a training network and cannot segment a target from the perspective of multi-scale and multi-resolution.
In order to achieve the above object, in a first aspect, an embodiment of the present invention provides a lymph node detection method based on an improved SegNet segmentation network, including the following steps:
s1, dividing a lymph node image data set into a training set and a testing set;
s2, constructing a segNet segmentation network based on cavity convolution operation;
s3, training the SegNet segmentation network by using a training set, and optimizing the SegNet segmentation network by minimizing a sine-cosine cross entropy loss function into a network optimization objective function;
and S4, identifying and segmenting the lymph nodes in the lymph image to be identified by using the trained SegNet segmentation network.
Specifically, the SegNet segmentation network based on the hole convolution operation comprises a feature extraction module and a lymph node segmentation module;
the characteristic extraction module is used for extracting a multilayer characteristic diagram of the lymph node, the input is a lymph node image, and the output is a series of characteristic diagrams;
the lymph node segmentation module is used for identifying and segmenting lymph nodes layer by layer according to the features extracted by the feature extraction module, inputting a series of feature maps extracted by the feature extraction module, and outputting a binary image with thresholded probability of each pixel point, wherein the size of the binary image is the same as that of an original lymph node image.
Specifically, the feature extraction module is composed of 5 coding units connected in series, and each coding unit is composed of a 2-by-one void convolution layer, a 1-by-one normalization layer, a 1-by-one linear correction unit layer and a 1-by-one down-sampling layer connected in series; the lymph node segmentation module is formed by connecting 5 decoding units and 1 Softmax layer in series, and each decoding unit is formed by connecting 1 sampling layer, 2 void convolution layer, 1 group normalization layer and 1 linear correction unit layer in series.
Specifically, the void convolution layer is used for extracting lymph node image features, and a void convolution operation formula is as follows:
wherein x is an input lymph node image, w is a cavity convolution kernel, r is an expansion rate of the cavity convolution, k is a moving range of the center of the convolution kernel, i is a position of a pixel point, and y is an output characteristic diagram.
Specifically, the group normalization operation formula is as follows:
wherein x is(k)The value of the input feature map for the kth hole convolution layer, E the expectation of all feature values in the input feature map, Var (-) the variance of all input feature values,
the characteristic value is normalized.
Specifically, the linear correction unit operation formula is as follows:
and performing linear correction on the input characteristic value, wherein when the input characteristic value is greater than 0, the output characteristic value is unchanged, and when the input characteristic value is less than 0, the output characteristic value is 0.
Specifically, the operational formula of the Softmax layer in the lymph node segmentation module is as follows:
wherein z isiFor linearly correcting the value of the pixel at position i in the output feature map of the unit layer, Softmax (z)i) Is a pixel point z on the original lymph node imageiIs the probability of a lymph node.
Specifically, the sine and cosine cross entropy loss function formula is expressed as follows:
Lcs＝-γ(cos(p)-αsin(p))log(p)
wherein, gamma and alpha are both hyper-parameters, and p is the prediction probability of the segmentation network to the pixel points of the trained image.
Specifically, the lymph node detection method further comprises the step S5 of performing post-processing on the lymph node segmentation result by adopting a Markov random field.
In a second aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the lymph node detecting method according to the first aspect is implemented.
Generally, compared with the prior art, the above technical solution conceived by the present invention has the following beneficial effects:
1. compared with the general convolution operation, the cavity convolution obtains a plurality of characteristic graphs under different scales by using the cavity convolution to extract the characteristics in the SegNet network and by adopting the cavity convolution operation with different expansion rates, and compared with the general convolution operation, the cavity convolution avoids the loss of information under the condition of under-sampling by increasing the receptive field area of the cavity convolution under the condition of not increasing extra calculated amount and parameters, thereby improving the identification rate and the segmentation accuracy of lymph nodes and solving the problem of the reduction of the resolution of sampled images caused by the pooling operation.
2. The invention designs a sine and cosine cross entropy loss function, uses the loss function as a measurement function for measuring network errors, and adds a sine and cosine factor in front of the cross entropy loss function, thereby realizing that a sample with small prediction error is given a weight smaller than the cross entropy loss function, and a sample with large prediction error is given a weight larger than the cross entropy loss function, and solving the problem of training the imbalance of positive and negative samples.
