CN111104834A - Application method of cross-contrast neural network in intelligent detection of heart sound - Google Patents
Application method of cross-contrast neural network in intelligent detection of heart sound Download PDFInfo
- Publication number
- CN111104834A CN111104834A CN201811281412.1A CN201811281412A CN111104834A CN 111104834 A CN111104834 A CN 111104834A CN 201811281412 A CN201811281412 A CN 201811281412A CN 111104834 A CN111104834 A CN 111104834A
- Authority
- CN
- China
- Prior art keywords
- neural network
- cross
- contrast
- image
- application method
- 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
- 238000000034 method Methods 0.000 title claims abstract description 24
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 19
- 238000001514 detection method Methods 0.000 title claims abstract description 9
- 238000001228 spectrum Methods 0.000 claims abstract description 5
- 230000005236 sound signal Effects 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 3
- 238000003062 neural network model Methods 0.000 claims 1
- 238000002474 experimental method Methods 0.000 abstract description 4
- 238000009826 distribution Methods 0.000 description 7
- 239000013598 vector Substances 0.000 description 7
- 238000012549 training Methods 0.000 description 6
- 230000004044 response Effects 0.000 description 5
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000004422 calculation algorithm Methods 0.000 description 2
- 208000019425 cirrhosis of liver Diseases 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 238000007619 statistical method Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000010191 image analysis Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000011524 similarity measure Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- 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
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
Abstract
The invention discloses an application method of a cross-contrast neural network in intelligent detection of heart sound. The invention combines the deep neural network and the IBS theory, and provides a new network architecture, namely a cross-contrast neural network. The new network architecture utilizes a convolution filter to extract the texture features of the sound frequency spectrum, and then calculates the similarity of the features of the two images in a comparison mode. Experiments on the phonocardiogram database show that the method can improve the classification accuracy and simultaneously give intuitive statistical explanation.
Description
Technical Field
The invention relates to application of statistical analysis and deep learning in intelligent detection of heart sounds.
Background
In 2012, the AlexNet network makes a historical breakthrough in ILSVRC, and the effect greatly surpasses that of the traditional image classification method, which is the first time that the deep neural network is used for large-scale image classification. After the AlexNet network, a series of convolutional neural network models emerged, constantly refreshing the performance on the ImageNet dataset. The deep learning and the neural network are rapidly developed in signal analysis, particularly image analysis, and the accuracy of analysis is greatly improved.
The application of neural networks to medical images presents two major problems. On the one hand, the neural network structure is very complex, and the phenomenon of overfitting is easy to occur on a small sample set, especially on a medical sample. On the other hand, the task of classifying medical images is not only concerned with improving accuracy, but also requires enhancing interpretability. Unlike many classification/segmentation tasks in the field of computer vision, medical computer-aided applications often require approval of clinical knowledge for large-scale use. Based on this, we propose a new network architecture, cross-contrast neural network (CCNN). The new network architecture utilizes a convolution filter to extract the texture features of the sound frequency spectrum, and then calculates the similarity of the features of the two images in a comparison mode. This method on the one hand promotes the use of the sample and on the other hand explains in part from a statistical point of view why we can judge the class of an image. Experiments on the phonocardiogram database show that the method can improve the classification accuracy and simultaneously give intuitive statistical explanation.
Disclosure of Invention
In view of the above disadvantages of the prior art methods, the present invention is directed to a new network architecture, cross-contrast neural network (CCNN). The new network architecture utilizes a convolution filter to extract the texture features of the sound frequency spectrum, and then calculates the similarity of the features of the two images in a comparison mode. Firstly, the heart sound signal is converted into an image and then input into a subsequent network for processing. There are many ways to convert the collected heart sound signals into images, and the scheme adopts a way of converting the heart sound into the images through short-time Fourier transform.
The network architecture has two parts:
the first part is the architectural part of the network, divided into a four-layer structure. Each layer is composed of a convolutional layer composed of several filters. In the second part, image features obtained by a k-th layer network are extracted for training, and the method comprises the following steps:
step 1, carrying out probability statistics on the extracted features;
step 2, obtaining the standardized MIBS value by using the modified IBS formula, combining the MIBS value with the label, and calculating
A loss value;
and 3, optimizing the loss value, and performing model training until the training is finished.
The invention has the beneficial effects that: the deep neural network is combined with the statistical analysis IBS theory, images are input into the network pairwise in a combined mode, the traditional IBS theory is modified and added into network judgment as priori knowledge, and the problems that the sample size is small, accuracy is difficult to guarantee and statistical characteristics cannot be searched are solved. This method on the one hand promotes the use of the sample and on the other hand explains in part from a statistical point of view why we can judge the class of an image. Experiments on the phonocardiogram database show that the method can improve the classification accuracy and simultaneously give intuitive statistical explanation.
Drawings
FIG. 1 is a network architecture diagram of the method of the present invention
FIG. 2 is a diagram of a convolutional neural network model
FIG. 3 is a flow chart of a model algorithm
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in detail with reference to the accompanying drawings and detailed description.
The image multi-task classification method based on the cross-contrast neural network is applied to the intelligent detection of the heart sounds. As shown in fig. 1, a network configuration of the present invention is configured such that a captured heart sound signal is first converted into an image, and the following short-time fourier transform method, for example, can be used. The sound signals are converted into normalized time-frequency graphs through short-time Fourier transform, and the images are subjected to feature extraction through a cross-contrast neural network. Response to different inputs. The specific formula is as follows:
(for example, parameters can be used such that fs is 2 kHZ, window size is 512, Hamming distance is 256)
As shown in fig. 2, the convolutional neural network structure is adopted, and it is assumed that we extract the convolution result of the ith layer, and the number of filters corresponding to this layer is t. Each filter can be seen as a pattern of some texture and the result of the convolution represents the response of the input image over the t texture patterns. In order to measure the response strength of the input image to the t filters, a threshold value is set, and the number of pixel points of which the pixel point values are larger than the threshold value on the t characteristic images is counted to serve as extracted texture characteristic values. And comparing the pixel value of each pixel point in the t images with the threshold, wherein the threshold is 1 when the pixel value is larger than or equal to the threshold, and otherwise, the threshold is 0. Thus, the number of 1 in each of the t images is obtained, and a series of probabilities w are obtainedi(i＝1，2，3，....，N)。
In at least one embodiment of the invention, the set is made up of a number of images of certain categories selected from a database, where the number of images of each category in the database may be slightly different but not biased too much, otherwise the trained network may be biased.
As shown in fig. 3, which is a flow chart of the algorithm, we use the image features output from any layer as the input of the subsequent IBS formula processing stage. For example, we can take the output of the i-th layer, and for a set of inputs (two time-frequency graphs of heart sounds), we set a threshold, and can reduce the feature map of the 4-th convolutional layer to 2 feature vectors of 1 × t. The magnitude of the feature vector represents the magnitude of the response of the input image to the t filters. Therefore, how to measure the similarity of the textures of the two images can be converted into measuring the distance between the two vectors. Since we do not care about the magnitude of the values, but only about whether the response strength of the two images to the t textures has obvious rank correlation, we do not choose a sensitive method of euclidean distance/manhattan distance and other values here. We used a modified IBS model to measure the energy required to carry the break point to the 45 degree line. We update the parameters of the entire network with the cross entropy between IBS values and the real label as a loss function.
