CN106203327A - 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
CN106203327A
CN106203327A CN201610534805.3A CN201610534805A CN106203327A CN 106203327 A CN106203327 A CN 106203327A CN 201610534805 A CN201610534805 A CN 201610534805A CN 106203327 A CN106203327 A CN 106203327A
Authority
CN
China
Prior art keywords
image
stacking
temperature
image block
sampling
Prior art date
Application number
CN201610534805.3A
Other languages
Chinese (zh)
Other versions
CN106203327B (en
Inventor
徐葳
冯迭乔
Original Assignee
清华大学
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 清华大学 filed Critical 清华大学
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, lung tumor identification system based on convolutional neural networks includes: receive along presetting several stacking images that dimension gathers;According to the identical movement rule of the same window, each stacking image is carried 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 each characteristic image block is carried out down-sampling;Each characteristic image block after down-sampling is up-sampled;And may be true based on the characteristic image block after at least one up-sampling corresponding of each pixel in temperature prognostic chart or be false probability, draw temperature prognostic chart.The present invention is effectively increased 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, particularly relate to a kind of lung tumors based on convolutional neural networks and know Other system and method.
Background technology
Convolutional neural networks had presented its advantage in characteristics of image identification mission in the last few years.Such as, two Dimension image, during 3-D view builds, utilizes convolutional neural networks that two dimensional image block carries out bad block screening.Current convolution Neutral net is to utilize degree of depth learning style mostly, image is belonged to true or false and classifies.Full convolutional neural networks is made For a kind of structure somewhat certain moduli type, it accepts to have made great progress in identification mission, including object boundary frame in local Detection, object key position and the prediction etc. of key point.These prediction mode have still continued to use the side learning grader Formula, improves and classifies image recognition.The shortcoming of this kind of mode is: the local identified due to full convolutional neural networks is special Levy and be in piece image, have no way of associating the relevant information of testing image so that recognition correct rate is limited by bottleneck.
Summary of the invention
The shortcoming of prior art in view of the above, it is an object of the invention to provide a kind of based on convolutional neural networks Lung tumor identification system and method, is used for solving in prior art when piece image carrying out local and identifying, identification is correct The problem that rate is low.
For achieving the above object and other relevant purposes, the present invention provides a kind of image temperature prognoses system, including: image Receive device, for receiving along presetting several stacking images that dimension gathers;Downsampling device based on convolutional network, is used for pressing According to the identical movement rule of the same window, each described stacking image is carried out piecemeal, and by each stacking of corresponding the same window position Image block in image merges convolution, to obtain characteristic image block, and each characteristic image block is carried out down-sampling;On Sampling apparatus, for carrying out up-sampling process by each characteristic image block after down-sampling so that at least adjacent window apertures position institute is right It is overlapping that each characteristic image block after the up-sampling answered has pixel;Temperature prediction means, after utilizing up-sampling at least One characteristic image block is true or is that false probability is to draw the color of each pixel in temperature prognostic chart;Wherein, described heat Each pixel in degree prognostic chart belongs at least one the characteristic image block after up-sampling.
In some embodiments, described downsampling device based on convolutional network also includes: normalization unit, and being used for will Several stacking images described are normalized, and move according still further to uniform window, by several stacking figures described in after normalization As carrying out piecemeal.
In some embodiments, described downsampling device based on convolutional network comprises: least one set by filter and The structure of downsampling unit composition;Wherein, when the quantity of described structure is multiple, each group structure cascades with one another;Wherein, with institute The filter stating image received device connection has the reception passage consistent with stacking amount of images, described reception passage transmission one Width stacking image;Described filter is for merging volume by each image block of the same window position corresponding in each stacking image Long-pending, obtain should the characteristic image block of the window's position;In other groups, filter receives the defeated of passage and previous stage filter Go out passage cascade, for the characteristic image block received is carried out full convolution;The output channel of every grade of filter is provided with institute Stating downsampling unit, for the characteristic image block received being carried out down-sampling, and the characteristic image block after down-sampling being sent into Next stage filter or the reception passage of up-sampling device.
In some embodiments, described filter for according to Each image block of the same window position corresponding in each stacking image is merged convolution;Wherein, WkFor kth convolution kernel, bkFor The side-play amount that corresponding kth convolution kernel adds on each passage, c is the port number of stacking image, and (u v) is WkSize, (i, j) For pixel position in image block.
In some embodiments, described filter include following at least one: filter based on color, based on lines Filter and filter based on gray scale.
In some embodiments, described up-sampling device is for according to described window size, to the characteristic pattern received As block carries out up-sampling based on quadratic interpolation.
In some embodiments, described up-sampling device includes: up-sampling unit, for based on the characteristic pattern received As each pixel value in block, it is filled in default expansion window, with the characteristic image block once up-sampled;Smoothing processing list Unit, for based on the default convolution kernel consistent with window size, carries out convolution to the described characteristic image block once up-sampled, Obtain the characteristic image block of the secondary up-sampling of corresponding described window size.
In some embodiments, described temperature prediction means is for weighing based on the true value that comprises obtained through learning in advance Weight and the loss function of falsity weight, carried out one by one the corresponding each characteristic image block received of each pixel of temperature prognostic chart Temperature evaluation, and draw described temperature prognostic chart based on obtained temperature evaluation.
In some embodiments, described loss function obtains based on learning in advance and comprises true value weight and vacation The cross entropy function of value weight.
In some embodiments, several stacking view data described are continuous print odd number image.
Based on above-mentioned purpose, the present invention also provides for a kind of lung tumor identification system based on convolutional neural networks, including: Image library, for spatially several lung CT stacking images of sequential storage;Arbitrary described image temperature prognoses system;Institute State image temperature prognoses system according to corresponding several lung CT stacking figures described of several lung CT stacking images output received The temperature prognostic chart of picture;Control device, for judging whether described image temperature prognoses system is extracted all lungs stored Portion CT stacking image, if it is not, then control described image temperature prognoses system to extract several lung CT layers by several times according to spatial order Folded image, till determining that described image acquiring apparatus is extracted all lung CT stacking images stored.
In some embodiments, described control device span based on piece image extracts several CT stacking figures by several times Picture.
Based on above-mentioned purpose, the present invention also provides for a kind of image temperature Forecasting Methodology, including: receive along presetting dimension collection Several stacking images;Described many good fortune stacking image is sent in downsampling device based on convolutional network, by described based on volume Each described stacking image, according to the identical movement rule of the same window, is carried out piecemeal by the downsampling device of long-pending network, and by correspondence Image block in each stacking image of the same window position merges convolution, to obtain characteristic image block, and to each spy Levy image block and carry out down-sampling;Each characteristic image block after down-sampling is carried out up-sampling process so that at least adjacent window apertures position Put each characteristic image block corresponding, after up-sampling and there is pixel overlap;Utilize at least one characteristic pattern after up-sampling It is true as block or is that false probability is to draw the color of each pixel in temperature prognostic chart;Wherein, in described temperature prognostic chart Each pixel belong at least one the characteristic image block after up-sampling.
In some embodiments, after receiving several stacking images gathered along default dimension, also include: by described Several stacking images are normalized;Move according to uniform window, several stacking images described in after normalization are carried out Piecemeal.
In some embodiments, described downsampling device based on convolutional network includes: least one set by filter and The structure of downsampling unit composition;Wherein, when the quantity of described structure is multiple, each group structure cascades with one another;Wherein, with institute The filter stating image received device connection has the reception passage consistent with stacking amount of images, described reception passage transmission one Width stacking image;Each image block of the same window position corresponding in each stacking image is merged convolution by described filter, To should the characteristic image block of the window's position;The output channel receiving passage and previous stage filter of filter in other groups Cascade, carries out full convolution by the characteristic image block received;Described downsampling unit in the output channel of every grade of filter, The characteristic image block received is carried out down-sampling, and the characteristic image block after down-sampling is sent into next stage filter or carries out Up-sampling operation.
In some embodiments, each image block of the same window position corresponding in each stacking image is entered by described filter Row merges convolution, obtain to should the characteristic image block of the window's position, including: described filter according toEach image block by the same window position corresponding in each stacking image Merge convolution;Wherein, WkFor kth convolution kernel, bkThe side-play amount added on each passage for corresponding kth convolution kernel, c For the port number of stacking image, (u v) is WkSize, (i j) is pixel position in image block.
In some embodiments, described filter include following at least one: filter based on color, based on lines Filter and filter based on gray scale.
In some embodiments, described each characteristic image block after down-sampling is carried out up-sampling process, including: according to Described window size, carries out up-sampling based on quadratic interpolation to the characteristic image block received.
In some embodiments, described according to window size, the characteristic image block received is carried out based on secondary-plug The up-sampling of value includes: based on each pixel value in the characteristic image block received, be filled in default expansion window, to obtain The characteristic image block once up-sampled;Based on the default convolution kernel consistent with window size, to the described spy once up-sampled Levy image block and carry out convolution, obtain the characteristic image block of the secondary up-sampling of corresponding described window size.
In some embodiments, described at least one characteristic image block utilized after up-sampling is true or is false probability Draw the color of each pixel in temperature prognostic chart;Wherein, each pixel in described temperature prognostic chart belongs to up-sampling After at least one characteristic image block, including: based on through learn in advance and obtain comprise true value weight and the damage of falsity weight Lose function, the corresponding each characteristic image block received of each pixel of temperature prognostic chart is carried out temperature evaluation one by one, and base Described temperature prognostic chart is drawn in obtained temperature evaluation.
In some embodiments, described loss function obtains based on learning in advance and comprises true value weight and vacation The cross entropy function of value weight.
In some embodiments, several stacking view data described are continuous print odd number image.
Based on above-mentioned purpose, the present invention also provides for a kind of lung tumor identification method based on convolutional neural networks, including: The most spatially several lung CT stacking images of sequential storage;According to any of the above-described described image temperature Forecasting Methodology, according to The temperature prognostic chart of corresponding several lung CT stacking images described of several lung CT stacking images output received;Judge described Whether image temperature prognoses system is extracted all lung CT stacking images stored, if it is not, then control described image temperature Prognoses system extracts several lung CT stacking images, until determining that described image acquiring apparatus is extracted by several times according to spatial order Till all lung CT stacking images stored.
In some embodiments, described control device span based on piece image extracts several CT stacking figures by several times Picture.
As it has been described above, the lung tumor identification system and method based on convolutional neural networks of the present invention, having following has Benefit effect: extracted each characteristic image block of the common trait that can represent several stacking images by convolutional neural networks, so Obtain each image and include the characteristic information of dimension outside place plane and plane, and then by adopting on each characteristic image block Sample and the prediction of temperature figure, be effectively increased the discrimination of regional area for the stacking image with spatial correlation.Meanwhile, Downsampling unit is set in convolutional neural networks, advantageously reduces the amount of calculation of convolutional neural networks, be effectively improved tumor and know Other efficiency.It addition, use the mode of quadratic interpolation when up-sampling, it is possible to the color transition of smooth each image block, improve and know Other accuracy.It addition, use odd number continuous print stacking image, not only contribute to the characteristic information of retaining space dimension, moreover it is possible to Ensure the Balance Evaluation of image between forward and backward image pair.It addition, it is the most optimum to embody model training because of loss function, Cannot function as last accurate result index.Calculate through the later stage of experiment and contrast the position of each tumor and calculate False Positive and False Negative.The present invention has descended to close to 5% on False Negative, at False 20% is have decreased on Positive.As can be seen here, compared with prior art, the present invention has been achieved for breakthrough.
Accompanying drawing explanation
Fig. 1 is shown as the structural representation of the image temperature prognoses system of the present invention.
Fig. 2 is shown as the figure in the image temperature prognoses system of the present invention in each stacking image of the merged convolution of filter Schematic diagram as block.
Fig. 3 is shown as the structural representation of the another kind of image temperature prognoses system of the present invention.
Fig. 4 is shown as the structural representation of the lung tumor identification system based on convolutional neural networks of the present invention.
Fig. 5 is shown as the flow chart of the image temperature Forecasting Methodology of the present invention.
Fig. 6 is shown as the flow chart of the lung tumor identification method based on convolutional neural networks of the present invention.
Fig. 7 is shown as the flow chart of another kind lung tumor identification based on the convolutional neural networks method of the present invention.
Fig. 8 is shown as lung CT stacking gradation of image figure, the tumor figure of handmarking the most successively and utilizes described Comparison diagram between the lung tumors temperature prognostic chart three that image temperature prognoses system is identified.
Detailed description of the invention
Below by way of specific instantiation, embodiments of the present invention being described, those skilled in the art can be by this specification Disclosed content understands other advantages and effect of the present invention easily.The present invention can also be by the most different concrete realities The mode of executing is carried out or applies, the every details in this specification can also based on different viewpoints and application, without departing from Various modification or change is carried out under the spirit of the present invention.
Refer to Fig. 1-Fig. 8.It should be noted that the diagram provided in the present embodiment illustrates this most in a schematic way Bright basic conception, the most graphic in component count time only display with relevant assembly in the present invention rather than is implemented according to reality, