CN110097974A - A kind of nasopharyngeal carcinoma far-end transfer forecasting system based on deep learning algorithm - Google Patents
A kind of nasopharyngeal carcinoma far-end transfer forecasting system based on deep learning algorithm Download PDFInfo
- Publication number
- CN110097974A CN110097974A CN201910407356.XA CN201910407356A CN110097974A CN 110097974 A CN110097974 A CN 110097974A CN 201910407356 A CN201910407356 A CN 201910407356A CN 110097974 A CN110097974 A CN 110097974A
- Authority
- CN
- China
- Prior art keywords
- far
- end transfer
- nasopharyngeal carcinoma
- deep learning
- feature
- 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
- 238000012546 transfer Methods 0.000 title claims abstract description 102
- 206010061306 Nasopharyngeal cancer Diseases 0.000 title claims abstract description 72
- 208000002454 Nasopharyngeal Carcinoma Diseases 0.000 title claims abstract description 71
- 201000011216 nasopharynx carcinoma Diseases 0.000 title claims abstract description 70
- 238000013135 deep learning Methods 0.000 title claims abstract description 36
- 230000001575 pathological effect Effects 0.000 claims abstract description 55
- 238000007781 pre-processing Methods 0.000 claims abstract description 9
- 238000004458 analytical method Methods 0.000 claims abstract description 5
- 238000012549 training Methods 0.000 claims description 51
- 206010028980 Neoplasm Diseases 0.000 claims description 38
- 238000000034 method Methods 0.000 claims description 17
- 230000006870 function Effects 0.000 claims description 10
- 238000013136 deep learning model Methods 0.000 claims description 6
- 239000000284 extract Substances 0.000 claims description 6
- 230000007170 pathology Effects 0.000 claims description 6
- 210000004881 tumor cell Anatomy 0.000 claims description 6
- 238000001914 filtration Methods 0.000 claims description 4
- 230000008569 process Effects 0.000 claims description 4
- 238000012790 confirmation Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 230000035945 sensitivity Effects 0.000 claims description 3
- 238000010200 validation analysis Methods 0.000 claims description 3
- 238000011496 digital image analysis Methods 0.000 claims description 2
- 210000001989 nasopharynx Anatomy 0.000 claims 1
- 230000008961 swelling Effects 0.000 claims 1
- 238000005516 engineering process Methods 0.000 abstract description 2
- 108090000623 proteins and genes Proteins 0.000 abstract description 2
- 238000012163 sequencing technique Methods 0.000 abstract description 2
- 238000013528 artificial neural network Methods 0.000 description 5
- 238000010606 normalization Methods 0.000 description 5
- 201000010099 disease Diseases 0.000 description 4
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 4
- 238000013507 mapping Methods 0.000 description 4
- 238000005070 sampling Methods 0.000 description 4
- 238000012216 screening Methods 0.000 description 4
- 206010027476 Metastases Diseases 0.000 description 3
- 201000011510 cancer Diseases 0.000 description 3
- 238000001514 detection method Methods 0.000 description 3
- 238000000605 extraction Methods 0.000 description 3
- 238000000513 principal component analysis Methods 0.000 description 3
- 238000004393 prognosis Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 2
- 208000001894 Nasopharyngeal Neoplasms Diseases 0.000 description 2
- 208000009565 Pharyngeal Neoplasms Diseases 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000011835 investigation Methods 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 210000001331 nose Anatomy 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 201000008006 pharynx cancer Diseases 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000001574 biopsy Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000002512 chemotherapy Methods 0.000 description 1
- 238000004195 computer-aided diagnosis Methods 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 238000002059 diagnostic imaging Methods 0.000 description 1
- 238000013399 early diagnosis Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 238000007477 logistic regression Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 230000009401 metastasis Effects 0.000 description 1
- 230000001394 metastastic effect Effects 0.000 description 1
- 206010061289 metastatic neoplasm Diseases 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000001823 molecular biology technique Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000008520 organization Effects 0.000 description 1
- 238000001959 radiotherapy Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 206010041823 squamous cell carcinoma Diseases 0.000 description 1
- 230000001225 therapeutic effect Effects 0.000 description 1
- 238000013526 transfer learning Methods 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16B—BIOINFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR GENETIC OR PROTEIN-RELATED DATA PROCESSING IN COMPUTATIONAL MOLECULAR BIOLOGY
- G16B30/00—ICT specially adapted for sequence analysis involving nucleotides or amino acids
-
- 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
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/30—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Medical Informatics (AREA)
- Evolutionary Biology (AREA)
- Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Public Health (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Analytical Chemistry (AREA)
- Proteomics, Peptides & Aminoacids (AREA)
- Biotechnology (AREA)
- Biophysics (AREA)
- Spectroscopy & Molecular Physics (AREA)
- Chemical & Material Sciences (AREA)
- Primary Health Care (AREA)
- Epidemiology (AREA)
- Pathology (AREA)
- Databases & Information Systems (AREA)
- Biomedical Technology (AREA)
- Image Analysis (AREA)
Abstract
The nasopharyngeal carcinoma far-end transfer forecasting system based on deep learning algorithm that the invention discloses a kind of, the forecasting system comprise the following modules: image capture module, image pre-processing module, patient information collection module, the far-end transfer prediction module based on deep learning algorithm and predictive information output module；System of the present invention realizes the risk profile of Nasopharyngeal Carcinoma Patients far-end transfer in Pathological levels, compared with conventional molecular biological technology and gene sequencing etc. are the analysis Diagnosis-treat Model of main means, the efficiency for obtaining diagnostic result is improved, the prediction judgement that clinician can not independently carry out is also achieved.
Description
Technical field
The invention belongs to computer-aided diagnosis system fields, more specifically, are related to a kind of based on deep learning calculation
The nasopharyngeal carcinoma far-end transfer forecasting system of method.
Background technique
Nasopharyngeal carcinoma refers to the malignant tumour betided at the top of nasopharyngeal cavity with side wall, is one of high-incidence malignant tumour in China.Generation
Boundary's health organization investigation report, there are 80% Nasopharyngeal Carcinoma Patients in the whole world in China.The disease incidence of nasopharyngeal carcinoma with China south compared with
It is high.It is reported that living in the middle part of Guangdong Province and saying the male of Guangdong place language, disease incidence is 30/,100,000~50/,100,000.
For the whole nation, the disease incidence of nasopharyngeal carcinoma is gradually decreased by south to north, and most northern disease incidence is not higher than 2/,100,000~3/,100,000.
Its occurrence and development is related to many factors.
About 98% nasopharyngeal carcinoma category poorly differentiated squamous cell carcinoma, the preferred therapeutic scheme with radiotherapy cooperation chemotherapy.Nose
The local recurrence and far-end transfer in pharynx cancer later period are the major causes of death of Nasopharyngeal Carcinoma Patients, and about 20% patient is in the standard of progress
Far-end transfer occurs after changing 2 years for the treatment of.Although the Molecular Biology Mechanism that clinically far-end transfer occurs at present has more
In-depth study, but at present there is no reliable method can to many prognosis informations including distant metastasis of human into
Row accurate judgement.Therefore, the artificial intelligence system for accurately and effectively judging nasopharyngeal carcinoma far-end transfer occurrence risk is found and establishes,
Far-end transfer whether occurs by the data Accurate Prediction Nasopharyngeal Carcinoma Patients obtained in early diagnosis, and turns to distal end may occur
The patient of shifting gives timely Individual treatment and treatment, has a very important significance for reducing NPC mortality.
Summary of the invention
It is an object of the present invention to provide a kind of nasopharyngeal carcinoma far-end transfer forecasting system based on deep learning algorithm, the system base
In deep learning algorithm, by delineating, cutting and pre-processing etc. that obtain can be defeated to nasopharyngeal carcinoma tumor pathological section digital picture
Enter the digital picture of deep neural network, detectable nasopharyngeal carcinoma tumor pathological image distal end is then trained by training data and is turned
The weight and scoring of shifting, and the Nasopharyngeal Carcinoma Patients clinical information through screening is combined, determine whether the digital picture of input has far
Transfer characteristic is held, to realize the prediction to nasopharyngeal carcinoma far-end transfer.
To achieve the goals above, The technical solution adopted by the invention is as follows:
A kind of nasopharyngeal carcinoma far-end transfer forecasting system based on deep learning algorithm, including far-end transfer prediction module and pre-
Measurement information output module；
Far-end transfer prediction module, is used for: processed nasopharyngeal carcinoma tumor pathological section digital picture and patient are faced
Bed information is trained and optimizes, and passes through deep learning algorithm process, cuts to each processed nasopharyngeal carcinoma tumor pathology
Piece digital picture is classified, and judging that patient has the pathological image quantity of far-end transfer risk is a, does not have far-end transfer
The pathological image quantity of risk is b, show that the scoring S=a/ (a+b) of far-end transfer, S value occur for Nasopharyngeal Carcinoma Patients tumour cell
Section is 0-1；
Predictive information output module, is used for: establishing Nasopharyngeal Carcinoma Patients and far-end transfer risk and far-end transfer prediction mould occurs
The functional relation of gained scoring S, i.e. P=f (S) in block, wherein P is that deep learning calculates belonging to each clinical information of gained
The numerical value that weight obtains；P value is calculated separately according to the scoring S of training data in deep learning algorithm training set, to generation and not
The P value that far-end transfer occurs carries out section respectively and is classified as (Pma,Pmb) and (Pna,Pnb)；Wherein PmaIt is remote to have in training set
Hold the minimum P value of the image of transfer characteristic, PmbFor the maximum P value of the image in training set with far-end transfer feature, PnaFor instruction
Practice the minimum P value for concentrating the image without far-end transfer feature, PnbNot have the image of far-end transfer feature in training set
Maximum P value, and Pma>Pnb；If P ∈ (Pma,Pmb) 1 is then exported, representative has far-end transfer risk, if P ∈ (Pna,Pnb) then defeated
Out 0, it represents without far-end transfer risk.
Further, further include image pre-processing module, be used for: to being delineated in Nasopharyngeal Carcinoma Patients pathological section digital picture
Tumor region cut, and background mistake is carried out using the digital picture of gray threshold method or Otsu threshold method to cutting
Noise and dirty data are eliminated in filter；It is then normalized, obtains processed nasopharyngeal carcinoma tumor pathological section digitized map
Picture is sent to far-end transfer prediction module.