3. According to the invention, the lymph nodes obtained by SegNet segmentation are subjected to post-processing by a Markov random field method, so that the edge part of the segmented object is further refined, and the sensitivity and the specificity of lymph node segmentation can be improved.
4. The lymph node segmentation method established by the invention can be used for automatically detecting and segmenting lymph nodes in lymph node CT images, and the segmentation result can be used for clinical reference and medical auxiliary diagnosis, so that the working efficiency of clinical diagnosis of doctors can be greatly improved.
Drawings
Fig. 1 is a flowchart of a lymph node detection method based on an improved SegNet segmentation network according to an embodiment of the present invention;
fig. 2 is a schematic diagram of a SegNet segmentation network structure based on a hole convolution operation according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
To facilitate understanding of the present invention, the terms involved are first explained:
conditional random field: a discriminant probability model represents a Markov random field of a set of output random variables (segmentation classes) given the set of input random variables (pixel points of an image input), i.e., a conditional random field assumes that the output random variables constitute the Markov random field.
As shown in fig. 1, a lymph node detection method based on an improved SegNet segmentation network includes the following steps:
s1, dividing a lymph node image data set into a training set and a testing set;
s2, constructing a segNet segmentation network based on cavity convolution operation;
s3, training the SegNet segmentation network by using a training set, and optimizing the SegNet segmentation network by minimizing a sine-cosine cross entropy loss function into a network optimization objective function;
and S4, identifying and segmenting the lymph nodes in the lymph image to be identified by using the trained SegNet segmentation network.
And S1, dividing a lymph node image data set into a training set and a testing set.
The lymph node image dataset includes N human medical images, each image being H x H1, where H x H represents the size of the height and width of each input image, typically 512 x 512, and 1 is the number of channels. Lymph node regions were manually marked with the assistance of a professional clinical radiologist, and saved as N sets of binary images (contour maps) as gold standards for a training set and a test set. The N individuals may include normal individuals (normal lymph nodes) and may also include lymph node cancer patients (diseased lymph nodes). There are a plurality of CT images with lymph nodes marked on each person in different numbers, and the lymph nodes are not limited to a specific part and may include a chest lymph node, a neck lymph node, and the like. The lymph node image data set obtained by labeling can be used for establishing a CT lymph node image library.
The lymph node image data set is randomly divided into a training set and a testing set, the training set has CT image data of M persons, and the rest N-M personal data form the testing set.
And S2, constructing a segNet segmentation network based on the hole convolution operation.
As shown in fig. 2, the SegNet segmentation network based on the hole convolution operation includes a feature extraction module and a lymph node segmentation module.
The characteristic extraction module is used for extracting a multi-layer characteristic diagram of the lymph node, the input is a CT lymph node image, and the output is a series of characteristic diagrams containing the lymph node. It consists of 5 coding units, each of which consists of 2-by-one void convolution layer (in series), 1-by-one normalization layer, 1-by-one linear correction unit layer and 1-by-one sampling layer, the output of the former is the input of the latter. Input to each layer is a constantly learned feature of the lymph node from abstract to concrete. The experimental result shows that the characteristics of lymph nodes can be enriched by the convolution operation of the two groups of cavities, and the performance of lymph node detection and segmentation is improved. The down-sampling layer is used to halve the feature map size.
The use of hole convolution for SegNet networks is proposed to address the problem of reduced resolution of sampled images caused by pooling operations (downsampling layers). The cavity convolution layer is used for extracting lymph node image features, and a cavity convolution operation formula is as follows:
wherein x is an input lymph node image, w is a cavity convolution kernel, r is an expansion rate of the cavity convolution, k is a moving range of the center of the convolution kernel, i is a position of a pixel point, and y is an output characteristic diagram. Setting is performed according to the size of the core, if k is 3, the summation range is 3, and these are parameters provided when setting the network. After the image is input to the hole convolution layer, the hole convolution operation is performed four times in parallel for each layer according to the set expansion rate of the hole convolution, and then the obtained different feature maps are input to the group normalization layer. The expansion ratio r here is chosen to be 1,2,4,8 for parallel operation.
