CN106203327B - Lung tumor identification system and method based on convolutional neural networks - Google Patents

Lung tumor identification system and method based on convolutional neural networks Download PDF

Info

Publication number
CN106203327B
CN106203327B CN201610534805.3A CN201610534805A CN106203327B CN 106203327 B CN106203327 B CN 106203327B CN 201610534805 A CN201610534805 A CN 201610534805A CN 106203327 B CN106203327 B CN 106203327B
Authority
CN
China
Prior art keywords
image
temperature
image block
sampling
filter
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
Application number
CN201610534805.3A
Other languages
Chinese (zh)
Other versions
CN106203327A (en
Inventor
徐葳
冯迭乔
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tsinghua University
Original Assignee
Tsinghua University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Tsinghua University filed Critical Tsinghua University
Priority to CN201610534805.3A priority Critical patent/CN106203327B/en
Publication of CN106203327A publication Critical patent/CN106203327A/en
Application granted granted Critical
Publication of CN106203327B publication Critical patent/CN106203327B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
    • G06K9/00Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
    • G06K9/00885Biometric patterns not provided for under G06K9/00006, G06K9/00154, G06K9/00335, G06K9/00362, G06K9/00597; Biometric specific functions not specific to the kind of biometric
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
    • G06K9/00Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
    • G06K9/36Image preprocessing, i.e. processing the image information without deciding about the identity of the image
    • G06K9/46Extraction of features or characteristics of the image
    • G06K9/4604Detecting partial patterns, e.g. edges or contours, or configurations, e.g. loops, corners, strokes, intersections
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
    • G06K9/00Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
    • G06K9/36Image preprocessing, i.e. processing the image information without deciding about the identity of the image
    • G06K9/46Extraction of features or characteristics of the image
    • G06K9/4652Extraction of features or characteristics of the image related to colour
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
    • G06K9/00Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
    • G06K9/62Methods or arrangements for recognition using electronic means
    • G06K9/6267Classification techniques
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/0002Inspection of images, e.g. flaw detection
    • G06T7/0012Biomedical image inspection
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/10Image acquisition modality
    • G06T2207/10072Tomographic images
    • G06T2207/10081Computed x-ray tomography [CT]
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/30Subject of image; Context of image processing
    • G06T2207/30004Biomedical image processing
    • G06T2207/30061Lung
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/30Subject of image; Context of image processing
    • G06T2207/30004Biomedical image processing
    • G06T2207/30061Lung
    • G06T2207/30064Lung nodule

Abstract

The present invention provides a kind of lung tumor identification system and method based on convolutional neural networks.Wherein, the lung tumor identification system based on convolutional neural networks includes: several stacking images for receiving and acquiring along default dimension;According to the identical movement rule of the same window, each stacking image is subjected to piecemeal, and the image block in each stacking image of corresponding the same window position is merged into convolution, to obtain characteristic image block, and down-sampling is carried out to each characteristic image block;Each characteristic image block after down-sampling is up-sampled;And it may be probability that is true or being vacation that the characteristic image block after at least one up-sampling is corresponded to based on each pixel in temperature prognostic chart, draw temperature prognostic chart.The present invention effectively increases the discrimination of the regional area for the stacking image with spatial correlation.

