CN110459303B - Medical image abnormity detection device based on depth migration - Google Patents
Medical image abnormity detection device based on depth migration Download PDFInfo
- Publication number
- CN110459303B CN110459303B CN201910568282.8A CN201910568282A CN110459303B CN 110459303 B CN110459303 B CN 110459303B CN 201910568282 A CN201910568282 A CN 201910568282A CN 110459303 B CN110459303 B CN 110459303B
- Authority
- CN
- China
- Prior art keywords
- medical image
- image
- medical
- feature
- attention
- 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
- 238000001514 detection method Methods 0.000 title claims abstract description 89
- 238000013508 migration Methods 0.000 title claims abstract description 11
- 230000005012 migration Effects 0.000 title claims abstract description 11
- 238000012549 training Methods 0.000 claims abstract description 30
- 230000005856 abnormality Effects 0.000 claims abstract description 27
- 238000000605 extraction Methods 0.000 claims abstract description 18
- 238000004364 calculation method Methods 0.000 claims abstract description 8
- 238000004590 computer program Methods 0.000 claims abstract description 8
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 claims description 32
- 201000010099 disease Diseases 0.000 claims description 23
- 208000035977 Rare disease Diseases 0.000 claims description 15
- 238000000034 method Methods 0.000 claims description 14
- 230000006870 function Effects 0.000 claims description 11
- 239000011159 matrix material Substances 0.000 claims description 10
- 230000007246 mechanism Effects 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 5
- 238000010276 construction Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 4
- 230000009466 transformation Effects 0.000 claims description 3
- 230000009286 beneficial effect Effects 0.000 claims description 2
- 238000006243 chemical reaction Methods 0.000 claims description 2
- 239000000126 substance Substances 0.000 claims description 2
- 238000009966 trimming Methods 0.000 claims description 2
- 230000002159 abnormal effect Effects 0.000 description 8
- 238000013527 convolutional neural network Methods 0.000 description 8
- 230000000694 effects Effects 0.000 description 5
- 238000013526 transfer learning Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 3
- 238000003745 diagnosis Methods 0.000 description 3
- 239000000284 extract Substances 0.000 description 3
- 210000000245 forearm Anatomy 0.000 description 3
- 210000002758 humerus Anatomy 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 238000007792 addition Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 210000000481 breast Anatomy 0.000 description 2
- 238000004422 calculation algorithm Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 238000007781 pre-processing Methods 0.000 description 2
- 206010006187 Breast cancer Diseases 0.000 description 1
- 208000026310 Breast neoplasm Diseases 0.000 description 1
- 206010007759 Cataract nuclear Diseases 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 210000000988 bone and bone Anatomy 0.000 description 1
- 201000011510 cancer Diseases 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000002059 diagnostic imaging Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 230000003902 lesion Effects 0.000 description 1
- 210000004072 lung Anatomy 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000002156 mixing Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 208000029552 nuclear cataract Diseases 0.000 description 1
- 210000004205 output neuron Anatomy 0.000 description 1
- 230000007170 pathology Effects 0.000 description 1
- 238000012913 prioritisation Methods 0.000 description 1
- 238000013138 pruning Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
- 210000002832 shoulder Anatomy 0.000 description 1
- 210000002027 skeletal muscle Anatomy 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 210000000707 wrist Anatomy 0.000 description 1
Images
Classifications
-
- 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
- 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
-
- G06T5/70—
-
- G06T5/73—
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H30/00—ICT specially adapted for the handling or processing of medical images
- G16H30/40—ICT specially adapted for the handling or processing of medical images for processing medical images, e.g. editing
Abstract
The invention discloses a medical image abnormity detection device based on depth migration, which comprises a computer memory, a computer processor and a computer program, wherein a medical image abnormity detection model is stored in the computer memory, and the computer processor executes the computer program to realize the following steps: denoising and sharpening the medical image to be tested to obtain a medical input image; inputting the medical input image into a medical image abnormity detection model, and outputting the medical input image to a detection result through calculation; the medical image anomaly detection model is obtained by training a medical image anomaly detection network consisting of a feature extraction unit, an attention unit and a classification unit by using a training sample. The medical image abnormality detection device can quickly detect the abnormality of the medical image.