And inputting the feature graph output by the void convolution layer into a group normalization layer, and performing normalization processing on the features of each layer by the group normalization layer so as to better control the range of data. The input is a feature map output by the hole convolution layer, and the number of the feature maps depends on the number of the set hole convolution kernels. The group normalization operation formula is as follows:
wherein x is(k)The value of the input feature map for the kth hole convolution layer, E the expectation of all feature values in the input feature map, Var (-) the variance of all input feature values,
the characteristic value is normalized.
The output characteristic diagram obtained after the group normalization is input into a linear correction unit layer, which is also commonly called an activation function layer, and is used for performing linear correction on the input characteristic value, wherein the input value smaller than zero is zero, and the input value larger than zero is kept unchanged. The linear correction unit operation formula is as follows:
and performing linear correction on the input characteristic value, wherein when the input characteristic value is greater than 0, the output characteristic value is unchanged, and when the input characteristic value is less than 0, the output characteristic value is 0.
And the lymph node segmentation module is used for identifying and segmenting lymph nodes layer by layer according to the characteristics extracted by the characteristic extraction module. The input is the feature of the relevant lymph node extracted by the feature extraction module (encoding unit), and the final output is a binary image of the lymph node segmentation result having the same size as the input image. The device consists of 5 decoding units, wherein each decoding unit consists of an upsampling layer, a hole convolution layer, a group normalization layer and a linear correction unit layer. The upsampling layer restores the feature map to the original size.
And inputting the characteristic diagram predicted for each type after the upsampling and convolution operation, the group normalization operation and the linear correction unit operation into the Softmax layer. The number of feature maps corresponds to the number of classes that need to be segmented. The operation of the Softmax layer is carried out so that the value of each output predicted point is between 0 and 1, and the probability that the point is a certain class label is represented. Here, the number of classes is 2, i.e., lymph nodes and tissue organs (non-lymph nodes). The operation is as follows:
wherein z isiFor linearly correcting the value of a pixel point at position i in the output characteristic diagram of the unit layer, zi＝f(xi)。
And S3, training the SegNet segmentation network by using a training set, and optimizing the SegNet segmentation network by using the minimization of sine-cosine cross entropy loss function as a network optimization objective function.
Configuring network training parameters: and (3) selecting a random gradient descent method, and initializing learning rate, momentum parameters, regularization coefficients, maximum times of training and minimum picture quantity input in each training. Specifically, the present embodiment sets the initial learning rate to 0.001; the regularization coefficient was 0.0005; the maximum number of training is 50; the image data read at one time is 4; and optimally training the model by using a random gradient descent method. In order to increase the training data volume, make the network more stable and improve the performance of network segmentation, the X-direction and Y-direction translation range from-10 to 10 pixels using a data expansion strategy.
The method aims to solve the problem of training the imbalance of the positive and negative samples, realize the calculation of network errors and reduce the errors through a reverse error propagation algorithm. The output error is calculated using a sine and cosine cross entropy loss function and the performance of the model is improved by a back propagation algorithm (BP algorithm). The sine and cosine cross entropy loss layer is composed of two main functions: and finally, calculating the average error and the average gradient of each point input after the loss error and the error gradient obtained by calculation of the derivatives of the sine-cosine cross entropy focusing function and the sine-cosine cross entropy focusing function are averaged. The sine and cosine cross entropy loss function formula is as follows:
Lcs＝-γ(cos(p)-αsin(p))log(p)
wherein, gamma and alpha are both hyper-parameters, and p is the prediction probability of the segmentation network to the pixel points of the trained image. The hyper-parameter gamma is generally selected to be greater than 1; to ensure that the output is positive, the hyperparameter a is typically greater than 0.64 and is used to calculate the error loss between the predicted value and the true tag value obtained from the network segmentation. The optimization objective is to minimize the loss function.
And S4, identifying and segmenting the lymph nodes in the lymph image to be identified by using the trained SegNet segmentation network.
And S5, performing post-processing on the lymph node segmentation result by adopting a Markov random field.
And taking the output result of the segmentation module as an area selected by a seed point of the conditional random field, and then realizing further refinement treatment on the edge part of the segmented object by establishing a Markov chain random field.
The image post-processing of the conditional random field comprises the following steps:
(1) using 8-neighborhood Gibbs sampling for the binary image after SegNet segmentation, and calculating the prior probability of lymph node pixel points in the sampled binary image;
U(x)＝∑c∈CVc(xc)
wherein, x is a class label of a binary image after SegNet segmentation, Z is a normalization constant, beta is a positive constant and is used for controlling the size of the cluster, U (x) is a Gibbs potential function, C is 8 neighborhoods of selected pixel points, and x iscIs the pixel value of the point c in the neighborhood, VCIs a step function.
(2) Calculating a likelihood function whether each pixel point in the original CT image belongs to a lymph node;
wherein x is a category label of the binary image after SegNet segmentation; y denotes the pixel value, μ, of the original CT imagesIs the mean, σ, of the image in the 8 neighborhoodssIs the variance in the 8 neighborhood.
(3) An original CT image and a segNet segmented binary image are input, the segNet segmented binary image is used as an initially selected segmentation region, and the original CT image is segmented again by using an iterative condition mode algorithm (ICM algorithm).
The post-processing is optimized based on the last a posteriori probability, where the likelihood function is represented by a gaussian distribution and the prior class of each point in the image is represented by a gibbs distribution. The maximum a posteriori probability can be represented by the following sub-formula:
the latter two terms represent the logarithm of the likelihood function and the logarithm of the prior probability, respectively. The ICM algorithm may obtain an approximate solution to the maximum a posteriori probability estimate, with an optimized objective function as follows:
and L is a set of all pixel points in the binary image after SegNet segmentation.
The optimization function obtains the maximum posterior probability of whether each pixel point in the image belongs to the lymph node in an iteration mode. The segmentation result can be used for clinical reference, and the working efficiency of clinical diagnosis of doctors can be greatly improved.
The above description is only for the preferred embodiment of the present application, but the scope of the present application is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present application should be covered within the scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.
Claims (8)
1. A lymph node detection method based on an improved SegNet segmentation network, characterized by comprising the following steps:
s1, dividing a lymph node image data set into a training set and a testing set;
s2, constructing a segNet segmentation network based on cavity convolution operation;
s3, training the SegNet segmentation network by using a training set, and optimizing the SegNet segmentation network by minimizing a sine-cosine cross entropy loss function into a network optimization objective function;
s4, using the trained SegNet segmentation network to identify and segment the lymph nodes in the lymph image to be identified,
the SegNet segmentation network based on the cavity convolution operation comprises a feature extraction module and a lymph node segmentation module;
the feature extraction module is formed by connecting 5 coding units in series, and each coding unit is formed by connecting 2-by-one hole convolution layers, 1-by-one group normalization layers, 1-by-one linear correction unit layers and 1-by-one down-sampling layers in series; the lymph node segmentation module is formed by connecting 5 decoding units and 1 Softmax layer in series, and each decoding unit is formed by connecting 1 × up-sampling layer, 2 × hollow convolution layer, 1 × group normalization layer and 1 × linear correction unit layer in series;
the sine and cosine cross entropy loss function formula is as follows:
Lcs＝-γ(cos(p)-αsin(p))log(p)
wherein, gamma and alpha are both hyper-parameters, and p is the prediction probability of the segmentation network to the pixel points of the trained image.
2. The lymph node detecting method according to claim 1, wherein the feature extracting module is configured to extract a multi-layer feature map of the lymph node, the input is a lymph node image, and the output is a series of feature maps;
the lymph node segmentation module is used for identifying and segmenting lymph nodes layer by layer according to the features extracted by the feature extraction module, inputting a series of feature maps extracted by the feature extraction module, and outputting a binary image with thresholded probability of each pixel point, wherein the size of the binary image is the same as that of an original lymph node image.
3. The lymph node detecting method according to claim 1, wherein the void convolution layer is used for extracting a lymph node image feature, and a void convolution operation formula is as follows:
wherein x is an input lymph node image, w is a cavity convolution kernel, r is an expansion rate of the cavity convolution, k is a moving range of the center of the convolution kernel, i is a position of a pixel point, and y is an output characteristic diagram.
4. The lymph node detection method of claim 1 wherein the group normalization operation is formulated as follows:
wherein x is(k)The values of the input feature map for the kth hole convolution layer, E the expectation of all feature values in the input feature map,
for the variance of all the input feature values,
the characteristic value is normalized.
5. The lymph node detecting method according to claim 1, wherein the linear correction unit operation formula is as follows:
and performing linear correction on the input characteristic value, wherein when the input characteristic value is greater than 0, the output characteristic value is unchanged, and when the input characteristic value is less than 0, the output characteristic value is 0.
6. The lymph node detection method of claim 1, wherein the operation formula of the Softmax layer in the lymph node segmentation module is as follows:
wherein z isiFor linearly correcting the value of the pixel at position i in the output feature map of the unit layer, Softmax (z)i) Is a pixel point z on the original lymph node imageiIs the probability of a lymph node.
7. The lymph node detecting method according to claim 1, wherein the lymph node detecting method further comprises a step s5 of post-processing the lymph node segmentation result using a markov random field.
8. A computer-readable storage medium, characterized in that the computer-readable storage medium has stored thereon a computer program which, when being executed by a processor, carries out the lymph node detection method according to any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910152279.8A CN109949276B (en) | 2019-02-28 | 2019-02-28 | Lymph node detection method for improving SegNet segmentation network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910152279.8A CN109949276B (en) | 2019-02-28 | 2019-02-28 | Lymph node detection method for improving SegNet segmentation network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109949276A CN109949276A (en) | 2019-06-28 |
| CN109949276B true CN109949276B (en) | 2021-06-11 |
Family
ID=67006541
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910152279.8A Active CN109949276B (en) | 2019-02-28 | 2019-02-28 | Lymph node detection method for improving SegNet segmentation network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109949276B (en) |
Families Citing this family (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110599452B (en) * | 2019-08-07 | 2022-02-22 | 全球能源互联网研究院有限公司 | Rust detection system, method, computer device and readable storage medium |
| CN110751175A (en) * | 2019-09-12 | 2020-02-04 | 上海联影智能医疗科技有限公司 | Method and device for optimizing loss function, computer equipment and storage medium |
| CN110717907A (en) * | 2019-10-06 | 2020-01-21 | 浙江大学 | Intelligent hand tumor detection method based on deep learning |
| CN111105424A (en) * | 2019-12-19 | 2020-05-05 | 广州柏视医疗科技有限公司 | Lymph node automatic delineation method and device |
| CN111340825B (en) * | 2020-03-05 | 2023-05-09 | 上海市肺科医院（上海市职业病防治院） | Method and system for generating mediastinum lymph node segmentation model |
| CN111428785B (en) * | 2020-03-23 | 2023-04-07 | 厦门大学 | Puffer individual identification method based on deep learning |
| CN111582111B (en) * | 2020-04-29 | 2022-04-29 | 电子科技大学 | Cell component segmentation method based on semantic segmentation |
| CN112686896B (en) * | 2021-03-12 | 2021-07-06 | 苏州鼎纳自动化技术有限公司 | Glass defect detection method based on frequency domain and space combination of segmentation network |
| CN113592885B (en) * | 2021-06-29 | 2024-03-12 | 中南大学 | SegNet-RS network-based large obstacle contour segmentation method |
| CN113808140B (en) * | 2021-09-24 | 2024-04-09 | 昆明理工大学 | Aluminum-silicon alloy microscopic image segmentation method for gap region perception |
| CN117197166B (en) * | 2023-11-06 | 2024-02-06 | 中南大学 | Polyp image segmentation method and imaging method based on edge and neighborhood information |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107657257A (en) * | 2017-08-14 | 2018-02-02 | 中国矿业大学 | A kind of semantic image dividing method based on multichannel convolutive neutral net |
| CN108062756A (en) * | 2018-01-29 | 2018-05-22 | 重庆理工大学 | Image, semantic dividing method based on the full convolutional network of depth and condition random field |
| CN108268870A (en) * | 2018-01-29 | 2018-07-10 | 重庆理工大学 | Multi-scale feature fusion ultrasonoscopy semantic segmentation method based on confrontation study |
| CN109035252A (en) * | 2018-06-29 | 2018-12-18 | 山东财经大学 | A kind of super-pixel method towards medical image segmentation |
| CN109191471A (en) * | 2018-08-28 | 2019-01-11 | 杭州电子科技大学 | Based on the pancreatic cell image partition method for improving U-Net network |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106845357B (en) * | 2016-12-26 | 2019-11-05 | 银江股份有限公司 | A kind of video human face detection and recognition methods based on multichannel network |