Description

Lung tumor identification system and method based on convolutional neural networks
Technical field
The present invention relates to a kind of image processing techniques, know more particularly to a kind of lung tumors based on convolutional neural networks Other system and method.
Background technique
Convolutional neural networks presented its advantage in characteristics of image identification mission in recent years.For example, two Image is tieed up to during 3-D image building, bad block screening is carried out to two dimensional image block using convolutional neural networks.Current convolution Neural network is to belong to true or false in the way of deep learning to image and classify mostly.Full convolutional neural networks are made For a kind of structure slightly certain moduli type, receive to have made great progress in identification mission locally, including object boundary frame Detection, object key position and prediction of key point etc..These prediction modes have still continued to use the side learnt to classifier Formula classifies to image recognition to improve.The shortcomings that this kind of mode, is: since the part that full convolutional neural networks are identified is special Sign is in piece image, the relevant information for being associated with testing image of having no way of, so that recognition correct rate is limited by bottleneck.
Summary of the invention
In view of the foregoing deficiencies of prior art, the purpose of the present invention is to provide a kind of based on convolutional neural networks Lung tumor identification system and method, for solving in the prior art when carrying out part identification to piece image, identification is correct The low problem of rate.
In order to achieve the above objects and other related objects, the present invention provides a kind of temperature figure forecasting system, comprising: image connects Receiving apparatus, for receiving several stacking images along the acquisition of default dimension;Downsampling device based on convolutional network, for according to Each stacking image is carried out piecemeal, and each stacking of corresponding the same window position is schemed by the identical movement rule of the same window Image block as in carries out convolution and merges again, to obtain characteristic image block, and adopt to each characteristic image block Sample;Device is up-sampled, for each characteristic image block after down-sampling to be carried out up-sampling treatment, so that at least adjacent window apertures position Each characteristic image block after corresponding up-sampling is overlapped with pixel;Temperature prediction meanss, after using up-sampling At least one characteristic image block is probability that is true or being vacation to draw the color of each pixel in temperature prognostic chart;Wherein, institute It states each pixel in temperature prognostic chart and belongs at least one characteristic image block after up-sampling.
In some embodiments, the downsampling device based on convolutional network further include: normalization unit, being used for will Several described stacking images are normalized, mobile according still further to uniform window, several stackings described in after normalization are schemed As carrying out piecemeal.
In some embodiments, the downsampling device based on convolutional network includes: it is at least one set of by filter and The structure of downsampling unit composition;Wherein, when the quantity of the structure is multiple, each group structure cascades with one another;Wherein, with institute The filter for stating image received device connection has and the stacking consistent receiving channel of amount of images, the receiving channel transmitting one Image is laminated in width;The filter be used for by it is each stacking image in correspond to the same window position each image block progress convolution again into Row merges, and obtains the characteristic image block for corresponding to the window's position;The receiving channel of filter and previous stage filter in other groups Output channel cascade, for the received characteristic image block of institute to be carried out a full convolution;It is set in the output channel of every grade of filter There is the downsampling unit, for carrying out a down-sampling to the received characteristic image block of institute, and by the characteristic image block after down-sampling It is sent into next stage filter or up-samples the receiving channel of device.
In some embodiments, the filter be used for according to Each image block for corresponding to the same window position in each stacking image is carried out convolution to merge again;Wherein, WkFor k-th of convolution Core, bkFor the offset that corresponding k-th of convolution kernel adds on each channel, c is the port number that image is laminated, and (u, v) is WkRuler Very little, (i, j) is pixel position in image block.
In some embodiments, the filter comprises at least one of the following: filter based on color is based on lines Filter and filter based on gray scale.
In some embodiments, the up-sampling device is used for according to the window size, to the received characteristic pattern of institute As block carries out the up-sampling based on quadratic interpolation.
In some embodiments, the up-sampling device includes: up-sampling unit, for based on a received characteristic pattern As each pixel value in block, it is filled in preset expansion window, with the characteristic image block once up-sampled;Smoothing processing list Member, for carrying out convolution to the characteristic image block once up-sampled based on the preset and consistent convolution kernel of window size, Obtain corresponding to the characteristic image block of the secondary up-sampling of the window size.
In some embodiments, the temperature prediction meanss are used for based on through weighing obtained from preparatory learnt comprising true value The loss function of weight and falsity weight, the received each characteristic image block of institute corresponding to each pixel of temperature prognostic chart carry out one by one Temperature evaluation, and the temperature prognostic chart is drawn based on the evaluation of obtained temperature.
In some embodiments, the loss function is to be based on obtained from study in advance and include true value weight and vacation It is worth the intersection entropy function of weight.
In some embodiments, several described stacking image datas are continuous odd number image.
Based on above-mentioned purpose, the lung tumor identification system based on convolutional neural networks that the present invention also provides a kind of, comprising: Image is laminated for spatially several lung CTs of sequential storage in image library;As above any temperature figure forecasting system;It is described Temperature figure forecasting system according to received several lung CTs stacking image output corresponding several lung CTs stacking image Temperature prognostic chart;Control device, for judging whether the temperature figure forecasting system is extracted all lung CT layers stored Folded image, if it is not, then control the temperature figure forecasting system extracts several lung CTs stacking image according to spatial order by several times, directly Until determining that described image extraction element is extracted all lung CTs stacking image stored.
In some embodiments, span of the control device based on piece image extracts several CT stacking figures by several times Picture.
Based on above-mentioned purpose, the present invention also provides a kind of temperature figure prediction techniques, comprising: what reception was acquired along default dimension Several stacking images;Several described stacking images are sent into the downsampling device based on convolutional network, are based on convolution by described The downsampling device of network carries out piecemeal according to the identical movement rule of the same window, by each stacking image, and will be corresponding same Image block in each stacking image of one the window's position carries out convolution and merges again, to obtain characteristic image block, and to every A characteristic image block carries out down-sampling;Each characteristic image block after down-sampling is subjected to up-sampling treatment, so that at least adjacent windows Each characteristic image block corresponding to mouth position, after up-sampling is overlapped with pixel;It is special using at least one after up-sampling Sign image block is probability that is true or being vacation to draw the color of each pixel in temperature prognostic chart;Wherein, the temperature prediction Each pixel in figure belongs at least one characteristic image block after up-sampling.
In some embodiments, it is receiving after several stacking images that default dimension acquires, further includes: will be described Several stacking images are normalized;It is mobile according to uniform window, several stacking images described in after normalization are carried out Piecemeal.
In some embodiments, the downsampling device based on convolutional network include: it is at least one set of by filter and The structure of downsampling unit composition;Wherein, when the quantity of the structure is multiple, each group structure cascades with one another;Wherein, with institute The filter for stating image received device connection has and the stacking consistent receiving channel of amount of images, the receiving channel transmitting one Image is laminated in width;Each image block that the same window position is corresponded in each stacking image is carried out convolution and closed again by the filter And obtain corresponding to the characteristic image block of the window's position;The receiving channel of filter and previous stage filter is defeated in other groups Channel cascades out, and the received characteristic image block of institute is carried out full convolution;It is adopted under described in the output channel of every grade of filter Sample unit carries out down-sampling to the received characteristic image block of institute, and the characteristic image block after down-sampling is sent into next stage filtering Device carries out up-sampling operation.
In some embodiments, the filter will be corresponded in each stacking image each image block of the same window position into Row convolution merges again, obtains the characteristic image block for corresponding to the window's position, comprising: the filter according toEach image block of the same window position will be corresponded in each stacking image Convolution is carried out to merge again;Wherein, WkFor k-th of convolution kernel, bkAdd on each channel for corresponding k-th of convolution kernel inclined Shifting amount, c are the port number that image is laminated, and (u, v) is WkSize, (i, j) are pixel position in image block.
In some embodiments, the filter comprises at least one of the following: filter based on color is based on lines Filter and filter based on gray scale.
In some embodiments, each characteristic image block by after down-sampling carries out up-sampling treatment, comprising: according to The window size, to received characteristic image block carry out the up-sampling based on quadratic interpolation.
In some embodiments, described according to window size, the received characteristic image block of institute is carried out based on secondary-plug The up-sampling of value include: based on each pixel value in received characteristic image block, preset expansion window is filled in, to obtain The characteristic image block once up-sampled;Based on the preset and consistent convolution kernel of window size, to the spy once up-sampled It levies image block and carries out convolution, obtain the characteristic image block for the secondary up-sampling for corresponding to the window size.
In some embodiments, described at least one characteristic image block using after up-sampling is true or is false probability To draw the color of each pixel in temperature prognostic chart;Wherein, each pixel in the temperature prognostic chart belongs to up-sampling At least one characteristic image block afterwards, comprising: based on the damage for obtained from being learnt in advance including true value weight and falsity weight Function is lost, temperature evaluation one by one, and base are carried out to the received each characteristic image block of the corresponding institute of each pixel of temperature prognostic chart The temperature prognostic chart is drawn in the evaluation of obtained temperature.
In some embodiments, the loss function is to be based on obtained from study in advance and include true value weight and vacation It is worth the intersection entropy function of weight.
In some embodiments, several described stacking image datas are continuous odd number image.
Based on above-mentioned purpose, the lung tumor identification method based on convolutional neural networks that the present invention also provides a kind of, comprising: Spatially image is laminated in several lung CTs of sequential storage in advance;According to any of the above-described temperature figure prediction technique, according to institute The temperature prognostic chart of received corresponding several lung CTs stacking image of several lung CTs stacking image output;Judge the heat Whether degree figure forecasting system is extracted all lung CTs stacking image stored, if it is not, then controlling temperature figure prediction system System extracts several lung CTs stacking image according to spatial order by several times, is stored until determining that described image extraction element is extracted All lung CTs stacking image until.
In some embodiments, span of the control device based on piece image extracts several CT stacking figures by several times Picture.
As described above, the lung tumor identification system and method for the invention based on convolutional neural networks, has with following Beneficial effect: each characteristic image block that can indicate the common trait of several stacking images is extracted by convolutional neural networks, so Obtained include in each image dimension except place plane and plane characteristic information, and then by being adopted on each characteristic image block Sample and the prediction of temperature figure effectively increase the discrimination of the regional area for the stacking image with spatial correlation.Meanwhile Downsampling unit is set in convolutional neural networks, advantageously reduces the calculation amount of convolutional neural networks, effectively improves tumour knowledge Other efficiency.In addition, capableing of the color transition of smooth each image block in up-sampling by the way of quadratic interpolation, improving and know Other accuracy.In addition, image is continuously laminated using odd number, the characteristic information of retaining space dimension is not only contributed to, moreover it is possible to Guarantee the Balance Evaluation of image between forward and backward image pair.In addition, because loss function whether can only embody model training benign, It cannot function as last accurate result index.Later period through testing calculates and compares the position of each tumour and calculate False Positive and False Negative.The present invention is had descended on False Negative close to 5%, in False 20% is had decreased on Positive.It can be seen that compared with prior art, the present invention has been achieved for breakthrough.
Detailed description of the invention
Fig. 1 is shown as the structural schematic diagram of temperature figure forecasting system of the invention.
Fig. 2 is shown as the image in temperature figure forecasting system of the invention in each stacking image of the merged convolution of filter The schematic diagram of block.
Fig. 3 is shown as the structural schematic diagram of another temperature figure forecasting system of the invention.
Fig. 4 is shown as the structural schematic diagram of the lung tumor identification system of the invention based on convolutional neural networks.
Fig. 5 is shown as the flow chart of temperature figure prediction technique of the invention.
Fig. 6 is shown as the flow chart of the lung tumor identification method of the invention based on convolutional neural networks.
Fig. 7 is shown as the flow chart of another lung tumor identification method based on convolutional neural networks of the invention.
Fig. 8 is successively shown as from left to right described in lung CT stacking image grayscale figure, the tumour figure of handmarking and utilization Comparison diagram between the lung tumors temperature prognostic chart three that temperature figure forecasting system is identified.
Specific embodiment