Description
Technical Field
The invention belongs to the field of medical image abnormity detection, and particularly relates to a medical image abnormity detection device based on depth migration.
Background
The anomaly detection of medical images, such as skeletal muscle X-ray film anomaly detection, has important clinical application value. First, the anomaly detection model can be used for prioritization of worklists, allowing more rapid diagnosis and treatment of severe patients by advancing detected anomalies during image interpretation. In addition, automatic anomaly detection may help alleviate radiologists 'fatigue, and one study shows that radiologists have a significant drop in the efficiency of fracture detection and the start of work near the end of the day's work. A model capable of carrying out anomaly detection can carry out importance degree sequencing on medical images to be detected, and the images of patients most likely to be detected as patients are preferentially provided for clinicians, so that doctors are helped to improve film reading efficiency and quality. At present, a method for assisting a doctor in an abnormality detection task by using artificial intelligence is mainly used for modeling based on deep learning according to medical images, and becomes a research hotspot of the artificial intelligence in the medical field. The difficulty of using deep learning to predict the abnormality of the medical image is as follows: firstly, how to extract effective medical image features; secondly, for some unusual diseases, the problem of insufficient medical image data exists, an effective detection model cannot be constructed, the detection effect is directly influenced, and the obstacle in the aspect is reduced.
In order to extract effective medical image features, the current most important methods include conventional image processing techniques, and depth model extraction. The former includes utilizing the outer edge of a shape based on boundaries, utilizing the entire area of a shape based on regions, and texture features. The latter directly inputs the medical image into the depth model for feature extraction so as to directly detect or serve as the feature of the detection model. With the continuous development of depth models based on Convolutional Neural Networks (CNN), the performance of feature extraction using the depth models is often better than that of traditional image features. Arevalo et al propose a feature learning framework for breast cancer diagnosis, and classify the breast X-ray photograph lesions by adopting CNN to automatically learn distinctive features; the method has the advantages that characteristics of shear wave elastic images are extracted by constructing Boltzmann machines and RBM depth models by Zhang et al of Shanghai university, and better classification performance of benign and malignant tumors is realized; spampipinato et al used deep CNN to assess skeletal bone age. Some other methods combine the CNN with a Recurrent Neural Network (RNN), such as Gao et al, which extracts the bottom layer local feature information in slit-lamp images by using the CNN, and extracts the high layer features by combining with the RNN, thereby grading the nuclear cataract. The method mainly extracts the characteristics of different medical images based on the depth model of the CNN or RNN structure, and achieves certain effect in respective clinical fields.
In addition, to solve the problems of small data volume of some unusual diseases and high cost of expert labeling, transfer learning is introduced into medical image detection. The medical image detection method based on transfer learning can be mainly divided into two strategies, one of which uses a pre-training network as a feature extractor. Research finds that although medical images are very different from natural images, the network model trained by ImageNet can still be used for medical image recognition, such as a breast pathology recognition feature generator used by Bar et al to pre-train the model; ginneken et al combine CNN features with manual features to improve nodule detection system performance. The second is to use the target medical data to fine-tune the pre-trained network, to use the limited labeled data set related to the task to fine-tune the network parameters in a supervision mode, to adjust several layers or all parameters of the network, such as Ciompi et al, which uses ImageNet to pre-train the coiling machine neural network, to use a small amount of labeled CT data to fine-tune the network in a supervision mode, to automatically detect the nodules around the lung.
In general, the current medical image anomaly detection algorithm has the following problems: in the process of anomaly detection of medical images, how to extract key features helps to solve the problem of anomaly detection; the transfer learning can be applied to the abnormal detection of medical images, and how to enable the transfer learning algorithm to fully utilize the useful information of common disease images, thereby improving the detection performance of the unusual diseases.
Disclosure of Invention
In view of the above, the present invention provides a medical image abnormality detection apparatus based on depth migration, which is capable of quickly detecting an abnormality of a medical image.
The technical scheme of the invention is as follows:
a medical image abnormality detection apparatus based on depth migration, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, the computer memory having a medical image abnormality detection model stored therein, the computer processor implementing the following steps when executing the computer program:
denoising and sharpening the medical image to be tested to obtain a medical input image;
inputting the medical input image into a medical image abnormity detection model, and outputting the medical input image to a detection result through calculation;
the medical image anomaly detection model is obtained by training a medical image anomaly detection network consisting of a feature extraction unit, an attention unit and a classification unit by using a training sample.
The invention adopts a medical image anomaly detection model based on deep migration, which fully utilizes the strong characterization capability of deep learning to obtain deep image characteristics, and combines a migration learning mechanism to increase the mobility of the characteristics among diseases, thereby finally optimizing the anomaly detection performance among different diseases.
The construction method of the medical image anomaly detection model comprises the following steps:
constructing a training sample, carrying out denoising and sharpening on the obtained medical image containing common diseases and rare diseases, and carrying out data enhancement on the denoised and sharpened medical image to obtain the training sample;
constructing a medical image abnormity detection network, wherein the medical image abnormity detection network comprises a feature extraction unit, an attention unit and a classification unit, wherein the feature extraction unit comprises at least 2 layers of convolution layers which are sequentially connected and is used for extracting image features of a medical image; the attention unit comprises an attention layer of a self-attention mechanism and is used for weighting the image characteristics to obtain weighted characteristics; the classification unit comprises at least 2 layers of all-connected layers which are connected in sequence and is used for performing classification prediction on the weighted features;
training the medical image anomaly detection network, training the medical image anomaly detection network by using a training sample, and determining network parameters when the training is ended to obtain a medical image anomaly detection model.
The medical image denoising method adopts a median filtering technology to denoise the medical image, and after the denoising treatment, the denoised image needs to be sharpened, so that the image contour is enhanced, and the quality loss of the medical image after transmission or conversion is compensated.
When a training sample is constructed, after a medical image is sharpened, data enhancement needs to be performed on the medical image to enhance information expression of an original medical image, specifically, performing data enhancement on the medical image subjected to denoising and sharpening includes:
and respectively converting the medical image according to horizontal turning, random trimming and scale transformation so as to realize data enhancement of the medical image.
In another embodiment, in the medical image abnormality detection network, the feature extraction unit includes 5 convolutional layers connected in sequence; the classification unit comprises 4 fully-connected layers which are connected in sequence.
Specifically, during training, a classification loss is constructed by using a classification unit to the cross entropy of the classification predicted value of the training sample and the real label of the training sample;
constructing MMD loss as the maximum mean difference of common and rare condition samples;
constructing a loss function of the medical image abnormity detection network according to the classification loss and the MMD loss, and updating parameters of the medical image abnormity detection network according to the loss function;
wherein, the MMD loss calculation formula is as follows:
wherein src is a common disorder set, tar is a rare disorder set, n1Number of samples of common diseases, n2For rare disease samples, srciIs the ith common disorder sample, tariIs the ith rare disorder sample.
When the medical image abnormality detection model is applied, only the disease images of a large number of samples, the class marks of the disease images and target diseases without the class marks are needed to be used as input of the medical image abnormality detection model. The requirement on rare disease data can be reduced, model fine adjustment is not needed under the supervision of class labels, and the obtained result can assist a clinician in diagnosis.
Compared with the prior art, the invention has the following technical effects:
the medical image abnormity detection device provided by the invention can realize detection of few diseases of a sample.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.
Fig. 1 is a block diagram of a construction process of a medical image abnormality detection model according to the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be further described in detail with reference to the accompanying drawings and examples. It should be understood that the detailed description and specific examples, while indicating the scope of the invention, are intended for purposes of illustration only and are not intended to limit the scope of the invention.
In order to implement abnormality detection on medical images, the embodiment provides a medical image abnormality detection apparatus based on depth migration, which includes a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, wherein a medical image abnormality detection model is stored in the computer memory, and the computer processor implements the following steps when executing the computer program:
denoising and sharpening the medical image to be tested to obtain a medical input image;
inputting the medical input image into a medical image abnormity detection model, and outputting the medical input image to a detection result through calculation;
the medical image anomaly detection model is obtained by training a medical image anomaly detection network consisting of a feature extraction unit, an attention unit and a classification unit by using a training sample.
The medical image abnormality detection device can realize abnormality detection of the medical image.
The construction method of the medical image anomaly detection model comprises the following steps:
and constructing a training sample, carrying out denoising and sharpening on the obtained medical image, and carrying out data enhancement on the denoised and sharpened medical image to obtain the training sample.
Specifically, a median filtering method is adopted to denoise medical images of common diseases and rare diseases, namely, a median of nine pixels is used to replace all pixel values in a 3-by-3 pixel region in the medical image so as to denoise the medical image; and then, sharpening the denoised medical image to highlight the contour of the medical image.
Considering that for rare diseases, the number of samples is not enough, the samples need to be further subjected to data enhancement, and specifically, medical images are respectively converted according to horizontal inversion, random pruning and scale transformation, so that the original data set is expanded, and the obstacle of insufficient data volume is overcome.
On the basis of well-constructed training samples, constructing a medical image anomaly detection network, specifically, the medical image anomaly detection network comprises a feature extraction unit, an attention unit and a classification unit, wherein the feature extraction unit comprises 5 layers of convolution layers which are sequentially connected and is used for extracting image features of a medical image; the attention unit comprises an attention layer of a self-attention mechanism and is used for weighting the image characteristics to obtain weighted characteristics; the classification unit comprises 4 fully-connected layers which are connected in sequence and is used for classifying and predicting the weighted features.
After the medical image anomaly detection network is constructed, training the medical image anomaly detection network by using a training sample, wherein the specific process comprises the following steps:
(a) scaling the preprocessed medical images of common diseases and rare diseases to obtain 256 × 256-sized pictures, randomly cutting out sub-pictures with the size of 224 × 224 as input, and simultaneously inputting the sub-pictures into the first five layers of convolution layers (conv1-conv5) of AlexNet pre-trained by ImageNet as a feature extraction unit for extracting image features of the medical images;
(b) the extracted image features are taken as an attention layer introduced into a self-attention mechanism, so that the features beneficial to classification in the extracted image features are weighted heavily, and the specific process is as follows:
(b-1) performing convolution operation of 1 x 1 on the image feature x obtained by the feature extraction unit to obtain an image feature f (x), an image feature g (x) and an image feature h (x), wherein the difference is that the image feature h (x) still has the same size as the image feature x, and the size of the image feature g (x) is 1/8 depth of the image feature x;
(b-2) matrix multiplying the special matrix of the image characteristics f (x) with the image characteristics g (x), namely Sij＝f(xi)T*g(xj) Obtaining characteristic autocorrelation matrixes S, i and j between a single pixel point and all pixel points as pixel indexes;
(b-3) processing the autocorrelation feature S to obtain an attention weight matrix having each value in the range of 0 to 1;
wherein S isijIs the pixel value, α, at the (i, j) position in the autocorrelation matrix Sj，iIs the attention weight at the (i, j) position in the attention weight matrix α;
(b-4) the weighted features of the attention layer output are:
xAttention＝o+h(x)
wherein the content of the first and second substances,
(c) each fully connected layer is learning a nonlinear mapping
Wherein
Is an implicit expression characteristic x of the l < th > layer of the full link layeri，wlAnd blWeight and offset, f, of the l-th fully-connected layer, respectivelylFor the activation function of the l-th fully-connected layer, a Leaky ReLU function is adopted, namely
fl(x)＝max(0，a*x)
Where a is set to 0.8.
The model comprises four fully-connected layers, wherein the first three layers are fully-connected layers of AlexNet, and the last fully-connected layer has two output neurons because the anomaly detection is a two-classification problem in essence.