-
2019
- 2019-02-28 CN CN201910152279.8A patent/CN109949276B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107657257A (en) * | 2017-08-14 | 2018-02-02 | 中国矿业大学 | A kind of semantic image dividing method based on multichannel convolutive neutral net |
| CN108062756A (en) * | 2018-01-29 | 2018-05-22 | 重庆理工大学 | Image, semantic dividing method based on the full convolutional network of depth and condition random field |
| CN108268870A (en) * | 2018-01-29 | 2018-07-10 | 重庆理工大学 | Multi-scale feature fusion ultrasonoscopy semantic segmentation method based on confrontation study |
| CN109035252A (en) * | 2018-06-29 | 2018-12-18 | 山东财经大学 | A kind of super-pixel method towards medical image segmentation |
| CN109191471A (en) * | 2018-08-28 | 2019-01-11 | 杭州电子科技大学 | Based on the pancreatic cell image partition method for improving U-Net network |
Non-Patent Citations (5)
| Title |
|---|
| "DILATED CONVOLUTIONS";Fisher Yu 等;《arXiv:1511.07122v2》;20160304;正文第1-9页 * |
| "Focal Loss for Dense Object Detection";Tsung-Yi Lin 等;《arXiv:1708.02002v2》;20180207;全文 * |
| "Group Normalization";Yuxin Wu 等;《arXiv:1803.08494v3》;20180611;正文第1-10页 * |
| "语义分割中的深度学习方法全解：从FCN、SegNet到各版本DeepLab ";GADFLYGIS;《https://blog.csdn.net/notHeadache/article/details/82183503》;20180711;全文 * |
| 解决样本不均衡的问题-调整类权重 修改交叉熵loss;notHeadache;《https://blog.csdn.net/notHeadache/article/details/82183503》;20180830;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109949276A (en) | 2019-06-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109949276B (en) | Lymph node detection method for improving SegNet segmentation network | |
| US9968257B1 (en) | Volumetric quantification of cardiovascular structures from medical imaging | |
| US20230186476A1 (en) | Object detection and instance segmentation of 3d point clouds based on deep learning | |
| CN107610194B (en) | Magnetic resonance image super-resolution reconstruction method based on multi-scale fusion CNN | |
| CN108898160B (en) | Breast cancer histopathology grading method based on CNN and imaging omics feature fusion | |
| CN110969626B (en) | Method for extracting hippocampus of human brain nuclear magnetic resonance image based on 3D neural network | |
| US20230104173A1 (en) | Method and system for determining blood vessel information in an image | |
| CN110930397A (en) | Magnetic resonance image segmentation method and device, terminal equipment and storage medium | |
| CN111145181B (en) | Skeleton CT image three-dimensional segmentation method based on multi-view separation convolutional neural network | |
| CN107766874B (en) | Measuring method and measuring system for ultrasonic volume biological parameters | |
| CN109492668B (en) | MRI (magnetic resonance imaging) different-phase multimode image characterization method based on multi-channel convolutional neural network | |
| CN113450397B (en) | Image deformation registration method based on deep learning | |
| CN110717518A (en) | Persistent lung nodule identification method and device based on 3D convolutional neural network | |
| CN113362310A (en) | Medical image liver segmentation method based on unsupervised learning | |
| CN110570394A (en) | medical image segmentation method, device, equipment and storage medium | |
| CN114202545A (en) | UNet + + based low-grade glioma image segmentation method | |
| CN111932495B (en) | Medical image detection method, device and storage medium | |
| CN115375711A (en) | Image segmentation method of global context attention network based on multi-scale fusion | |
| CN111091575B (en) | Medical image segmentation method based on reinforcement learning method | |
| CN116563533A (en) | Medical image segmentation method and system based on target position priori information | |
| Liu et al. | AHU-MultiNet: adaptive loss balancing based on homoscedastic uncertainty in multi-task medical image segmentation network | |
| CN112396605B (en) | Network training method and device, image recognition method and electronic equipment | |
| CN113539402B (en) | Multi-mode image automatic sketching model migration method | |
| Tao et al. | Tooth CT Image Segmentation Method Based on the U-Net Network and Attention Module. | |
| CN113379770B (en) | Construction method of nasopharyngeal carcinoma MR image segmentation network, image segmentation method and device |
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 | ||
| GR01 | Patent grant |