Illustrate embodiments of the present invention below by way of specific specific example, those skilled in the art can be by this specification Other advantages and efficacy of the present invention can be easily understood for disclosed content.The present invention can also pass through in addition different specific realities The mode of applying is embodied or practiced, the various details in this specification can also based on different viewpoints and application, without departing from Various modifications or alterations are carried out under spirit of the invention.
Please refer to Fig. 1-Fig. 8.It should be noted that diagram provided in the present embodiment only illustrates this hair in a schematic way Bright basic conception, only shown in schema then with related component in the present invention rather than component count when according to actual implementation, Shape and size are drawn, when actual implementation kenel, quantity and the ratio of each component can arbitrarily change for one kind, and its component Being laid out kenel may also be increasingly complex.
Embodiment one
As shown in Figure 1, the present invention provides a kind of structural schematic diagram of temperature figure forecasting system.The temperature figure forecasting system Including the software and hardware being mounted in computer equipment.Wherein, the hardware in the computer equipment includes: input unit, Processing unit, storage unit, caching and display unit etc., wherein may include in the processing unit and be exclusively used in convolutional Neural The chip or integrated circuit of network and include convolutional neural networks algorithm computer program.The processing unit passes through journey The timing of sequence setting distributes the operation of each hardware, to execute the function of following each devices.Wherein, the computer equipment include but It is not limited to: single server, server cluster of multiple servers cooperation operation etc..
The temperature figure forecasting system 1 include: image received device 11, the downsampling device 12 based on convolutional network, on Sampling apparatus 13 and temperature prediction meanss 14.
Described image reception device 11 is used to receive several stacking images along the acquisition of default dimension.Here, described image Reception device 11 may include processing unit, caching and the interface being connected with the image library of storage stacking image.Described image Processing unit in reception device 11 is indicated according to the timing of program, is read from image library by interface in advance along default dimension Several stacking images of acquisition.Wherein, several described stacking images are from captured by the preset time dimension or Spatial Dimension , can overlap to embody and read in the image set of captured dimensional characteristics relevance.
For example, described image reception device 11 reads several continuous CT from the CT image library in CT image library Image is laminated, wherein the quantity of the CT stacking image is odd number.For the ease of subsequent several stacking images received to institute Temperature prediction is carried out, the size of several stacking images is consistent.
In a specific example, several described stacking images are, for example, h × 512 × 512 sizes lung CT figure Piece, wherein h is the scanning number of plies of CT.
The downsampling device 12 based on convolutional network is used for according to the same window and identical movement rule, will be each The stacking image carries out piecemeal, and the image block in each stacking image of corresponding the same window position is merged convolution, To obtain characteristic image block, and down-sampling is carried out to each characteristic image block.
Convolutional network in the downsampling device 12 based on convolutional network is convolutional neural networks (Convolutional Neural Network, abbreviation CNN), the convolutional neural networks include convolutional layer (alternating Convolutional layer) and pond layer (pooling layer).Wherein, the convolutional layer can be considered subsequent and will retouch in detail Each filter stated, the pond layer can be considered the subsequent downsampling unit that will be described in detail.
Here, the downsampling device 12 based on convolutional network includes: the processing list for being capable of handling volume machine network algorithm Member and matching caching.The processing unit can be shared with the processing unit in image received device 11, may also mean that packet Unit containing the chip or integrated circuit for being exclusively used in convolutional neural networks.If the downsampling device 12 based on convolutional network In include said chip or integrated circuit.It receives and fills to improve image processing speed, the chip or integrated circuit and image It sets and is connected between 11 by the image receiving channel of hardware.Provided by the parallel reception image received device 11 of each receiving channel Each stacking image.
Specifically, the downsampling device 12 based on convolutional network utilizes above-mentioned each hardware, composition comprising filter and The structure type of downsampling unit.Wherein, the quantity of the filter can be one, be also possible to multiple.Each filter and Downsampling unit can be integrated in the chip/integrated circuit.Described image reception device 11 is by each receiving channel and at least One filter is connected.Wherein, the filter comprises at least one of the following: filter based on color is based on lines Filter and filter based on gray scale.Wherein, the window size of the window is less than the picture size of stacking image.It is described Window is unified in the movement rule of each stacking image.Processing unit in the downsampling device 12 based on convolutional network Start first from the same starting pixels point (such as [0,0]) in each width stacking image, according to same movement rule, (such as span is 1 Movement rule), to each width stacking image carry out piecemeal.
For example, the processing unit is divided into image block a11, a12 ..., a1n, a21, a22 ... for image A is laminated, and amn;Stacking image B is divided into image block b11, b12 ..., b1n, b21, b22 ... and bmn;Stacking image C is divided into image Block c11, c12 ..., c1n, c21, c22 ... and cmn;Stacking image D is divided into image block d11, d12 ..., d1n, d21, B22 ... and dmn;Stacking image E is divided into image block e11, e12 ..., e1n, e21, e22 ... and emn.Wherein, m is figure As the line number of block, n is the columns of image block.
It should be noted that the window can be 1 pixel column/pixel according to span according to actual temperature forecast demand The piecemeal that the movement rule of column carries out blanket type to each width lamination image is handled, and can also be divided for preset image-region Block processing.It is existing or from now in technical solution if any based on the present embodiment enlighten and carry out partitioned mode of the present invention Within the scope of.Then, the image block of footmark same in each stacking image (i.e. footmark is all [m, n]) is sent by the processing unit The same filter merges convolution.Wherein, the merging convolution refer to by each image block with the convolution kernel in filter It carries out convolution and is merged into a characteristic image block, so that the feature letter that each pixel is reacted in filtered characteristic image block Breath being capable of the feature of each image block that is inputted of integrating representation.
For example, as shown in Fig. 2, filter receives the image block that footmark in each stacking image is all [m, n], and by each image Block carries out convolution with the convolution kernel in filter respectively, and takes and obtain to integrate the result after each pixel convolution Embody the pixel characteristic information of each image block characteristics.
Alternatively, the blocking process can be a part in filter.The downsampling device based on convolutional network Convolutional network in 12 is made of being stacked as several layers, and general next layer of output is the input of higher.Picture is by most Bottom input, top output is final result.The structure of more specifically neural network can be a digraph, be used to Complete some special tasks.The input of each filter layer is all three-dimensional array h, w, a d, and wherein h and w is input Size, d indicate the special of characteristic pattern or channel.More specifically, for inputting picture, h and w are exactly the height and width of picture, and d is Input picture color channel number (the Color Channel number of conventional RGB figure is 3, grayscale image 1).When each stacking image is according to signal Input and output access when entering filter, the convolution kernel size in filter is window size.In the present embodiment, according to complete The filter type of convolution comes out piecemeal according to window movement rule by filter in each stacking image input filter Each image block of corresponding the same window position merges convolution.
In a kind of optinal plan, the downsampling device 12 based on convolutional network includes: normalization unit, and being used for will Several described stacking images are normalized, according still further to the same window and identical movement rule, by the institute after normalization It states several stacking images and carries out piecemeal.
Specifically, the normalization unit normalizes to the pixel value of each stacking image between [0,1].It again will normalization Each stacking image input filter afterwards.
Be in alternative dispensing means, the filter be used for according to Each image block that the same window position is corresponded in each stacking image is merged into convolution;Wherein, WkFor k-th of convolution kernel, bkFor The offset that corresponding k-th of convolution kernel adds on each channel, c are the port number that image is laminated, and (u, v) is WkSize, (i, j) For pixel position in image block.
In order to preferably extract the characteristic information for capableing of each image block of integrating representation, the down-sampling based on convolutional network Filter in device 12 is multiple, and each filter cascades.As shown in Figure 3.
For example, filter F ilter11, Filter12 and Filter13 be respectively intended to extract the color of image block, lines and Gray feature.And filter F ilter11 is connected with image received device 11, to receive the stacking image of each receiving channel.Filtering Successively grade is associated in after Filter11 by device Filter12 and Filter13, and the characteristic image block exported to Filter11 carries out It is filtered for the correlation of lines and gray scale.
Then, the downsampling unit chooses the modes such as extreme value or average value according to from the grouping of preset pixel, right Each characteristic image block of filter output carries out down-sampling.Wherein, in every level structure the convolution kernel size of filter to be not more than The size that pixel is grouped in downsampling unit is preferred, i.e. 0≤u≤s;0≤v≤s.Wherein, (u, v) be convolution kernel size, (s, S) be in downsampling unit pixel be grouped size.
It can be found that the port number of data remains unchanged after the down-sampling layer for being s by a window size, length and width difference It is reduced to original 1/s, so that total data scale falls below original 1/s2.Down-sampling layer is also an advantage that, just Be if being added without down-sampling layer, convolutional neural networks can position to object it is especially sensitive, even the movement of very little amplitude The intermediate result of network can be allowed great deviation occur.Convolutional neural networks can better adapt to after down-sampling layer is added Being displaced bring influences, and keeping intermediate result is not in very big deviation.After data size continuously reduces, opened because calculating The reduction of pin, we suitably can increase the number of characteristic pattern in the deep layer of network, convolutional neural networks is enabled to learn to arrive More picture features.
In addition, in order to improve operation efficiency, each filter can be using multiple CPU (or GPU) in server to each window The image block of position it is parallel merge convolution.
For example, the downsampling unit is grouped according to 4 × 4 pixel, in the received characteristic image block of institute every 4 × 4 The characteristic value of pixel is averaged, and the average value is assigned to the corresponding pixel points after down-sampling in image block.
In order to improve the processing speed of subsequent up-sampling treatment, a kind of optinal plan is, described based under convolutional network The characteristic information of each image block is further concentrated in sampling apparatus 12.
Specifically, the downsampling device 12 based on convolutional network includes: multiple groups are by filter and downsampling unit group At structure, and each group structure cascades with one another.
Wherein, the filter connecting with described image reception device 11 has to be led to consistent receive of stacking amount of images Road, the receiving channel transmit a width and image are laminated;The filter is for will correspond to the same window position in each stacking image Each image block merge convolution, obtain the characteristic image block for corresponding to the window's position.The reception of filter is logical in other groups The output channel of road and previous stage filter cascades, for the received characteristic image block of institute to be carried out full convolution.It is filtered at every grade The output channel of device is equipped with the downsampling unit, for carrying out down-sampling to the received characteristic image block of institute, and adopts by under Characteristic image block after sample is sent into next stage filter or up-samples the receiving channel of device 13.
For example, in the downsampling device 12 based on convolutional network comprising 4 grades of above structure Structure1, Structure2, Structure3 and Structure4.It wherein, include multiple filters and a down-sampling list in every level structure Member, the filter in each structure are also cascade connection, what the last one filter that downsampling unit is organized where receiving was exported Characteristic image block, and carry out down-sampling processing.The filter being connected in structure Structure1 with described image reception device 11 Each image block for correspond to the same window position in each stacking image is merged into convolution, and by each characteristic pattern after merging convolution Next filter is sent into as block and carries out feature extraction, until the last one filter is by each characteristic pattern in structure Structure1 It is finished as block filters, each characteristic image block is sent into downsampling unit and carries out down-sampling.Structure Structure1 will be after down-sampling Each characteristic image block be sequentially sent to the filter in structure Structure2 and carry out feature extraction one by one, and in structure After the last one filter is by the filtering of each characteristic image block in Structure2, by the downsampling unit in the structure to each Characteristic image block carries out down-sampling again.When each stacking image is by the feature of the downsampling device 12 based on convolutional network It extracts with after down-sampling, the size of each characteristic image block remains original graph to be much smaller than the size of original picture block As the characteristic information of block.
Then, each characteristic image block is admitted to up-sampling device 13.The up-sampling device 13 is used for will be after down-sampling Each characteristic image block carries out up-sampling treatment, so that each characteristic image block at least after up-sampling corresponding to adjacent window apertures position It is overlapped with pixel.
Specifically, the up-sampling device 13 can be according to institute received characteristic image block and original picture block dimensionally Gap answers the value of each pixel according to the region in original picture block corresponding to each pixel in each characteristic image block It makes in corresponding region.
For example, the up-sampling device 13 utilizesTo be up-sampled.Wherein, the region Size be (s, s).As can be seen that each input pixel has been duplicated into the rectangle that a size is s*s in the output, in this way Up-sampling the space structures of former data is maintained while amplification data scale.
In order to make finally obtained temperature prognostic chart more accurately describe the common trait in each width stacking image, Yi Zhongke Selecting scheme is that the up-sampling device 13 carries out based on secondary-plug the received characteristic image block of institute according to the window size The up-sampling of value.
Specifically, the up-sampling device 13 includes: up-sampling unit and smooth processing unit.