Further, further include image capture module, be used for: (1) the nasopharyngeal carcinoma tumor pathology for reading complete scan is cut
Piece digital picture sketches out the tumour cell region in image, is sent to image pre-processing module；(2) a large amount of known turns of storage
Training set of the nasopharyngeal carcinoma tumor pathological section digital picture of information as deep learning algorithm is moved, far-end transfer prediction is supplied to
Module.
Further, the far-end transfer prediction module completes processed nasopharyngeal carcinoma tumor pathological section digital picture
After analysis, processed nasopharyngeal carcinoma tumor pathological section digital picture is returned into training set, and whether send out in confirmation far-end transfer
After life, the processed nasopharyngeal carcinoma tumor pathological section digital picture and its corresponding result are added to training set.
Further, further include patient information collection module, be used for: the Detailed clinical data of Nasopharyngeal Carcinoma Patients is read
It takes and filtering screening, removes the incomplete case-data of clinical information, the patient clinical information screened.
Further, the far-end transfer prediction module contains convolutional layer, pond layer, local articulamentum, full articulamentum
With the multilayered structure of classifier, the last layer is the structure of one two classification prediction, uses mode of learning for non-supervisory formula
It practises.
Further, the function of the far-end transfer prediction module includes: that tumor region picture TSI is extracted and made and is based on
The training and validation data set of small slice；An improved deep learning model is established, and trains and can extract small slice feature
Model Weight；All slice features of one TSI figure are polymerize, the feature of TSI figure is obtained；To the spy of whole TSI figure
Sign trains cascade classifier, judges whether TSI figure has transfer attribute；Various assessments are carried out to training pattern.
Further, the far-end transfer prediction module based on deep learning algorithm carries out training pattern multi-party
The assessment in face, the i.e. analysis to processed nasopharyngeal carcinoma tumor pathological section digital picture and classification performance are weighed from various dimensions
Amount, including precision ACC, sensitivity S EN, specificity SPC and recipient's operating characteristics (ROC) area under the curve AUC；Wherein, SEN
Indicate ratio of the correctly predicted pathological picture without transfer, SPC indicates that correctly predicted pathological picture has the ratio of transfer, these measurements
Formula is defined as follows, and wherein TP represents the quantity of the pathological picture with transfer correctly judged, and TN represents the nothing correctly judged
The quantity of the pathological picture of transfer, FP represent the quantity of the pathological picture with transinformation of false judgment, and FN represents mistake
The quantity of the pathological picture without transinformation of judgement；
ACC=(TP+TN)/(TP+FP+TN+FN)
SEN=TN/ (TN+FP)
SPC=TP/ (TP+FN)
AUC=(SEN+SPC)/2
。
Compared with prior art, the beneficial effects of the present invention are:
(1) deep learning algorithm has been applied to pathology in the prediction of prognosis for the first time by the present invention, utilizes artificial intelligence
Technology realizes the prediction to prognosis；
(2) present invention realizes the risk profile of Nasopharyngeal Carcinoma Patients far-end transfer in Pathological levels for the first time, with conventional molecular
Biology techniques and gene sequencing etc. are compared for the analysis Diagnosis-treat Model of main means, improve the efficiency for obtaining diagnostic result,
The cost for reducing diagnosis also achieves the prediction judgement that clinician can not independently carry out；
(3) self-optimization is constantly carried out in such a way that test set returns to training set in deep learning algorithm of the invention,
Being continuously replenished by test set, identifies pathological image with far-end transfer feature etc. more on the basis of existing accuracy rate
For fine feature, the continuous self-optimization of prediction and perfect is realized, the final of prediction nasopharyngeal carcinoma far-end transfer is also improved
Accuracy rate.
Detailed description of the invention
Fig. 1 is the composition figure of the nasopharyngeal carcinoma far-end transfer prediction based on deep learning algorithm；
Fig. 2 is the algorithm flow chart of the embodiment of the present invention；
Fig. 3 a, Fig. 3 b, the original image that Fig. 3 c is the embodiment of the present invention are sliced；
Image slice after Fig. 3 d, Fig. 3 e, the standardization that Fig. 3 f is the embodiment of the present invention.
Specific embodiment
In a kind of nasopharyngeal carcinoma far-end transfer prediction based on deep learning algorithm proposed by the present invention, the specific of each module is led to
It crosses following manner and realizes its function:
Image capture module is by reading the nasopharyngeal carcinoma tumor pathological section digital picture of Nasopharyngeal Carcinoma Patients complete scan (again
>=400 times of number, resolution ratio >=12000 × 12000), and by artificial or software in Nasopharyngeal Carcinoma Patients pathological section digital picture
Tumour cell region (resolution ratio >=1000 × 1000) delineated, the stored a large amount of known transinformation noses of acquisition module
Pharynx cancer cancer pathology is sliced training set of the digital picture as deep learning algorithm；
Image pre-processing module: cutting the tumor region delineated in image capture module, and uses gray threshold
Method and Otsu threshold method carry out filtering background to the digital picture of cutting, eliminate noise and dirty data；Then carry out normalizing
Change processing, obtains processed nasopharyngeal carcinoma tumor pathological section digital picture, and image pixel is in 128 × 128 or 256 × 256
One kind；It is medium to be analyzed that deep neural network is input to through the obtained digital picture of series of preprocessing.
Patient information collection module is read out to the Detailed clinical data of nasopharyngeal cancer patient and filtering screening, and removal is clinical
The imperfect and false nasopharyngeal cancer patient case-data of information, the patient clinical information screened, and it is entered into depth
It is medium to be analyzed to spend neural network；
Far-end transfer prediction module based on deep learning algorithm contains convolutional layer, pond layer, local articulamentum, Quan Lian
The multilayered structure for connecing layer and classifier uses mode of learning to learn for non-supervisory formula.This exports image pre-processing module
The patient clinical letter of the screening of processed nasopharyngeal carcinoma tumor pathological section digital picture and the output of patient information collection module
Breath is trained and optimizes, and passes through deep learning algorithm process, to each processed nasopharyngeal carcinoma tumor pathological section number
Word image is classified, and judging that patient has the pathological image quantity of far-end transfer risk is a, does not have far-end transfer risk
Pathological image quantity be b, obtain Nasopharyngeal Carcinoma Patients tumour cell occur far-end transfer scoring S=a/ (a+b), S value section
For 0-1；Meanwhile after completing processed nasopharyngeal carcinoma tumor pathological section digital image analysis, by pathological section digital picture
Training set is returned, and after whether manual confirmation far-end transfer occurs, the nasopharyngeal carcinoma tumor pathology in image capture module is cut
Piece digital picture and its corresponding result are added to training set.
Nasopharyngeal Carcinoma Patients are established in predictive information output module, and far-end transfer risk and far-end transfer prediction module occurs
The functional relation of middle gained scoring S, i.e. P=f (S), wherein P is that deep learning calculates each clinical information ownership of gained
Heavy numerical value out；According to the information that pathological image acquisition module is sorted out, occur and do not occur the P value point of far-end transfer to it
Not carry out section be classified as (Pma,Pmb) and (Pna,Pnb)；Wherein PmaMost for the image in training set with far-end transfer feature
Small P value, PmbFor the maximum P value of the image in training set with far-end transfer feature, PnaNot have far-end transfer in training set
The minimum P value of the image of feature, PnbNot have the maximum P value of the image of far-end transfer feature, and general P in training setma
>Pnb；If P ∈ (Pma,Pmb) 1 is then exported, representative has far-end transfer risk, if P ∈ (Pna,Pnb) 0 is then exported, it represents and turns without distal end
Move risk.
In the far-end transfer prediction module based on deep learning algorithm:
Algorithm flow chart is as shown in Figure 2.The algorithm flow includes several main detection nasopharyngeal carcinoma transition phases.Firstly,
It is manually cut out tumor region (TSI) from whole image (WSI) respectively, and adaptive selecting may include metastases
The small slice in region and the small slice of tumour for not including metastases region.Data cleansing is carried out to all slices extracted
With pretreatment.Second step, establishes an improved deep learning model, and by feed training set data train it is detectable
The Model Weight of tumor region transfer.The feature being sliced by the model extraction.Third step, to all slices of a TSI figure
Feature is polymerize, and the feature of TSI figure is obtained.4th step trains cascade classifier to the feature of whole TSI figure, and judgement should
Whether TSI figure has transfer attribute.Finally, the present embodiment carries out various assessments to training pattern, deep learning model is aobvious
Its powerful performance in detection tumor region branch problem is shown.
Step 1: TSI is extracted and is made training and validation data set based on small slice:
The flat resolution of WSI image is about 12000 × 12000.In order to reduce the complexity of calculating and extract effective
Training data, artificial cutting cuts benefit tumor biopsy image (TSI) from WSI, and resolution ratio is about 300 × 300 to arrive
Between 2000 × 3000.Since the factors such as scanning circumstance, technological means influence, have in original nasopharyngeal carcinoma tumor WSI image a large amount of
Noise and dirty data will certainly make in training set and test set if directly extracting small slice using original WSI
Comprising a large amount of dirty data, model training effect is influenced.In order to extract effective slice from TSI, the present embodiment uses gray scale
(Otsu threshold method can also be used) in threshold method, and numerical value is set as 200, and when gray value is less than 200, the slice is chosen to be added
Enter training set.In order to reduce computation complexity while improve the validity of data, data are randomly selected from the TSI of data set,
And sampling ratio is set according to the resolution sizes of TSI.The positive and negative small slice sample of 256 × 256 sizes.The sample size calculates
It is as follows:
Num=WidthTSI/Widthpatch*HeightTSI/Heightpatch*sample_factor
Wherein num is the positive/negative sample number that the TSI can finally be got, WidthTSIFor the width of TSI figure, WidthpatchIt is small
The width of slice, HeightTSIFor the length of TSI figure, HeightpatchFor the length of small slice, sample_factor is sampling ratio, if
TSI is negative sample (not shifting), then sampling ratio is 1, if TSI is positive sample (transfer), then sampling ratio is 2.Successively
Achieve the purpose that positive and negative sample is balanced.It is special with transfer that 26,732 are extracted from 676 original pathological image data sets in total
Sign small slice and 27,688 without transfer characteristic small slice.
Data normalization: deep learning has the ability learnt from the data of nonstandardized technique.However, in some practical feelings
Under condition, image is indicated by multidimensional characteristic.If be directly trained using initial data, different scale features may
Result is had an impact.By normalization, different features can be become into identical scale.During this investigation it turned out, simply
Using two kinds of image normalization strategies: zero-mean normalization and z-score standardization.Zero-mean normalization, is widely used in
In machine learning algorithm, it is used in data set.Then z-score standardization is introduced on data set.Its standardized algorithm
It is as follows:
(1) zero-mean normalizes: for sequence of pictures x1,x2,…,xn, n indicates small number of sections here, and the present embodiment has
Here y1,y2,…,yn, indicate newly-generated sequence of pictures.
(2) z-score is normalized: for sequence of pictures x1,x2,…,xn, the present embodiment has
Here y1,y2,…,yn, indicate newly-generated sequence of pictures, and mean value is 0.Original image slice is described in Fig. 3
Slice (Fig. 3 d/e/f) after (Fig. 3 a b c) and standardization.
Second step establishes an improved deep learning model, and trains the Model Weight that can extract small slice feature:
Convolutional neural networks (CNN) have been widely used in medical imaging task.It is special that there is powerful extraction to be layered for it
The ability of sign.Although deep learning model also can guarantee its robustness when biggish data set, data set has very
Big noise.Therefore, the present embodiment is built by extracting suitable slice from TSI image to avoid those noisy pixels
Large-scale data set is found.
The disaggregated model of the present embodiment selection is existing to be all widely used in various Computer Vision Tasks.In this implementation
In the Detection task of example, other network structures are not changed, the knot in addition to network the last layer to be transformed into one two classification prediction
Structure.
It is very small since slice training set is compared to for the data set of ImageNet.Therefore, the present embodiment is adopted
The data of the present embodiment are trained with transfer learning.Reach the present embodiment training by modifying ResNet and extracts the mesh of feature
's.Table 1 illustrates specific neural network structure.Unlike original ResNet, the M-ResNet of the present embodiment is modified
The rear end part of original ResNet joined more full articulamentums, carry out Feature Selection.In training, original ResNet
Part is without training, that is to say, that the learning rate of this part is set as 0.The present embodiment directly uses the original part ResNet
Weight, and training modification part.In the data being previously mentioned, the division proportion of training set and verifying collection is 8:2.
The network structure of Table 1.M-ResNet
The present embodiment alleviates the training of network using the learning framework of residual error, and the network proposed is than being previously proposed
Network is deeper.This embodiment introduces residual blocks to accelerate training process.The architecture solves the explosion of neural network gradient
The problem of.
Depth residual block: if those layers are identical mappings behind deep layer network, model is just degenerated shallow for one
Layer network.What that be solved is exactly to learn identical mapping function.But some layers is directly allowed to go fitting one potentially identical
Mapping function H (x)=x, it is relatively difficult, trained reason is difficult to here it is deep layer network.If being H (x)=F network design
(x)+x can be converted to study one residual error function F (x)=H (x)-x.As long as F (x)=0, an identical mapping is constituted
H (x)=x.Moreover, regression criterion is more easier certainly.
The present embodiment defines loss function:
Loss function uses cross entropy cost function, and cross entropy is mainly used for calculating the error of actual value and predicted value.It is false
If x(i)For i-th of sample, y(i)For i-th actual value, hθ(x(i)) indicate i-th predicted value probability, m is sample number
Amount, the cross entropy of logistic regression can be expressed as:
If the image block sample of given one group of tape label, (x(i),y(i)) indicate i-th group of data label corresponding with it,
In the present invention, nasopharyngeal carcinoma prediction is considered to be two classification problems, y(i)Belong to 0 or 1, for Model Weight parameter θ, letter
Number is defined as:
For two classification problems:
When taking logarithm to have (1-3), (1-4):
In i-th group of sample, log probability is combined are as follows:
Then the cross entropy of entire sample may be expressed as:
Y herein(i)For i-th actual value, hθ(x(i)) indicate i-th predicted value probability, m is sample number
Amount is applied to the nasopharyngeal carcinoma prediction of two classification herein, then it is as follows to intersect entropy function:
Step 3: polymerizeing to all slice features of a TSI figure, the feature of TSI figure is obtained:
One TSI figure to be predicted is concentrated for three data, the present embodiment carries out sequence cutting to it, obtains to be predicted
256 × 256 sizes set of slices.These set of slices are sent into the neural network model that training finishes and extract feature, are obtained
To the dimensional feature of M × 512.M depends on the size of TSI.Since the M of every TSI is of different sizes, so if needing to obtain whole
The characteristic pattern of TSI, it is necessary to carry out feature selecting.There is used herein Principal Component Analysis (PCA) dimensionality reductions
Method, as follows
fj=PCA (Fj),j∈{1,...,512},
Wherein fjIt is upper j-th of the feature of TSI, FjIndicate that j-th of feature in this way, will be each in all M samples
Small slice feature in TSI figure aggregates into 512 dimensional features to indicate that TSI schemes.
Step 4: training cascade classifier to the feature of whole TSI figure, judge whether TSI figure has transfer attribute:
Due to TSI be it is unbalanced, screen out some non-diverting TSI figures, finally obtain in data set to pre-
The TSI of survey schemes, and after carrying out feature extraction described in third step respectively above-mentioned data set, and has carried out 5 foldings to data set and has intersected
Verifying.Training is simultaneously predicted using cascade classifier of the decision tree in conjunction with Adaboost.Final mask is obtained to differentiate TSI
Whether there is metastatic.
Step 5: carrying out various assessments to training pattern:
The present embodiment precision ACC, sensitivity S EN, specificity SPC and recipient's operating characteristics (ROC) area under the curve
AUC；Wherein, SEN indicates ratio of the correctly predicted pathological picture without transfer, and SPC indicates that correctly predicted pathological picture has transfer
Ratio, these measure formulas are defined as follows, and wherein TP represents the quantity of the pathological picture with transfer correctly judged, and TN is represented
The quantity of the pathological picture without transfer correctly judged, FP represent the number of the pathological picture with transinformation of false judgment
Amount, FN represent the quantity of the pathological picture without transinformation of false judgment.
ACC=(TP+TN)/(TP+FP+TN+FN)
SEN=TN/ (TN+FP)
SPC=TP/ (TP+FN)
AUC=(SEN+SPC)/2
。
In conclusion the contents of the present invention are not limited in the above embodiments, the knowledgeable people in same area can
Can propose other embodiments easily within technological guidance's thought of the invention, but this embodiment is included in this hair
Within the scope of bright.
Claims (8)
1. a kind of nasopharyngeal carcinoma far-end transfer forecasting system based on deep learning algorithm, which is characterized in that pre- including far-end transfer
Survey module and predictive information output module；
Far-end transfer prediction module, is used for: processed nasopharyngeal carcinoma tumor pathological section digital picture and patient clinical are believed
Breath is trained and optimizes, and passes through deep learning algorithm process, to each processed nasopharyngeal carcinoma tumor pathological section number
Word image is classified, and judging that patient has the pathological image quantity of far-end transfer risk is a, does not have far-end transfer risk
Pathological image quantity be b, obtain Nasopharyngeal Carcinoma Patients tumour cell occur far-end transfer scoring S=a/ (a+b), S value section
For 0-1；
Predictive information output module, is used for: establishing Nasopharyngeal Carcinoma Patients and occurs in far-end transfer risk and far-end transfer prediction module
The functional relation of gained scoring S, i.e. P=f (S), wherein P is that deep learning calculates each affiliated weight of clinical information of gained
The numerical value obtained；P value is calculated separately according to the scoring S of training data in deep learning algorithm training set, is not occurred to generation and
The P value of far-end transfer carries out section respectively and is classified as (Pma,Pmb) and (Pna,Pnb)；Wherein PmaIn training set to there is distal end to turn
Move the minimum P value of the image of feature, PmbFor the maximum P value of the image in training set with far-end transfer feature, PnaFor training set
In do not have far-end transfer feature image minimum P value, PnbNot have the image of far-end transfer feature in training set most
Big P value, and Pma>Pnb；If P ∈ (Pma,Pmb) 1 is then exported, representative has far-end transfer risk, if P ∈ (Pna,Pnb) 0 is then exported,
It represents without far-end transfer risk.
2. a kind of nasopharyngeal carcinoma far-end transfer forecasting system based on deep learning algorithm according to claim 1, feature
Be, further include image pre-processing module, be used for: to the tumor region delineated in Nasopharyngeal Carcinoma Patients pathological section digital picture into
Row cutting, and filtering background is carried out to the digital picture of cutting using gray threshold method or Otsu threshold method, eliminate noise
And dirty data；It is then normalized, obtains processed nasopharyngeal carcinoma tumor pathological section digital picture, be sent to distal end
Branch prediction module.
3. a kind of nasopharyngeal carcinoma far-end transfer forecasting system based on deep learning algorithm according to claim 2, feature
It is, further includes image capture module, be used for: (1) reads the nasopharyngeal carcinoma tumor pathological section digital picture of complete scan, delineate
Tumour cell region in image out, is sent to image pre-processing module；(2) the nasopharynx cancerous swelling of a large amount of known transinformations of storage
Training set of the tumor pathological section digital picture as deep learning algorithm, is supplied to far-end transfer prediction module.
4. a kind of nasopharyngeal carcinoma far-end transfer forecasting system based on deep learning algorithm according to claim 1, feature
It is, the far-end transfer prediction module, will after completing processed nasopharyngeal carcinoma tumor pathological section digital image analysis
The nasopharyngeal carcinoma tumor pathological section digital picture of processing returns to training set, and after whether confirmation far-end transfer occurs, will be described
Processed nasopharyngeal carcinoma tumor pathological section digital picture and its corresponding result are added to training set.
5. a kind of nasopharyngeal carcinoma far-end transfer forecasting system based on deep learning algorithm according to claim 1, feature
It is, further includes patient information collection module, be used for: the Detailed clinical data of Nasopharyngeal Carcinoma Patients is read out and filter screen
Choosing removes the incomplete case-data of clinical information, the patient clinical information screened.
6. a kind of nasopharyngeal carcinoma far-end transfer forecasting system based on deep learning algorithm according to claim 1, feature
It is, the far-end transfer prediction module contains the more of convolutional layer, pond layer, local articulamentum, full articulamentum and classifier
Hierarchical structure, the last layer are the structure of one two classification prediction, and mode of learning is used to learn for non-supervisory formula.
7. a kind of nasopharyngeal carcinoma far-end transfer forecasting system based on deep learning algorithm according to claim 1, feature
It is, the function of the far-end transfer prediction module includes: that tumor region picture TSI is extracted and made the training based on small slice
With validation data set；An improved deep learning model is established, and trains the Model Weight that can extract small slice feature；It is right
All slice features of one TSI figure are polymerize, and the feature of TSI figure is obtained；Cascade point is trained to the feature of whole TSI figure
Class device, judges whether TSI figure has transfer attribute；Various assessments are carried out to training pattern.
8. a kind of nasopharyngeal carcinoma far-end transfer forecasting system based on deep learning algorithm according to claim 7, feature
It is, the far-end transfer prediction module based on deep learning algorithm, various assessments is carried out to training pattern, i.e., to
The analysis of the nasopharyngeal carcinoma tumor pathological section digital picture of processing and classification performance are measured from various dimensions, mainly include precision
Area AUC under ACC, sensitivity S EN, specificity SPC and recipient's operating characteristics ROC curve；Wherein, SEN is indicated correctly predicted
Ratio of the pathological picture without transfer, SPC indicate that correctly predicted pathological picture has the ratio of transfer, these measure formulas define such as
Under: wherein TP represents the quantity of the pathological picture with transfer correctly judged, and TN represents the pathology without transfer correctly judged
The quantity of picture, FP represent the quantity of the pathological picture with transinformation of false judgment, and the nothing that FN represents false judgment turns
Move the quantity of the pathological picture of information；
ACC=(TP+TN)/(TP+FP+TN+FN)
SEN=TN/ (TN+FP)
SPC=TP/ (TP+FN)
AUC=(SEN+SPC)/2.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910407356.XA CN110097974A (en) | 2019-05-15 | 2019-05-15 | A kind of nasopharyngeal carcinoma far-end transfer forecasting system based on deep learning algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910407356.XA CN110097974A (en) | 2019-05-15 | 2019-05-15 | A kind of nasopharyngeal carcinoma far-end transfer forecasting system based on deep learning algorithm |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110097974A true CN110097974A (en) | 2019-08-06 |
Family
ID=67448332
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910407356.XA Pending CN110097974A (en) | 2019-05-15 | 2019-05-15 | A kind of nasopharyngeal carcinoma far-end transfer forecasting system based on deep learning algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110097974A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111445996A (en) * | 2020-03-31 | 2020-07-24 | 杭州依图医疗技术有限公司 | Medical information processing method, processing system and storage medium |
| CN111653365A (en) * | 2020-07-23 | 2020-09-11 | 中山大学附属第一医院 | Nasopharyngeal carcinoma auxiliary diagnosis model construction and auxiliary diagnosis method and system |
| CN112712875A (en) * | 2019-10-25 | 2021-04-27 | 中国医药大学附设医院 | Transfer learning aided prediction system, method and computer program product |
| TWI726459B (en) * | 2019-10-25 | 2021-05-01 | 中國醫藥大學附設醫院 | Transfer learning aided prediction system, method and computer program product thereof |
| CN112884724A (en) * | 2021-02-02 | 2021-06-01 | 广州智睿医疗科技有限公司 | Intelligent judgment method and system for lung cancer histopathological typing |
| CN113128599A (en) * | 2021-04-23 | 2021-07-16 | 南方医科大学南方医院 | Machine learning-based head and neck tumor distal metastasis prediction method |
| CN113393938A (en) * | 2021-07-13 | 2021-09-14 | 华南理工大学 | Breast cancer risk prediction system fusing image and clinical characteristic information |
| CN114093512A (en) * | 2021-10-21 | 2022-02-25 | 杭州电子科技大学 | Survival prediction method based on multi-mode data and deep learning model |
| CN114141306A (en) * | 2021-11-19 | 2022-03-04 | 天津大学 | Distant metastasis identification method based on gene interaction mode optimization graph representation |
| CN114305324A (en) * | 2020-09-28 | 2022-04-12 | 台北医学大学 | Tumor recurrence prediction device and method |
| TWI775161B (en) * | 2020-09-28 | 2022-08-21 | 臺北醫學大學 | Tumor recurrence prediction device and method |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107610129A (en) * | 2017-08-14 | 2018-01-19 | 四川大学 | A kind of multi-modal nasopharyngeal carcinima joint dividing method based on CNN |
| US20190114770A1 (en) * | 2017-10-13 | 2019-04-18 | Shenzhen Keya Medical Technology Corporation | Systems and methods for detecting cancer metastasis using a neural network |
| CN109686426A (en) * | 2018-12-29 | 2019-04-26 | 上海商汤智能科技有限公司 | Medical imaging processing method and processing device, electronic equipment and storage medium |
-
2019
- 2019-05-15 CN CN201910407356.XA patent/CN110097974A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107610129A (en) * | 2017-08-14 | 2018-01-19 | 四川大学 | A kind of multi-modal nasopharyngeal carcinima joint dividing method based on CNN |
| US20190114770A1 (en) * | 2017-10-13 | 2019-04-18 | Shenzhen Keya Medical Technology Corporation | Systems and methods for detecting cancer metastasis using a neural network |
| CN109686426A (en) * | 2018-12-29 | 2019-04-26 | 上海商汤智能科技有限公司 | Medical imaging processing method and processing device, electronic equipment and storage medium |
Non-Patent Citations (1)
| Title |
|---|
| 《数字图像处理 基于XAVIS组态软件》: "《数据挖掘的应用与实践 大数据时代的案例分析》", 西安交通大学出版社, pages: 313 - 316 * |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112712875A (en) * | 2019-10-25 | 2021-04-27 | 中国医药大学附设医院 | Transfer learning aided prediction system, method and computer program product |
| TWI726459B (en) * | 2019-10-25 | 2021-05-01 | 中國醫藥大學附設醫院 | Transfer learning aided prediction system, method and computer program product thereof |
| CN112712875B (en) * | 2019-10-25 | 2022-12-02 | 台湾中国医药大学附设医院 | Auxiliary prediction system and method for transfer learning |
| CN111445996A (en) * | 2020-03-31 | 2020-07-24 | 杭州依图医疗技术有限公司 | Medical information processing method, processing system and storage medium |
| CN111653365A (en) * | 2020-07-23 | 2020-09-11 | 中山大学附属第一医院 | Nasopharyngeal carcinoma auxiliary diagnosis model construction and auxiliary diagnosis method and system |
| CN114305324A (en) * | 2020-09-28 | 2022-04-12 | 台北医学大学 | Tumor recurrence prediction device and method |
| TWI775161B (en) * | 2020-09-28 | 2022-08-21 | 臺北醫學大學 | Tumor recurrence prediction device and method |
| CN112884724B (en) * | 2021-02-02 | 2022-06-03 | 广州智睿医疗科技有限公司 | Intelligent judgment method and system for lung cancer histopathological typing |
| CN112884724A (en) * | 2021-02-02 | 2021-06-01 | 广州智睿医疗科技有限公司 | Intelligent judgment method and system for lung cancer histopathological typing |
| CN113128599A (en) * | 2021-04-23 | 2021-07-16 | 南方医科大学南方医院 | Machine learning-based head and neck tumor distal metastasis prediction method |
| CN113393938A (en) * | 2021-07-13 | 2021-09-14 | 华南理工大学 | Breast cancer risk prediction system fusing image and clinical characteristic information |
| CN114093512A (en) * | 2021-10-21 | 2022-02-25 | 杭州电子科技大学 | Survival prediction method based on multi-mode data and deep learning model |
| CN114093512B (en) * | 2021-10-21 | 2023-04-18 | 杭州电子科技大学 | Survival prediction method based on multi-mode data and deep learning model |
| CN114141306A (en) * | 2021-11-19 | 2022-03-04 | 天津大学 | Distant metastasis identification method based on gene interaction mode optimization graph representation |
| CN114141306B (en) * | 2021-11-19 | 2023-04-07 | 天津大学 | Distant metastasis identification method based on gene interaction mode optimization graph representation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110097974A (en) | A kind of nasopharyngeal carcinoma far-end transfer forecasting system based on deep learning algorithm | |
| CN110111895A (en) | A kind of method for building up of nasopharyngeal carcinoma far-end transfer prediction model | |
| Xue et al. | An application of transfer learning and ensemble learning techniques for cervical histopathology image classification | |
| CN107977671A (en) | A kind of tongue picture sorting technique based on multitask convolutional neural networks | |
| CN111985536A (en) | Gastroscope pathological image classification method based on weak supervised learning | |
| CN106096654A (en) | A kind of cell atypia automatic grading method tactful based on degree of depth study and combination | |
| CN109410194A (en) | A kind of cancer of the esophagus pathology image processing method based on deep learning | |
| Tang et al. | Segnet-based gland segmentation from colon cancer histology images | |
| CN101551854B (en) | A processing system of unbalanced medical image and processing method thereof | |
| CN109671053A (en) | A kind of gastric cancer image identification system, device and its application | |
| CN112365464A (en) | GAN-based medical image lesion area weak supervision positioning method | |
| CN101609485B (en) | Medical image diagnosing system and diagnosing method based on migrated nuclear matching tracing | |
| CN110543912A (en) | Method for automatically acquiring cardiac cycle video in fetal key section ultrasonic video | |
| CN115953666B (en) | Substation site progress identification method based on improved Mask-RCNN | |
| Zhang et al. | Automatic detection of invasive ductal carcinoma based on the fusion of multi-scale residual convolutional neural network and SVM | |
| CN113657449A (en) | Traditional Chinese medicine tongue picture greasy classification method containing noise labeling data | |
| CN114782753A (en) | Lung cancer histopathology full-section classification method based on weak supervision learning and converter | |
| CN115471701A (en) | Lung adenocarcinoma histology subtype classification method based on deep learning and transfer learning | |
| CN117036288A (en) | Tumor subtype diagnosis method for full-slice pathological image | |
| Kumar et al. | Colon cancer classification of histopathological images using data augmentation | |
| CN114782948A (en) | Global interpretation method and system for cervical liquid-based cytology smear | |
| Li et al. | High-order correlation-guided slide-level histology retrieval with self-supervised hashing | |
| CN116228759B (en) | Computer-aided diagnosis system and apparatus for renal cell carcinoma type | |
| Jiang et al. | Classifying cervical histopathological whole slide images via deep multi-instance transfer learning | |
| CN116664932A (en) | Colorectal cancer pathological tissue image classification method based on active learning |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190806 |
|
| RJ01 | Rejection of invention patent application after publication |