Next we will explain in detail the principle of the M-IBS model of similarity measure. Assuming that the two images belong to different classes, they should have different texture distributions; belonging to the same class, they should have similar texture characteristics. This is statistically represented as a significant correlation between the 1 x t-dimensional vectors extracted by the filter, and we therefore use IBS to measure this correlation, given two sets of extracted 1 x t-dimensional texture feature vectors (e.g., taking the fourth layer results in the following feature vectors):
we define the calculation of the NorIBS value as follows:
wherein
wkDenotes the convolution result, p, of the 4-1 layer filter ki(wk) The ratio of the number of pixels representing the output of the ith (i ═ 1, 2) image in a set of inputs (two MR images) on the kth filter of 4-1 layers is greater than a given threshold. It can be seen that the corresponding kth filter represents the probability that the texture appears on the original. Ri(wk) Indicating that for image i, the output of the k-th filter is sorted in size among 512 values. F (w)k) Showing the return of the k-th filter on both imagesThe normalized entropy. N represents the number of filters, and in this embodiment, N is 512. If the time-frequency graphs of two heart sounds belong to the same class, the ordering values of the texture features should be close, and the NorIBS value tends to be 0. If the feature vectors of 2 x 512 dimensions are plotted as scatter plots, the points should be distributed around the 45 degree line. On the contrary, if the time-frequency graphs of the two heart sounds do not belong to the same class, the NorIBS value is greater than 0, and if the filters with larger entropies have ordering differences, the NorIBS value is increased more greatly. The point distribution reflected on the scatter diagram is very scattered in a two-dimensional space and can be traced without regularity. For two time-frequency graphs of heart sounds we have given labels 0 or 1 when feeding them into the network, with label 0 representing the group of inputs originating from the same liver fibrosis stage and label 1 representing from different liver fibrosis stages. Since we want to conduct the cross entropy between the true label and the IBS predictor back as loss to compute the gradient update network, and the upper limit of IBS is not strictly 1, we modify the original IBS model. The specific method comprises the following steps: after calculating NorIBS, we chose to randomly shuffle the probability values. We define the distance after scrambling as RevIBS:
RevIBS＝NorIBS(random(p1(wk))，ramdom(p2(wk))) (6)
α is a learning factor, which can be set to 2-4.
For images belonging to the same class, NorIBS is close to 0, and its disturbed RevIBS value is much larger than NorIBS, this difference is further amplified by exponential operation, and ModIBS tends to 0. For the MR images of different stages, the original distribution of the MR images shows a scattered distribution, the distribution before and after the scattering does not change greatly, and the difference between the original distribution and the scattered distribution tends to 0, and the ModIBS tends to 1. So far, we have converted ModIBS to values between 0 and 1. We use cross entropy as a loss function and use a batch gradient training method to modify the model parameters.
loss＝(1-label)·log(1-IBS)+label·log(IBS) (8)
For a test picture to be staged, selecting partial samples of each category in a training set and the picture to be tested to form an input pair, calculating ModIBS of the partial samples and the picture to be tested, and selecting the category of the training set sample with the highest similarity as a prediction result.
The main innovation of this new network architecture is:
1) the new network architecture utilizes a filter to extract the texture features of the sound frequency spectrum, and calculates the similarity of the features of the two images in a comparison mode.
2) Experiments show that the method can improve the classification accuracy and simultaneously give visual statistical explanation.
It should be understood that although the present description refers to embodiments, not every embodiment contains only a single technical solution, and such description is for clarity only, and those skilled in the art should make the description as a whole, and the technical solutions in the embodiments can also be combined appropriately to form other embodiments understood by those skilled in the art.
The above-listed detailed description is only a specific description of a possible embodiment of the present invention, and they are not intended to limit the scope of the present invention, and equivalent embodiments or modifications made without departing from the technical spirit of the present invention should be included in the scope of the present invention.
Claims (4)
1. The invention discloses a technology of an application method of a cross-contrast neural network in the intelligent detection of heart sound, which is characterized in that the cross-contrast neural network and an IBS formula are applied, trained and tested.
2. The technology of the application method of the cross-contrast neural network in the intelligent detection of the heart sound as claimed in claim 1, is characterized in that collected heart sound signals are converted into images, the images are input into the network pairwise in a combined mode, the traditional IBS theory is modified and added into network judgment as priori knowledge, and the problems that the sample size is small, the accuracy is difficult to guarantee and statistical characteristics cannot be found are solved.
3. The technique of the application method of the cross-contrast neural network in the intelligent detection of the heart sound as claimed in claim 1, wherein the cross-contrast neural network model trained by the image is utilized, the image features output from any layer can be used as the input of the subsequent IBS formula processing stage, the texture features of the sound frequency spectrum are extracted by the filter, and the similarity of the two image features is calculated by means of contrast.