Wherein, the up-sampling unit be used for based on each pixel value in received characteristic image block, be filled in default Expansion window, with the characteristic image block once up-sampled.
The up-sampling unit can utilize formula:By it is each in received characteristic image block Pixel is duplicate to be filled into the default corresponding region for expanding window.Wherein, the size for expanding window is less than aforementioned use In the window size for carrying out piecemeal to image.In view of i/s and j/s may be real number, this example is using four nearest pixels Point row quadratic interpolation goes to approach the numerical value of each pixel.
The smoothing processing unit is used for based on the preset and consistent convolution kernel of window size, to by primary up-sampling Characteristic image block carry out convolution, obtain the characteristic image block for the secondary up-sampling for corresponding to the window size.
Using the up-sampling mode of the up-sampling unit and smooth processing unit, intuitively it is being an advantage over image directly What block was upsampled to, but it is very nearly the same with the former in the effect of actual experiment.If the factor of behind is exactly up-sampling layer Followed by a convolutional layer and convolutional layer filter core size is greater than s, then above-mentioned secondary after this filter Interpolation can be obtained by the study of filter parameter.Therefore above-mentioned quadratic interpolation is included in the latter in some sense In convolutional layer, although not fully of equal value.The acceptance region of above-mentioned filter layer have it is apparent intermesh, and each mistake Filter parameter is all shared in layer, therefore the propagated forward based on GPU and backpropagation calculating can become very efficiently.
Here, in the smoothing processing unit can only comprising with the consistent convolution kernel of window size, can also be in the base of convolution Such as weighted average is carried out to each characteristic image block on plinth to handle.Each characteristic image block after the smoothing processing cell processing There can be 1 span according to the window's position sequence.
Here, the characteristic image block after up-sampling treatment can be considered each stacking image of reaction in the respective window location drawing As the temperature of block.
Then, the temperature prediction meanss 14 are for being true using at least one characteristic image block after up-sampling or being false Probability draw the color of each pixel in temperature prognostic chart;Wherein, each pixel in the temperature prognostic chart belongs to At least one characteristic image block after up-sampling.
Here, the temperature prediction meanss 14 can share the hardware such as processing unit, the caching in aforementioned each device, and receive Each characteristic image block that up-sampling device 13 is exported.
Specifically, the temperature prediction meanss 14 are consistent or inconsistent with true value using the received each characteristic image block of institute Probability evaluate the temperature information of each pixel in temperature prognostic chart.
In view of convolutional neural networks will judge to last input temperature figure y, therefore, we will be to output y definition One loss function.Loss function can be output and true valueBetween quadratic sum.For example, the loss function isThe loss function is also possible to be arranged based on cross entropy.For example, the loss function is
In practical applications, since true value is only possible to value 0 or 1, loss function takes cross entropy just in this case It can more show the degree of closeness of output and true value.A kind of optional way is, the temperature prediction meanss 14 are according to preset Loss function carries out temperature evaluation one by one to the received each characteristic image block of the corresponding institute of each pixel of temperature prognostic chart, and Based on the corresponding relationship of preset temperature evaluation and temperature color, the temperature prognostic chart is drawn.Wherein, the loss function is used In the probability for indicating that each characteristic image block in pixel place is consistent or inconsistent with true value in temperature prognostic chart.For example, described Loss function is characterized the mean square deviation between image block and true value, then the temperature prediction meanss 14 are successively right as follows The temperature of pixel is evaluated in temperature prognostic chart:
In characteristic image block yi、yi+1... and yi+tIn comprising the pixel O (a, b) in temperature prognostic chart, then it is described Temperature prediction meanss 14 utilize formulaCalculate the temperature of pixel O (a, b).
It should be noted that above-mentioned loss function is only for example rather than limitation of the present invention.In fact, loss function is also It can be cross entropy or unknown losses function.
In order to more accurately predict that the temperature of each stacking image, the temperature prediction meanss 14 are based on being learnt and being obtained in advance The loss function comprising true value weight and falsity weight arrived, the received each feature of institute corresponding to each pixel of temperature prognostic chart Image block carries out temperature evaluation one by one, and draws the temperature prognostic chart based on the evaluation of obtained temperature.
Wherein, it can be based on above-mentioned mean square deviation formula and improved letter through loss function obtained from preparatory machine learning Number.Preferred loss function is based on obtained from preparatory study and the friendship comprising true value weight and falsity weight in the present embodiment Pitch entropy function:
Wherein, black_ratio indicates through preparatory machine learning the true value weight of determination, and white_ratio indicates warp The falsity weight of preparatory machine learning and determination.
Using above-mentioned loss function, the temperature prediction meanss 14 in temperature prognostic chart each pixel temperature carry out by One temperature evaluation, and then according to the corresponding relationship of preset temperature evaluation and color, draw temperature prognostic chart.
Balanced, crossover entropy can well equalize the true value of the true value of black matrix and white background, so that neural network be made to exist Training is just easier to restrain towards globally optimal solution.
Embodiment two
The basis of each optinal plan in above-described embodiment one, the present embodiment provides a kind of based on convolutional neural networks Lung tumor identification system.Wherein, the lung tumor identification system based on convolutional neural networks applies the CT figure in hospital As in scanner or the image processing equipment (such as server) being connected with CT image scanning instrument.For CT image scanner For device, CT scanning device is included at least.The CT scanning device is by the position to be checked (such as lung, brain or whole body) of living body along sky Between axially carry out the scanning of cross section type, obtain corresponding several CT stacking image.It is corresponding, it is described to be based on convolutional neural networks Lung tumor identification system 3 include: image library 31, the temperature that any optinal plan is composed in such as above-described embodiment one Figure forecasting system 32 and control device 33.As shown in Figure 4.
In the present embodiment, temporarily it is with the lung tumor identification system temperature figure forecasting system 32 based on convolutional neural networks Example is illustrated, the tumour of the lung tumor identification system 3 based on convolutional neural networks lung for identification.
Wherein, the temperature figure forecasting system 32 and CT scanning device 2 share an image library 31.CT scanning device 2 is by institute The CT stacking image of scanning is stored in image library 31 according to spatial order.The temperature figure forecasting system 32 is in control device 33 Control under, read several CT therein and image be laminated, and according to any scheme described in embodiment one, carry out several CT The temperature prediction of image is laminated, and according to the operation of doctor, obtained temperature prognostic chart is supplied to doctor.
Wherein, the control device 33 be used to judge the temperature figure forecasting system 32 whether be extracted stored it is all Image is laminated in CT, if it is not, then control the temperature figure forecasting system 32 extracts several CT stacking image according to spatial order by several times, Until determining that described image extraction element is extracted all CT stacking image stored.
For example, image is laminated according to the sequence that image is respectively laminated in image library 31, with 5 width CT as one in the control device 33 Group, gradation provide several stacking images to temperature figure forecasting system 32.When the temperature figure forecasting system 32 is according to being received 5 width CT stacking image output temperature prognostic chart when, the control device 33 judges whether to be extracted all CT layers stored Folded image if otherwise continuing to extract next group of CT stacking image, and inputs temperature figure prediction system if so, exiting circulation System 32, until having extracted all CT stacking images.
Wherein, the temperature prognostic chart that the temperature figure forecasting system 32 is exported can be considered what this group of CT stacking image shared Temperature prognostic chart.Preferably, the temperature prognostic chart that the temperature figure forecasting system 32 is exported is considered as this group of CT stacking image Temperature prognostic chart corresponding to image is laminated in one width CT of centre.
In order to improve the temperature precision of prediction of CT stacking image, in a kind of optinal plan, CT layers of traversal is not completed in judgement In the case where folded image, image is laminated based on the CT that the span of piece image extracts different groups in the control device 33, so may be used To obtain the temperature prognostic chart of all CT stacking images in addition to head and the tail two width CT stacking image.
For example, the control device 33 is first by { x1,x2,x3,x4,x55 width CT stacking image feeding temperature figure forecasting system 32 carry out for image x3Temperature prediction.When judging x5When not last of image is laminated in CT in image library 31, the control Device 33 processed continues to select { x2,x3,x4,x5,x65 width CT stacking image is sent into temperature figure forecasting system 32 and carries out for image x4 Temperature prediction.It so repeats, until judging to exit circulation comprising last in extracted one group of CT stacking image.
Process as shown in Figure 7, the course of work citing of the lung tumor identification system 3 based on convolutional neural networks is such as Under:
After CT image scanning instrument obtains several CT stacking figures of patient, patient data and several CT stacking are schemed Correspondence is stored in image library 31.As scanned instruction of the control device 33 based on CT image scanning instrument or doctor When life has sent the instruction for obtaining temperature prognostic chart to 33 place equipment of control device by operating terminal, the control device 33 Image is laminated according to several CT of scanning sequency storage according to corresponding find of patient data, and informs in temperature figure forecasting system 32 Image acquiring apparatus extracts 1-5 CT and image is laminated.
Described image extraction element extracts corresponding several CT stacking image from image library 31 and utilizes corresponding each CT The receiving channel of stacking image passes to the downsampling device based on convolutional network.The downsampling device based on convolutional network In filter F ilter11 in the first level structure according to unified window size and window movement rule, image is laminated in each CT Each image block of middle the same window position merges convolution, so obtains the characteristic image block for corresponding to each the window's position.? Further include in primary structure and at least one cascade filter of Filter11 (Filter12 ..., Filter1n).In the first order In structure, the convolution kernel size of each filter can be consistent with above-mentioned window size, in order to reduce calculation amount and optimization convolution The structure of neural network.After the downsampling unit in the first level structure is located at filter F ilter1n, by by feature Image block is grouped the mode of down-sampling, and the characteristic image block of each the window's position of correspondence is carried out down-sampling.It again will be after down-sampling Characteristic image block is sent at least one filter (Filter21 ..., Filter2n) in the second level structure.In second level structure Each filter by the way that the received each characteristic image block of institute is carried out full process of convolution respectively, obtain corresponding to each the window's position into The characteristic image block of one step, wherein the size of the convolution kernel in the second level structure be less than received characteristic image block size. Again by the downsampling unit in the second level structure, the size of each characteristic image block is further reduced.By m cascade structure, Characteristic image block corresponding to each the window's position is substantially reduced.
Then, by up-sampling device in the way of quadratic interpolation, each characteristic image block after diminution is restored, with Obtain corresponding to the temperature image block of each the window's position.
Then, have the characteristics that overlapping region using each temperature image block by temperature prediction meanss, it will be in temperature prognostic chart The temperature evaluation of each pixel transfers to the temperature image block for covering the pixel to determine.Specifically, the temperature prediction dress It sets and uses the loss function based on obtained from being learnt in advance including true value weight and falsity weight, to each of temperature prognostic chart The received each characteristic image block (i.e. temperature image block) of the corresponding institute of pixel carries out temperature evaluation one by one, and is based on obtained The temperature prognostic chart is drawn in temperature evaluation, and the temperature prognostic chart drawn is one width of centre in corresponding 5 width CT stacking image The temperature prognostic chart of CT stacking image.
When the control fills determine that temperature figure forecasting system 32 completes that the 5 width CTs that are controlled are laminated in image about centre When the temperature prognostic chart of image is laminated in one width CT, judge whether completed 5 width CT stacking image has included corresponding same patient Last, if so, exit circulation, if it is not, 2-6 CT stacking images are then transferred to image received device, and again Temperature figure forecasting system 32 is executed, to obtain the temperature prognostic chart of corresponding 4th width CT stacking image.Such iterative cycles, it is described Lung tumor identification system based on convolutional neural networks can obtain remaining other than each two width CT stacking image of head and the tail The temperature prognostic chart of each CT stacking image.
Embodiment three
As shown in figure 5, the present invention provides a kind of flow chart of temperature figure prediction technique.The temperature figure prediction technique is main It is executed by forecasting system.Wherein the forecasting system includes the software and hardware being mounted in computer equipment.Wherein, described Hardware in computer equipment includes: input unit, processing unit, storage unit, caching and display unit etc., wherein described May include in processing unit the chip for being exclusively used in convolutional neural networks or integrated circuit and include convolutional neural networks calculate The computer program of method.The processing unit distributes the operation of each hardware by the timing of program setting, to execute following each dresses The function of setting.Wherein, the computer equipment includes but is not limited to: the service that single server, the cooperation of multiple servers are run Device cluster etc..
The temperature figure prediction technique realizes the prediction to temperature figure by executing following steps.
Step S110, several stacking images acquired along default dimension are received.Here, the forecasting system may include place Reason unit, caching and the interface being connected with the image library of storage stacking image.The forecasting system refers to according to the timing of program Show, reads several stacking images in advance along the acquisition of default dimension from image library by interface.Wherein, several described stacking figures Seem from captured by the preset time dimension or Spatial Dimension, can overlap to embody captured dimensional characteristics association It is read in the image set of property.
For example, the forecasting system reads continuous several CT stacking figure from the CT image library in CT image library Picture, wherein the quantity of the CT stacking image is odd number.Heat is carried out for the ease of subsequent several stacking images received to institute The size of degree prediction, several stacking images is consistent.
In a specific example, several described stacking images are, for example, h × 512 × 512 sizes lung CT figure Piece, wherein h is the scanning number of plies of CT.
Step S120, based on the downsampling device of convolutional network according to the same window and identical in the described forecasting system Each stacking image is carried out piecemeal by movement rule, and by the image block in each stacking image of corresponding the same window position Convolution is merged, to obtain characteristic image block, and down-sampling is carried out to each characteristic image block.
Convolutional network in the downsampling device based on convolutional network is convolutional neural networks (Convolutional Neural Network, abbreviation CNN), the convolutional neural networks include convolutional layer (alternating convolutional Layer) and pond layer (pooling layer).Wherein, the convolutional layer can be considered the subsequent each filter that will be described in detail, The pond layer can be considered the subsequent downsampling unit that will be described in detail.
Convolutional network in the downsampling device based on convolutional network is made of being stacked as several layers, under general One layer of output is the input of higher.Picture is inputted by the bottom, and top output is final result.It is more specifically refreshing Structure through network can be a digraph, for completing some special tasks.The input of each filter layer is all One three-dimensional array h, w, d, wherein h and w is the size of input, and d indicates the special of characteristic pattern or channel.More specifically, right In inputting picture, h and w are exactly the height and width of picture, and d is input picture color channel number (the Color Channel number of conventional RGB figure It is 3, grayscale image 1).
Here, the downsampling device based on convolutional network include: the processing unit for being capable of handling volume machine network algorithm, And matching caching.The processing unit can be shared with the processing unit in image received device, be may also mean that and included It is exclusively used in the chip of convolutional neural networks or the unit of integrated circuit.If including in the downsampling device based on convolutional network Said chip or integrated circuit.In order to improve image processing speed, between the chip or integrated circuit and image received device It is connected by the image receiving channel of hardware.Respectively stacking is schemed provided by the parallel reception image received device of each receiving channel Picture.
Specifically, the downsampling device based on convolutional network utilizes above-mentioned each hardware, and composition is comprising filter under The structure type of sampling unit.Wherein, the quantity of the filter can be one, be also possible to multiple.Each filter is under Sampling unit can be integrated in the chip/integrated circuit.Described image reception device passes through each receiving channel and at least one The filter is connected.Wherein, the filter comprises at least one of the following: the filter based on color, the mistake based on lines Filter and filter based on gray scale.Wherein, the window size of the window is less than the picture size of stacking image.The window It is unified in the movement rule of each stacking image.Processing unit in the downsampling device based on convolutional network first from Same starting pixels point (such as [0,0]) in each width stacking image starts, according to (the movement that such as span is 1 of same movement rule Rule), piecemeal is carried out to each width stacking image.
For example, the processing unit is divided into image block a11, a12 ..., a1n, a21, a22 ... for image A is laminated, and amn;Stacking image B is divided into image block b11, b12 ..., b1n, b21, b22 ... and bmn;Stacking image C is divided into image Block c11, c12 ..., c1n, c21, c22 ... and cmn;Stacking image D is divided into image block d11, d12 ..., d1n, d21, B22 ... and dmn;Stacking image E is divided into image block e11, e12 ..., e1n, e21, e22 ... and emn.Wherein, m is figure As the line number of block, n is the columns of image block.
It should be noted that the window can be 1 pixel column/pixel according to span according to actual temperature forecast demand The piecemeal that the movement rule of column carries out blanket type to each width lamination image is handled, and can also be divided for preset image-region Block processing.It is existing or from now in technical solution if any based on the present embodiment enlighten and carry out partitioned mode of the present invention Within the scope of.
Then, the image block of footmark same in each stacking image (i.e. footmark is all [m, n]) is sent by the processing unit The same filter merges convolution.Wherein, the merging convolution refer to by each image block with the convolution kernel in filter It carries out convolution and is merged into a characteristic image block, so that the feature letter that each pixel is reacted in filtered characteristic image block Breath being capable of the feature of each image block that is inputted of integrating representation.
For example, as shown in Fig. 2, filter receives the image block that footmark in each stacking image is all [m, n], and by each image Block carries out convolution with the convolution kernel in filter respectively, and takes and obtain to integrate the result after each pixel convolution Embody the pixel characteristic information of each image block characteristics.
Alternatively, the blocking process can be a part in filter.The downsampling device based on convolutional network Convolutional network in 12 is made of being stacked as several layers, and general next layer of output is the input of higher.Picture is by most Bottom input, top output is final result.The structure of more specifically neural network can be a digraph, be used to Complete some special tasks.The input of each filter layer is all three-dimensional array h, w, a d, and wherein h and w is input Size, d indicate the special of characteristic pattern or channel.More specifically, for inputting picture, h and w are exactly the height and width of picture, and d is Input picture color channel number (the Color Channel number of conventional RGB figure is 3, grayscale image 1).When each stacking image is according to signal Input and output access when entering filter, the convolution kernel size in filter is window size.In the present embodiment, according to complete The filter type of convolution comes out piecemeal according to window movement rule by filter in each stacking image input filter Each image block of corresponding the same window position merges convolution.
In a kind of optinal plan, the downsampling device based on convolutional network includes: normalization unit, is used for institute It states several stacking images to be normalized, according still further to the same window and identical movement rule, described in after normalization Several stacking images carry out piecemeal.
Specifically, the normalization unit normalizes to the pixel value of each stacking image between [0,1].It again will normalization Each stacking image input filter afterwards.
Be in alternative dispensing means, the filter be used for according to Each image block that the same window position is corresponded in each stacking image is merged into convolution;Wherein, WkFor k-th of convolution kernel, bkFor The offset that corresponding k-th of convolution kernel adds on each channel, c are the port number that image is laminated, and (u, v) is WkSize, (i, j) For pixel position in image block.
In order to preferably extract the characteristic information for capableing of each image block of integrating representation, the down-sampling based on convolutional network Filter in device is multiple, and each filter cascades.As shown in Figure 3.
For example, filter F ilter11, Filter12 and Filter13 be respectively intended to extract the color of image block, lines and Gray feature.And filter F ilter11 is connected with image received device, to receive the stacking image of each receiving channel.Filter Successively grade is associated in after Filter11 Filter12 and Filter13, and the characteristic image block exported to Filter11 carries out needle Correlation filtering to lines and gray scale.
Then, the downsampling unit chooses the modes such as extreme value or average value according to from the grouping of preset pixel, right Each characteristic image block of filter output carries out down-sampling.Wherein, in every level structure the convolution kernel size of filter to be not more than The size that pixel is grouped in downsampling unit is preferred, i.e. 0≤u≤s;0≤v≤s.Wherein, (u, v) be convolution kernel size, (s, S) be in downsampling unit pixel be grouped size.
It can be found that the port number of data remains unchanged after the down-sampling layer for being s by a window size, length and width difference It is reduced to original 1/s, so that total data scale falls below original 1/s2.Down-sampling layer is also an advantage that, just Be if being added without down-sampling layer, convolutional neural networks can position to object it is especially sensitive, even the movement of very little amplitude The intermediate result of network can be allowed great deviation occur.Convolutional neural networks can better adapt to after down-sampling layer is added Being displaced bring influences, and keeping intermediate result is not in very big deviation.After data size continuously reduces, opened because calculating The reduction of pin, we suitably can increase the number of characteristic pattern in the deep layer of network, convolutional neural networks is enabled to learn to arrive More picture features.
In addition, in order to improve operation efficiency, each filter can be using multiple CPU (or GPU) in server to each window The image block of position it is parallel merge convolution.
For example, the downsampling unit is grouped according to 4 × 4 pixel, in the received characteristic image block of institute every 4 × 4 The characteristic value of pixel is averaged, and the average value is assigned to the corresponding pixel points after down-sampling in image block.
In order to improve the processing speed of subsequent up-sampling treatment, a kind of optinal plan is, described based under convolutional network The characteristic information of each image block is further concentrated in sampling apparatus.
Specifically, the downsampling device based on convolutional network includes: multiple groups are made of filter and downsampling unit Structure, and each group structure cascades with one another.
Wherein, the filter connecting with described image reception device has and the consistent receiving channel of amount of images is laminated, The receiving channel transmits a width and image is laminated;The filter is for will correspond to each of the same window position in each stacking image Image block merges convolution, obtains the characteristic image block for corresponding to the window's position.In other groups the receiving channel of filter with The output channel of previous stage filter cascades, for the received characteristic image block of institute to be carried out full convolution.In every grade of filter Output channel is equipped with the downsampling unit, for carrying out down-sampling to the received characteristic image block of institute, and will be after down-sampling Characteristic image block be sent into next stage filter or up-sample device receiving channel.
For example, in the downsampling device based on convolutional network comprising 4 grades of above structure Structure1, Structure2, Structure3 and Structure4.It wherein, include multiple filters and a down-sampling list in every level structure Member, the filter in each structure are also cascade connection, what the last one filter that downsampling unit is organized where receiving was exported Characteristic image block, and carry out down-sampling processing.The filter being connected in structure Structure1 with described image reception device will Each image block for correspond to the same window position in each stacking image merges convolution, and by each characteristic image after merging convolution Block is sent into next filter and carries out feature extraction, until the last one filter is by each characteristic image in structure Structure1 Block filtering finishes, and each characteristic image block is sent into downsampling unit and carries out down-sampling.Structure Structure1 will be after down-sampling Each characteristic image block is sequentially sent to the filter in structure Structure2 and carries out feature extraction one by one, and in structure After the last one filter is by the filtering of each characteristic image block in Structure2, by the downsampling unit in the structure to each Characteristic image block carries out down-sampling again.
When each stacking image is after the feature extraction of the downsampling device based on convolutional network and down-sampling, each spy The size of image block is levied to be much smaller than the size of original picture block, but still remains the characteristic information of original picture block.
Step S130, each characteristic image block is admitted to the up-sampling device in the forecasting system.The up-sampling device Each characteristic image block after down-sampling is subjected to up-sampling treatment, so that at least after up-sampling corresponding to adjacent window apertures position Each characteristic image block is overlapped with pixel.
Specifically, the up-sampling device can according to received characteristic image block and the difference of original picture block dimensionally Away from according to the region in original picture block corresponding to each pixel in each characteristic image block, by the value duplication of each pixel Into corresponding region.
For example, the up-sampling device utilizesTo be up-sampled.Wherein, the region Having a size of (s, s).As can be seen that each input pixel has been duplicated into the rectangle that a size is s*s in the output, it is such Up-sampling maintains the space structure of former data while amplification data scale.
In order to make finally obtained temperature prognostic chart more accurately describe the common trait in each width stacking image, Yi Zhongke Selecting scheme is that the up-sampling device carries out based on quadratic interpolation the received characteristic image block of institute according to the window size Up-sampling.
Specifically, the step S130 includes: S131, S132 (being unillustrated).
Step S131, based on each pixel value in received characteristic image block, be filled in preset expansion window, with To the characteristic image block once up-sampled.
Specifically, formula is utilized:By each pixel weight in received characteristic image block Multiple is filled into the default corresponding region for expanding window.Wherein, the size for expanding window, which is less than, is previously used for image Carry out the window size of piecemeal.In view of i/s and j/s may be real number, this example is secondary using four nearest pixel rows Interpolation goes to approach the numerical value of each pixel.
Step S132, based on the preset and consistent convolution kernel of window size, to the characteristic image by once up-sampling Block carries out convolution, obtains the characteristic image block for the secondary up-sampling for corresponding to the window size.
Using the up-sampling mode of step S131 and S132, directly image block is upsampled to being intuitively an advantage over, But it is very nearly the same with the former in the effect of actual experiment.If the factor of behind is exactly that up-sampling layer is rolled up followed by one Lamination and convolutional layer filter core size are greater than s, then quadratic interpolation above-mentioned after this filter can pass through The study of filter parameter obtains.Therefore above-mentioned quadratic interpolation is included in some sense in the latter convolutional layer, though It is so not fully of equal value.The acceptance region of above-mentioned filter layer have it is apparent intermesh, and each filter parameter is in layer It is all inside shared, therefore the propagated forward based on GPU and backpropagation calculating can become very efficiently.
Here, can only comprising with the consistent convolution kernel of window size, can also be on the basis of convolution to each characteristic image block Carry out such as weighted average processing.Each characteristic image block after the smoothing processing cell processing can according to the window's position sequence With 1 span.
Here, the characteristic image block after up-sampling treatment can be considered each stacking image of reaction in the respective window location drawing As the temperature of block.
Then, it is drawn step S140, using at least one characteristic image block after up-sampling is probability that is true or being vacation The color of each pixel in temperature prognostic chart;Wherein, each pixel in the temperature prognostic chart belongs to after up-sampling extremely A few characteristic image block.
Specifically, the forecasting system using received each characteristic image block probability consistent or inconsistent with true value To evaluate the temperature information of each pixel in temperature prognostic chart.
In view of convolutional neural networks will judge to last input temperature figure y, therefore, we will be to output y definition One loss function.Loss function can be output and true valueBetween quadratic sum.For example, the loss function isThe loss function is also possible to be arranged based on cross entropy.For example, the loss function is
In practical applications, since true value is only possible to value 0 or 1, loss function takes cross entropy just in this case It can more show the degree of closeness of output and true value.
A kind of optional way is that the forecasting system is according to preset loss function, to each pixel of temperature prognostic chart The corresponding received each characteristic image block of institute carries out temperature evaluation one by one, and pair based on preset temperature evaluation and temperature color It should be related to, draw the temperature prognostic chart.Wherein, the loss function is for indicating each spy where pixel in temperature prognostic chart Levy the image block probability consistent or inconsistent with true value.For example, the loss function is characterized between image block and true value Mean square deviation, then the temperature prediction meanss as follows successively comment the temperature of in temperature prognostic chart pixel Valence:
In characteristic image block yi、yi+1... and yi+tIn comprising the pixel O (a, b) in temperature prognostic chart, then it is described Forecasting system utilizes formulaCalculate the temperature of pixel O (a, b).
It should be noted that above-mentioned loss function is only for example rather than limitation of the present invention.In fact, loss function is also It can be cross entropy or unknown losses function.
In order to more accurately predict that the temperature of each stacking image, the forecasting system are based on packet obtained from being learnt in advance The loss function of weight containing true value and falsity weight, the received each characteristic image block of institute corresponding to each pixel of temperature prognostic chart Temperature evaluation one by one is carried out, and the temperature prognostic chart is drawn based on the evaluation of obtained temperature.
Wherein, it can be based on above-mentioned mean square deviation formula and improved letter through loss function obtained from preparatory machine learning Number.Preferred loss function is based on obtained from preparatory study and the friendship comprising true value weight and falsity weight in the present embodiment Pitch entropy function:
Wherein, black_ratio indicates through preparatory machine learning the true value weight of determination, and white_ratio indicates warp The falsity weight of preparatory machine learning and determination.
Using above-mentioned loss function, the forecasting system carries out heat one by one to the temperature of each pixel in temperature prognostic chart Degree evaluation, and then according to the corresponding relationship of preset temperature evaluation and color, draw temperature prognostic chart.
Balanced, crossover entropy can well equalize the true value of the true value of black matrix and white background, so that neural network be made to exist Training is just easier to restrain towards globally optimal solution.
Example IV
The basis of each optinal plan in above-described embodiment three, as shown in fig. 6, the present embodiment provides one kind to be based on convolution The lung tumor identification method of neural network.Wherein, the lung tumor identification method based on convolutional neural networks mainly by It is executed based on the lung tumor identification system of convolutional neural networks.The lung tumor identification system based on convolutional neural networks System applies CT image scanning instrument or the image processing equipment (such as server) that is connected with CT image scanning instrument in hospital In.For CT image scanning instrument, CT scanning device is included at least.The CT scanning device is by the position to be checked of living body (such as lung, brain or whole body) axially carries out the scanning of cross section type along space, obtains corresponding several CT stacking image.
In the present embodiment, the swollen of lung is identified using the lung tumor identification method based on convolutional neural networks Tumor.Heat described in each optinal plan in embodiment three is used in the lung tumor identification method based on convolutional neural networks Spend figure prediction technique.That is, being predicted in the lung tumor identification system based on convolutional neural networks comprising temperature figure System.
Wherein, the temperature figure forecasting system and CT scanning device share an image library.
Specifically, the CT scanned stacking image is stored in image library by CT scanning device according to spatial order.
Then, the lung tumor identification system based on convolutional neural networks starts temperature according to the control rule Figure forecasting system executes step S220: reading several CT therein and image is laminated, and according to any described in embodiment three Scheme carries out the temperature prediction of several CT stacking image, and according to the operation of doctor, obtained temperature prognostic chart is supplied to Doctor.Detailed process is shown in Fig. 7.Wherein, Fig. 8 is that lung CT is laminated figure grayscale image, manual identified lung tumors position and utilizes this The schematic diagram of the lung tumors position of embodiment identification.
Followed by the lung tumor identification system based on convolutional neural networks executes step according to the control rule Rapid S230: judging whether the temperature figure forecasting system is extracted all CT stacking image stored, if it is not, then described in control Temperature figure forecasting system extracts several CT stacking image according to spatial order by several times, until determining that described image extraction element extracts Until all CT stacking image stored.
For example, the lung tumor identification system based on convolutional neural networks is according to being respectively laminated the suitable of image in image library Sequence, with 5 width CT stacking image for one group, since being laminated image in the first width CT, grouping provides CT to temperature figure forecasting system Image is laminated.When the temperature figure forecasting system 5 width CT stacking image output temperature prognostic chart received according to institute, the control Device processed judges whether to be extracted all CT stacking image stored, if so, circulation is exited, if otherwise continuing to extract next Image is laminated in group CT, and inputs the temperature figure forecasting system, until having extracted all CT stacking images.
Wherein, the temperature prognostic chart that the temperature figure forecasting system is exported can be considered the shared heat of this group of CT stacking image Spend prognostic chart.Preferably, the temperature prognostic chart that the temperature figure forecasting system is exported is considered as in this group of CT stacking image Temperature prognostic chart corresponding to image is laminated in one width CT of centre.
In order to improve the temperature precision of prediction of CT stacking image, in a kind of optinal plan, CT layers of traversal is not completed in judgement In the case where folded image, the lung tumor identification system based on convolutional neural networks is in advance with the span of piece image to CT Stacking image is grouped, the temperature prediction of so available all CT stacking images in addition to head and the tail two width CT stacking image Figure.
For example, first by { x1,x2,x3,x4,x55 width CT stacking image is sent into temperature figure forecasting system and carries out for image x3 Temperature prediction.When judging x5When not last of image is laminated in CT in image library, the control device continues to select { x2, x3,x4,x5,x65 width CT stacking image is sent into temperature figure forecasting system and carries out for image x4Temperature prediction.It so repeats, directly To judging that in extracted one group of CT stacking image include last, circulation is exited.
As shown in fig. 7, the course of work of the lung tumor identification system based on convolutional neural networks is exemplified below:
After CT image scanning instrument obtains several CT stacking figures of patient, patient data and several CT stacking are schemed Correspondence is stored in image library.When the lung tumor identification system based on convolutional neural networks is based on CT image scanning instrument Scanned instruction or doctor the finger for obtaining temperature prognostic chart is had sent to equipment where control device by operating terminal When enabling, image is laminated according to several CT of scanning sequency storage according to corresponding find of patient data in the control device, and informs heat The image acquiring apparatus spent in figure forecasting system extracts 1-5 CT and image is laminated.
Image acquiring apparatus in the temperature figure forecasting system extracts corresponding several CT stacking image from image library And the downsampling device based on convolutional network is passed to using the receiving channel of each CT stacking image of correspondence.It is described to be based on convolution Filter F ilter11 in the downsampling device of network in the first level structure is according to unified window size and the mobile rule of window Then, each CT each image block that the same window position in image is laminated is merged into convolution, so obtains corresponding to each the window's position Characteristic image block.In the first level structure further include at least one cascade filter of Filter11 (Filter12 ..., Filter1n).In the first level structure, the convolution kernel size of each filter can be consistent with above-mentioned window size, in order to subtract Few calculation amount and the structure for optimizing convolutional neural networks.Downsampling unit in the first level structure is located at filter F ilter1n Later, in such a way that characteristic image block is grouped down-sampling, the characteristic image block of each the window's position of correspondence adopt Sample.Again by the characteristic image block after down-sampling be sent into the second level structure at least one filter (Filter21 ..., Filter2n).Each filter in second level structure by the way that the received each characteristic image block of institute is carried out full process of convolution respectively, Obtain corresponding to the further feature image block of each the window's position, wherein the size of the convolution kernel in the second level structure is less than institute The size of received characteristic image block.Again by the downsampling unit in the second level structure, each characteristic image block is further reduced Size.By multiple cascade structure, characteristic image block corresponding to each the window's position is substantially reduced.
Then, by the up-sampling device in the temperature figure forecasting system in the way of quadratic interpolation, after diminution Each characteristic image block is restored, to obtain corresponding to the temperature image block of each the window's position.
Then, there is overlapping region using each temperature image block by the temperature prediction meanss in the temperature figure forecasting system The characteristics of, transfer to the temperature image block for covering the pixel to determine the temperature evaluation of pixel each in temperature prognostic chart. Specifically, the temperature prediction meanss are used based on the loss for obtained from being learnt in advance including true value weight and falsity weight Function, the heat of each characteristic image block (i.e. temperature image block) progress received to the corresponding institute of each pixel of temperature prognostic chart one by one Degree evaluation, and the temperature prognostic chart is drawn based on the evaluation of obtained temperature, the temperature prognostic chart drawn is corresponding 5 width CT The temperature prognostic chart of one width CT of the centre stacking image in image is laminated.Fig. 8 is followed successively by lung CT stacking image ash from left to right Degree figure, the tumour figure of handmarking and the lung tumors temperature prognostic chart three that is identified using the temperature figure forecasting system Between comparison diagram.
When the control device determines that temperature figure forecasting system completes that the 5 width CT controlled are laminated in image about centre When the temperature prognostic chart of image is laminated in one width CT, judge whether completed 5 width CT stacking image has included corresponding same patient Last if it is not, 2-6 CT stacking images are then transferred to image received device, and predict if so, exit circulation The temperature prognostic chart of 4th CT stacking image.Such iterative cycles, the lung tumor identification system based on convolutional neural networks System can obtain the temperature prognostic chart of remaining each CT stacking image other than each two width CT stacking image of head and the tail.
In conclusion various embodiments of the present invention are extracted by convolutional neural networks can indicate the common of several stacking images Each characteristic image block of feature has so obtained the characteristic information including dimension except place plane and plane in each image, into And by predicting each characteristic image block up-sampling and temperature figure, it effectively increases for the stacking image with spatial correlation Regional area discrimination.Meanwhile downsampling unit is set in convolutional neural networks, advantageously reduce convolutional neural networks Calculation amount, effectively improve tumour identification efficiency.
In addition, capableing of the color transition of smooth each image block in up-sampling by the way of quadratic interpolation, improving identification Accuracy.
In addition, image is continuously laminated using odd number, the characteristic information of retaining space dimension is not only contributed to, moreover it is possible to protect Demonstrate,prove the Balance Evaluation of image between forward and backward image pair.
In addition, because loss function whether can only embody model training benign, cannot function as last accurate result index. Later period through testing calculates and compares the position of each tumour and calculate False Positive and False Negative.This Invention is had descended on False Negative close to 5%, has decreased to 20% on False Positive.Thus may be used See, compared with prior art, the present invention has been achieved for breakthrough.
So the present invention effectively overcomes various shortcoming in the prior art and has high industrial utilization value.
The above-described embodiments merely illustrate the principles and effects of the present invention, and is not intended to limit the present invention.It is any ripe The personage for knowing this technology all without departing from the spirit and scope of the present invention, carries out modifications and changes to above-described embodiment.Cause This, institute is complete without departing from the spirit and technical ideas disclosed in the present invention by those of ordinary skill in the art such as At all equivalent modifications or change, should be covered by the claims of the present invention.

Claims (22)

1. a kind of temperature figure forecasting system characterized by comprising
Image received device, for receiving several stacking images along the acquisition of default dimension;
Downsampling device based on convolutional network, for according to the same window and identical movement rule, each stacking to be schemed As carrying out piecemeal processing, and the image block for corresponding to the same window position in each stacking image is subjected to convolution and is merged again, obtained Down-sampling is carried out to characteristic image block, and to obtained each characteristic image block;
Device is up-sampled, for up-sampling each characteristic image block after down-sampling;
Temperature prediction meanss, for drawing heat using at least one characteristic image block after up-sampling is probability that is true or being vacation Spend the color of each pixel in prognostic chart;Wherein, each pixel in the temperature prognostic chart belongs to after up-sampling at least One characteristic image block.
2. temperature figure forecasting system according to claim 1, which is characterized in that the down-sampling dress based on convolutional network It sets further include:
Normalization unit, for several described stacking images to be normalized, according still further to the same window and identical shifting Several stacking images described in after normalization are carried out piecemeals by dynamic rule.
3. temperature figure forecasting system according to claim 1 or 2, which is characterized in that described based on being adopted under convolutional network Sampling device includes: the structure that at least one set is made of filter and downsampling unit;Wherein, the number of filters in the structure For at least one, when the number of filters in the structure is multiple, each filter in same group of structure is cascaded;When described When the quantity of structure is multiple, each group structure cascades with one another;
Wherein, the filter connecting with described image reception device has and the consistent receiving channel of amount of images is laminated, described Receiving channel transmits a width and image is laminated;The filter is used to that each image of the same window position will to be corresponded in each stacking image Block carries out convolution and merges again, obtains the characteristic image block for corresponding to the window's position;
The receiving channel of filter and the output channel of previous stage filter cascade in other groups, are used for the received characteristic pattern of institute As block carries out full convolution;
The downsampling unit is equipped in the output channel of every grade of filter, for carrying out down to the received characteristic image block of institute Sampling, and the characteristic image block after down-sampling is sent into next stage filter or up-samples the receiving channel of device.
4. temperature figure forecasting system according to claim 3, which is characterized in that the mistake being connect with image received device Filter be used for according toThe same window position will be corresponded in each stacking image Each image block set carries out convolution and merges again;Wherein, WkFor k-th of convolution kernel, bkIt is corresponding k-th of convolution kernel each The offset added on channel, c are the port number that image is laminated, and (u, v) is WkSize, (i, j) are pixel position in image block.
5. temperature figure forecasting system according to claim 3, which is characterized in that the filter in each structure include with down toward Few one kind: the filter based on color, the filter based on lines and the filter based on gray scale.
6. temperature figure forecasting system according to claim 1, which is characterized in that the up-sampling device is used for according to described Window size, to received characteristic image block carry out the up-sampling based on quadratic interpolation.
7. temperature figure forecasting system according to claim 6, which is characterized in that the up-sampling device includes:
Up-sampling unit, for based on each pixel value in received characteristic image block, be filled in preset expansion window, obtain To the characteristic image block once up-sampled;
Smoothing processing unit, for being based on the preset and consistent convolution kernel of window size, to the feature once up-sampled Image block carries out convolution, obtains the characteristic image block for the secondary up-sampling for corresponding to the window size.
8. temperature figure forecasting system according to claim 1, which is characterized in that the temperature prediction meanss are used for based on warp In advance include the loss function of true value weight and falsity weight obtained from study, institute is corresponded to each pixel of temperature prognostic chart Received each characteristic image block carries out temperature evaluation one by one, and draws the temperature prediction based on the evaluation of obtained temperature Figure.
9. temperature figure forecasting system according to claim 8, which is characterized in that the loss function is based on study in advance Obtained from and include true value weight and falsity weight intersection entropy function.
10. temperature figure forecasting system according to claim 1, which is characterized in that several described stacking image datas are to connect Continuous odd number image.
11. a kind of lung tumor identification system based on convolutional neural networks characterized by comprising
Image is laminated for spatially several lung CTs of sequential storage in image library;
Temperature figure forecasting system as described in any in claim 1-10;
The temperature figure forecasting system is according to received several the corresponding described lung CT layers of several lung CTs stacking image output of institute The temperature prognostic chart of folded image;
Image is laminated for judging whether the temperature figure forecasting system is extracted all lung CTs stored in control device, If it is not, then control the temperature figure forecasting system extracts several lung CTs stacking image according to spatial order by several times, until determining Until described image extraction element is extracted all lung CTs stacking image stored.
12. the lung tumor identification system according to claim 11 based on convolutional neural networks, which is characterized in that described Span of the control device based on piece image extracts several CT stacking image by several times.
13. a kind of temperature figure prediction technique characterized by comprising
Receive several stacking images acquired along default dimension;
Several described stacking images are sent into downsampling devices based on convolutional network, by described based on being adopted under convolutional network Sampling device carries out piecemeal according to the same window and identical movement rule, by each stacking image, and will be in each stacking image The image block of corresponding the same window position carries out convolution and merges again, to obtain characteristic image block, and to each characteristic pattern As block carries out down-sampling;
Each characteristic image block after down-sampling is subjected to up-sampling treatment;
Each picture in temperature prognostic chart is drawn using at least one characteristic image block after up-sampling is probability that is true or being vacation The color of vegetarian refreshments;Wherein, each pixel in the temperature prognostic chart belongs at least one characteristic image block after up-sampling.
14. temperature figure prediction technique according to claim 13, which is characterized in that receive acquire along default dimension it is more Width is laminated after image, further includes:
Several described stacking images are normalized;
According to the same window and identical movement rule, several stacking images described in after normalization are subjected to piecemeals.
15. temperature figure prediction technique described in 3 or 14 according to claim 1, which is characterized in that described based under convolutional network Sampling apparatus includes: at least one set of structure being made of filter and downsampling unit;Wherein, the filter number in the structure Amount is at least one, each filter cascade when the number of filters in the structure is multiple, in same group of structure;Work as institute State structure quantity be it is multiple when, each group structure cascades with one another;
Wherein, the filter connecting with described image reception device has and the consistent receiving channel of amount of images is laminated, described Receiving channel transmits a width and image is laminated;The filter will be corresponded in each stacking image each image block of the same window position into Row convolution merges again, obtains the characteristic image block for corresponding to the window's position;
The receiving channel of filter and the output channel of previous stage filter cascade in other groups, by the received characteristic image block of institute Carry out full convolution;
The downsampling unit in the output channel of every grade of filter carries out down-sampling to the received characteristic image block of institute, And the characteristic image block after down-sampling is sent into next stage filter or carries out up-sampling operation.
16. temperature figure prediction technique according to claim 15, which is characterized in that described to be connect with image received device Each image block that the same window position is corresponded in each stacking image is carried out convolution and merged again by filter, obtains corresponding to the window The characteristic image block of mouthful position, comprising: the filter according toIt will Each image block that the same window position is corresponded in each stacking image carries out convolution and merges again;Wherein, WkFor k-th of convolution Core, bkFor the offset that corresponding k-th of convolution kernel adds on each channel, c is the port number that image is laminated, and (u, v) is WkRuler Very little, (i, j) is pixel position in image block.
17. temperature figure prediction technique according to claim 15, which is characterized in that the filter in each structure includes following It is at least one: the filter based on color and/or the filter based on lines and the filter based on gray scale.
18. temperature figure prediction technique according to claim 13, which is characterized in that each characteristic pattern by after down-sampling As block carries out up-sampling treatment, comprising: according to the window size, carry out the received characteristic image block of institute based on quadratic interpolation Up-sampling.
19. temperature figure prediction technique according to claim 18, which is characterized in that it is described according to window size, to being connect The characteristic image block of receipts carries out the up-sampling based on quadratic interpolation
Based on each pixel value in received characteristic image block, preset expansion window is filled in, once to be up-sampled Characteristic image block;
The characteristic image block progress convolution once up-sampled is obtained with the consistent convolution kernel of window size based on preset To the characteristic image block of the secondary up-sampling of the correspondence window size.
20. temperature figure prediction technique according to claim 13, which is characterized in that at least one using after up-sampling A characteristic image block is probability that is true or being vacation to draw the color of each pixel in temperature prognostic chart, comprising: based on through pre- Include the loss function of true value weight and falsity weight obtained from first learning, each pixel of temperature prognostic chart is corresponded to and is adopted Each characteristic image block after sample carries out temperature evaluation one by one, and draws the temperature prediction based on the evaluation of obtained temperature Figure.
21. temperature figure prediction technique according to claim 20, which is characterized in that the loss function is based in advance Obtained from habit and include true value weight and falsity weight intersection entropy function.
22. temperature figure prediction technique according to claim 13, which is characterized in that several described stacking image datas are to connect Continuous odd number image.
CN201610534805.3A 2016-07-08 2016-07-08 Lung tumor identification system and method based on convolutional neural networks Active CN106203327B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201610534805.3A CN106203327B (en) 2016-07-08 2016-07-08 Lung tumor identification system and method based on convolutional neural networks