(d) Inputting the weighted features of the common disease samples into the full-connection layer for classification, and then the classification error of the model is
Where φ (-) represents the non-linear representation hidden by the network,
for the expression of the characteristics of a certain sample,
represents
Belong to class mark
F (-) is a cross entropy calculation function. This is taken as the classification loss function and is marked as LClassification；
(e) Inputting common disease and rare disease samples into a full-connection layer, calculating the maximum mean difference of each layer, adding the maximum mean differences, calculating the mean, and adjusting model parameters to enable sample distribution represented by features generated by a source item data set and a target data set to be similar as much as possible, wherein the maximum mean difference calculation mode is as follows:
LMMD＝dist(src，tar)
wherein src is a common disease data set, tar is a rare disease data set, n1Number of samples in the common disease data set, n2For rare disease data set sample number, srciThe ith sample in the common disease, tariThe ith sample in rare diseases. This is taken as the MMD loss function and is marked as LMMD；
(f) Mixing L withMMDAnd LClassificationA weighted addition is performed, and finally the loss function L,
L＝LClassification+λ*LMMD
wherein, λ is a penalty coefficient set by a human, and the effect is better when the penalty coefficient is set to 0.5. Finally, fine adjustment is carried out on the whole network according to the loss function so as to obtain a medical image abnormity detection model with a good rare disease detection effect.
After the medical image abnormity detection model is obtained, the medical image to be detected is input into the medical image abnormity detection model, and whether the medical image is abnormal or not can be judged.
According to the medical image preprocessing method, the medical image preprocessing is carried out by utilizing a traditional image processing mode, so that the characteristic extraction of the medical image by the characteristic extraction unit is facilitated; data expansion is achieved by using a plurality of rows of data enhancement technologies, so that the problems of unbalanced samples and small quantity in original data are solved; in addition, the sample distribution difference between common diseases and rare diseases is reduced as much as possible while focusing on important features by combining a self-attention mechanism, so that a medical image abnormality detection model capable of assisting rare disease detection by using common disease knowledge is constructed. F-measure is used as an evaluation index to measure the performance of the material. The calculation method of the F-measure comprises the following steps:
F-measure＝(2*P*R)/(P+R)
P＝TP/(TP+FP)
R＝TP/(TP+FN)
wherein, TP is the number of true samples, which represents the number of true abnormal samples in the samples predicted to have abnormality; FP is the number of false positive samples, which represents the number of actually normal samples in the samples predicted to have abnormality; FN is the false negative number of samples, representing the number of samples that are actually abnormal among the samples predicted to be normal. On the basis, the classification accuracy is calculated, wherein P is the proportion of the samples which are predicted to be abnormal and are actually abnormal, and the higher the value is, the more accurate the classifier is represented; r is the recall rate of classification, and refers to the proportion of samples which are predicted to have the abnormality in the samples which are actually abnormal, and the higher the value is, the more the samples with the abnormality are found. The F-measure is a harmonic average value of the accuracy and the recall rate, and the higher the value is, the more and more the depth model formed by the group of parameters is accurately classified to the abnormal samples, namely the better the prediction performance is.
40895X-ray musculoskeletal images from 14982 patients were examined using a medical imaging abnormality examination device, including seven sites including elbow, finger, forearm, hand, humerus, shoulder, and wrist, with relatively few cases of forearm and humerus. In the abnormality detection experiment for the forearm, the average F-measure is 0.704; in the test for detecting abnormality of humerus, the average F-measure was 0.691. The comprehensive experiment result shows that the medical image abnormity detection device can complete the detection of rare diseases in medical images.
The medical image anomaly detection method based on depth migration does not aim at treatment, and can be used for scientific research.
The above-mentioned embodiments are intended to illustrate the technical solutions and advantages of the present invention, and it should be understood that the above-mentioned embodiments are only the most preferred embodiments of the present invention, and are not intended to limit the present invention, and any modifications, additions, equivalents, etc. made within the scope of the principles of the present invention should be included in the scope of the present invention.
Claims (2)
1. A depth migration based medical image abnormality detection apparatus, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, wherein a medical image abnormality detection model is stored in the computer memory, and the computer processor executes the computer program to implement the following steps:
denoising and sharpening the medical image to be tested to obtain a medical input image;
inputting the medical input image into a medical image abnormity detection model, and outputting the medical input image to a detection result through calculation;
the construction method of the medical image anomaly detection model comprises the following steps:
constructing a training sample, carrying out denoising and sharpening on the obtained medical image containing common diseases and rare diseases, and respectively carrying out conversion on the denoised and sharpened medical image according to horizontal turning, random trimming and scale transformation so as to realize data enhancement on the medical image and obtain the training sample;
constructing a medical image abnormity detection network, wherein the medical image abnormity detection network comprises a feature extraction unit, an attention unit and a classification unit, wherein the feature extraction unit comprises 5 layers of convolution layers which are sequentially connected and is used for extracting image features of a medical image; the attention unit comprises an attention layer of a self-attention mechanism and is used for weighting the image characteristics to obtain weighted characteristics; the classification unit comprises 4 layers of all-connected layers which are connected in sequence and is used for performing classification prediction on the weighted features;
training a medical image anomaly detection network, wherein the medical image anomaly detection network is trained by using a training sample, and during training, a classification loss is constructed by using a classification unit to the cross entropy of a classification predicted value of the training sample and a real label of the training sample; constructing MMD loss as the maximum mean difference of common and rare condition samples; constructing a loss function of the medical image abnormity detection network according to the classification loss and the MMD loss, and updating parameters of the medical image abnormity detection network according to the loss function; when the training is ended, determining network parameters to obtain a medical image abnormity detection model;
in the attention layer, the feature which is beneficial to classification in the extracted image features is weighted heavily, and the specific process is as follows:
(b-1) performing convolution operation of 1 x 1 on the image feature x obtained by the feature extraction unit to obtain an image feature f (x), an image feature g (x) and an image feature h (x), wherein the difference is that the image feature h (x) still has the same size as the image feature x, and the size of the image feature g (x) is 1/8 depth of the image feature x;
(b-2) matrix multiplying the special matrix of the image characteristics f (x) with the image characteristics g (x), namely Sij＝f(xi)T*g(xj) Obtaining characteristic autocorrelation matrixes S, i and j between a single pixel point and all pixel points as pixel indexes;
(b-3) processing the autocorrelation feature S to obtain an attention weight matrix having each value in the range of 0 to 1;
wherein S isijIs the pixel value, α, at the (i, j) position in the autocorrelation matrix Sj,iIs the attention weight at the (i, j) position in the attention weight matrix α;
(b-4) the weighted features of the attention layer output are:
xAttention＝o+h(x)
wherein the content of the first and second substances,
2. the apparatus according to claim 1, wherein the MMD loss is calculated by the following formula:
wherein src is a common disorder set, tar is a rare disorder set, n1Number of samples of common diseases, n2Number of samples of rare disorders, srciIs the ith common disorder sample, tariIs the ith rare disorder sample.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910568282.8A CN110459303B (en) | 2019-06-27 | 2019-06-27 | Medical image abnormity detection device based on depth migration |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910568282.8A CN110459303B (en) | 2019-06-27 | 2019-06-27 | Medical image abnormity detection device based on depth migration |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110459303A CN110459303A (en) | 2019-11-15 |
| CN110459303B true CN110459303B (en) | 2022-03-08 |
Family
ID=68481215
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910568282.8A Active CN110459303B (en) | 2019-06-27 | 2019-06-27 | Medical image abnormity detection device based on depth migration |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110459303B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111192679B (en) * | 2019-12-25 | 2024-04-19 | 上海联影智能医疗科技有限公司 | Method, device and storage medium for processing image data abnormality |