4. The technique of the application method of the cross-contrast neural network in the intelligent detection of the cardioid sound as claimed in claim 1, wherein the classification of an image is determined by comparing a known classification image with each classification image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811281412.1A CN111104834A (en) | 2018-10-25 | 2018-10-25 | Application method of cross-contrast neural network in intelligent detection of heart sound |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811281412.1A CN111104834A (en) | 2018-10-25 | 2018-10-25 | Application method of cross-contrast neural network in intelligent detection of heart sound |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111104834A true CN111104834A (en) | 2020-05-05 |
Family
ID=70419372
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811281412.1A Pending CN111104834A (en) | 2018-10-25 | 2018-10-25 | Application method of cross-contrast neural network in intelligent detection of heart sound |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111104834A (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107122809A (en) * | 2017-04-24 | 2017-09-01 | 北京工业大学 | Neural network characteristics learning method based on image own coding |
| CN107693044A (en) * | 2017-11-15 | 2018-02-16 | 广东顺德西安交通大学研究院 | Surveillance of Coronary Heart diagnostic device |
| CN107945817A (en) * | 2017-11-15 | 2018-04-20 | 广东顺德西安交通大学研究院 | Heart and lung sounds signal sorting technique, detection method, device, medium and computer equipment |
-
2018
- 2018-10-25 CN CN201811281412.1A patent/CN111104834A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107122809A (en) * | 2017-04-24 | 2017-09-01 | 北京工业大学 | Neural network characteristics learning method based on image own coding |
| CN107693044A (en) * | 2017-11-15 | 2018-02-16 | 广东顺德西安交通大学研究院 | Surveillance of Coronary Heart diagnostic device |
| CN107945817A (en) * | 2017-11-15 | 2018-04-20 | 广东顺德西安交通大学研究院 | Heart and lung sounds signal sorting technique, detection method, device, medium and computer equipment |
Non-Patent Citations (1)
| Title |
|---|
| 张冰清: "医学影像在肝纤维化程度预判中的应用研究", 《中国优秀硕士学位论文全文数据库 (医药卫生科技辑)》 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112308158B (en) | Multi-source field self-adaptive model and method based on partial feature alignment | |
| CN108537743B (en) | Face image enhancement method based on generation countermeasure network | |
| CN107564025B (en) | Electric power equipment infrared image semantic segmentation method based on deep neural network | |
| CN109389171B (en) | Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology | |
| CN109002848B (en) | Weak and small target detection method based on feature mapping neural network | |
| CN111814704A (en) | Full convolution examination room target detection method based on cascade attention and point supervision mechanism | |
| CN108629370B (en) | Classification recognition algorithm and device based on deep belief network | |
| CN107169117B (en) | Hand-drawn human motion retrieval method based on automatic encoder and DTW | |
| CN108805151B (en) | Image classification method based on depth similarity network | |
| CN110322445B (en) | Semantic segmentation method based on maximum prediction and inter-label correlation loss function | |
| Guo et al. | JointPruning: Pruning networks along multiple dimensions for efficient point cloud processing | |
| CN108764019A (en) | A kind of Video Events detection method based on multi-source deep learning | |
| CN109871749B (en) | Pedestrian re-identification method and device based on deep hash and computer system | |
| CN108171119B (en) | SAR image change detection method based on residual error network | |
| Lin et al. | Hypergraph optimization for multi-structural geometric model fitting | |
| Shoohi et al. | DCGAN for Handling Imbalanced Malaria Dataset based on Over-Sampling Technique and using CNN. | |
| WO2020119624A1 (en) | Class-sensitive edge detection method based on deep learning | |
| CN110163130B (en) | Feature pre-alignment random forest classification system and method for gesture recognition | |
| Singh et al. | Wavelet based histogram of oriented gradients feature descriptors for classification of partially occluded objects | |
| CN112818774A (en) | Living body detection method and device | |
| Tiwari | A pattern classification based approach for blur classification | |
| Sarikabuta et al. | Impacts of layer sizes in deep residual-learning convolutional neural network on flower image classification with different class sizes | |
| CN109376619A (en) | A kind of cell detection method | |
| He et al. | Beyond ${\chi^ 2} $ Difference: Learning Optimal Metric for Boundary Detection | |
| CN111104834A (en) | Application method of cross-contrast neural network in intelligent detection of heart sound |
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 | ||
| DD01 | Delivery of document by public notice |
Addressee: Zhu Haochuan Document name: First notice of examination opinions |
|
| DD01 | Delivery of document by public notice | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20200505 |
|
| WD01 | Invention patent application deemed withdrawn after publication |