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201610534805.3A CN106203327B (en) 2016-07-08 2016-07-08 Lung tumor identification system and method based on convolutional neural networks

Publications (2)

Publication Number Publication Date
CN106203327A CN106203327A (en) 2016-12-07
CN106203327B true CN106203327B (en) 2019-04-19

Family

ID=57472713

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201610534805.3A Active CN106203327B (en) 2016-07-08 2016-07-08 Lung tumor identification system and method based on convolutional neural networks

Country Status (1)

Country Link
CN (1) CN106203327B (en)

Families Citing this family (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106778867B (en) * 2016-12-15 2020-07-07 北京旷视科技有限公司 Target detection method and device, and neural network training method and device
CN106778785B (en) * 2016-12-23 2019-09-17 东软集团股份有限公司 Construct the method for image Feature Selection Model and the method, apparatus of image recognition
CN106920227B (en) * 2016-12-27 2019-06-07 北京工业大学 The Segmentation Method of Retinal Blood Vessels combined based on deep learning with conventional method
CN107016665A (en) * 2017-02-16 2017-08-04 浙江大学 A kind of CT pulmonary nodule detection methods based on depth convolutional neural networks
CN107169485B (en) * 2017-03-28 2020-10-09 北京捷通华声科技股份有限公司 Mathematical formula identification method and device
CN107145756A (en) * 2017-05-17 2017-09-08 上海辉明软件有限公司 A kind of stroke types Forecasting Methodology and device
CN107292312A (en) * 2017-06-19 2017-10-24 中国科学院苏州生物医学工程技术研究所 Tumour recognition methods
CN107437110B (en) * 2017-07-11 2021-04-02 中国科学院自动化研究所 Block convolution optimization method and device of convolutional neural network
CN107292884B (en) * 2017-08-07 2020-09-29 杭州深睿博联科技有限公司 Method and device for identifying edema and hematoma in MRI (magnetic resonance imaging) image
CN108022238B (en) * 2017-08-09 2020-07-03 深圳科亚医疗科技有限公司 Method, computer storage medium, and system for detecting object in 3D image
CN108446724B (en) * 2018-03-12 2020-06-16 江苏中天科技软件技术有限公司 Fusion feature classification method
CN109285157A (en) * 2018-07-24 2019-01-29 深圳先进技术研究院 Myocardium of left ventricle dividing method, device and computer readable storage medium
CN109285158A (en) * 2018-07-24 2019-01-29 深圳先进技术研究院 Vascular wall patch dividing method, device and computer readable storage medium
CN109447969A (en) * 2018-10-29 2019-03-08 北京青燕祥云科技有限公司 Hepatic space occupying lesion recognition methods, device and realization device
CN109800873B (en) * 2019-01-29 2021-03-23 北京旷视科技有限公司 Image processing method and device

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103544506A (en) * 2013-10-12 2014-01-29 Tcl集团股份有限公司 Method and device for classifying images on basis of convolutional neural network
CN104346622A (en) * 2013-07-31 2015-02-11 富士通株式会社 Convolutional neural network classifier, and classifying method and training method thereof
CN104850836A (en) * 2015-05-15 2015-08-19 浙江大学 Automatic insect image identification method based on depth convolutional neural network
CN104933695A (en) * 2014-03-19 2015-09-23 联想(北京)有限公司 Information processing method and electronic equipment

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9400925B2 (en) * 2013-11-15 2016-07-26 Facebook, Inc. Pose-aligned networks for deep attribute modeling

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104346622A (en) * 2013-07-31 2015-02-11 富士通株式会社 Convolutional neural network classifier, and classifying method and training method thereof
CN103544506A (en) * 2013-10-12 2014-01-29 Tcl集团股份有限公司 Method and device for classifying images on basis of convolutional neural network
CN104933695A (en) * 2014-03-19 2015-09-23 联想(北京)有限公司 Information processing method and electronic equipment
CN104850836A (en) * 2015-05-15 2015-08-19 浙江大学 Automatic insect image identification method based on depth convolutional neural network

Also Published As

Publication number Publication date
CN106203327A (en) 2016-12-07

Similar Documents

Publication Publication Date Title
CN106940816B (en) CT image pulmonary nodule detection system based on 3D full convolution neural network
Melekhov et al. Relative camera pose estimation using convolutional neural networks
CN107220618B (en) Face detection method and device, computer readable storage medium and equipment
Bayramoglu et al. Deep learning for magnification independent breast cancer histopathology image classification
Yang et al. Building extraction at scale using convolutional neural network: Mapping of the united states
US20200175352A1 (en) Structure defect detection using machine learning algorithms
CN104751178B (en) Lung neoplasm detection means and method based on shape template matching combining classification device
CN104809426B (en) Training method, target identification method and the device of convolutional neural networks
Nhat-Duc et al. Automatic recognition of asphalt pavement cracks using metaheuristic optimized edge detection algorithms and convolution neural network
CN106570477B (en) Vehicle cab recognition model building method and model recognizing method based on deep learning
RU2720440C1 (en) Image segmentation method using neural network
CN104537647B (en) A kind of object detection method and device
CN106250870B (en) A kind of pedestrian's recognition methods again of joint part and global similarity measurement study
Utans et al. Selecting neural network architectures via the prediction risk: Application to corporate bond rating prediction
CN107016665A (en) A kind of CT pulmonary nodule detection methods based on depth convolutional neural networks
CN106529447B (en) Method for identifying face of thumbnail
CN107220980A (en) A kind of MRI image brain tumor automatic division method based on full convolutional network
WO2017215622A1 (en) Object segmentation method and apparatus and computing device
CN106504233A (en) Image electric power widget recognition methodss and system are patrolled and examined based on the unmanned plane of Faster R CNN
CN103593670B (en) A kind of copper plate/strip detection method of surface flaw based on online limit of sequence learning machine
CN106780612B (en) Object detecting method and device in a kind of image
CN103116763B (en) A kind of living body faces detection method based on hsv color Spatial Statistical Character
CN105046213B (en) A kind of method of augmented reality
CN104424634B (en) Object tracking method and device
CN105869173B (en) A kind of stereoscopic vision conspicuousness detection method

Legal Events

Date Code Title Description
PB01 Publication
C06 Publication
SE01 Entry into force of request for substantive examination
C10 Entry into substantive examination
GR01 Patent grant
GR01 Patent grant