| CN111754453A (en) * | 2020-05-11 | 2020-10-09 | 佛山市第四人民医院(佛山市结核病防治所) | Pulmonary tuberculosis detection method and system based on chest radiography image and storage medium |
| CN113707312A (en) * | 2021-09-16 | 2021-11-26 | 人工智能与数字经济广东省实验室(广州) | Blood vessel quantitative identification method and device based on deep learning |
Family Cites Families (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10565708B2 (en) * | 2017-09-06 | 2020-02-18 | International Business Machines Corporation | Disease detection algorithms trainable with small number of positive samples |
| US10671918B2 (en) * | 2017-10-24 | 2020-06-02 | International Business Machines Corporation | Attention based sequential image processing |
| CN108021916B (en) * | 2017-12-31 | 2018-11-06 | 南京航空航天大学 | Deep learning diabetic retinopathy sorting technique based on attention mechanism |
| CN108492273A (en) * | 2018-03-28 | 2018-09-04 | 深圳市唯特视科技有限公司 | A kind of image generating method based on from attention model |
| CN108491817B (en) * | 2018-03-30 | 2021-02-26 | 国信优易数据股份有限公司 | Event detection model training method and device and event detection method |
| KR101953752B1 (en) * | 2018-05-31 | 2019-06-17 | 주식회사 뷰노 | Method for classifying and localizing images using deep neural network and apparatus using the same |
| CN109165667B (en) * | 2018-07-06 | 2022-02-22 | 中国科学院自动化研究所 | Brain disease classification system based on self-attention mechanism |
| CN109635835A (en) * | 2018-11-08 | 2019-04-16 | 深圳蓝韵医学影像有限公司 | A kind of breast lesion method for detecting area based on deep learning and transfer learning |
| CN109523526B (en) * | 2018-11-08 | 2021-10-22 | 腾讯科技（深圳）有限公司 | Tissue nodule detection and model training method, device, equipment and system thereof |
| CN109493308B (en) * | 2018-11-14 | 2021-10-26 | 吉林大学 | Medical image synthesis and classification method for generating confrontation network based on condition multi-discrimination |
-
2019
- 2019-06-27 CN CN201910568282.8A patent/CN110459303B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN110459303A (en) | 2019-11-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Qi et al. | Automated diagnosis of breast ultrasonography images using deep neural networks | |
| CN108364006B (en) | Medical image classification device based on multi-mode deep learning and construction method thereof | |
| CN108257135A (en) | The assistant diagnosis system of medical image features is understood based on deep learning method | |
| CN110459303B (en) | Medical image abnormity detection device based on depth migration | |
| CN107203989A (en) | End-to-end chest CT image dividing method based on full convolutional neural networks | |
| CN106056595A (en) | Method for automatically identifying whether thyroid nodule is benign or malignant based on deep convolutional neural network | |
| CN112784879A (en) | Medical image segmentation or classification method based on small sample domain self-adaption | |
| CN111429407A (en) | Chest X-ray disease detection device and method based on two-channel separation network | |
| Lan et al. | Run: Residual u-net for computer-aided detection of pulmonary nodules without candidate selection | |
| CN114398979A (en) | Ultrasonic image thyroid nodule classification method based on feature decoupling | |
| CN112508884A (en) | Comprehensive detection device and method for cancerous region | |
| Pradhan et al. | Lung cancer detection using 3D convolutional neural networks | |
| Khan | Identification of lung cancer using convolutional neural networks based classification | |
| Wang et al. | A ResNet‐based approach for accurate radiographic diagnosis of knee osteoarthritis | |
| Kaur et al. | Analysis of brain tumor using pre-trained CNN models and machine learning techniques | |
| Cheng et al. | Quantitative evaluation of liver fibrosis based on ultrasound radio frequency signals: An animal experimental study | |
| Khan et al. | A novel approach to predict brain cancerous tumor using transfer learning | |
| CN109767429A (en) | A kind of image screening method and device | |
| CN115880266B (en) | Intestinal polyp detection system and method based on deep learning | |
| Yektaei et al. | Diagnosis of lung cancer using multiscale convolutional neural network | |
| CN114565786A (en) | Tomography image classification device and method based on channel attention mechanism | |
| Roopa et al. | Mortality Prediction of Lung Cancer from CT Images Using Deep Learning Techniques | |
| Reddy et al. | Different Algorithms for Lung Cancer Detection and Prediction | |
| Dodamani et al. | Transfer Learning-Based Osteoporosis Classification Using Simple Radiographs | |
| Hassan et al. | An Expedition to Intelligent Diagnosis of Bone Cancer and It's Direction from Capsule Network |
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 |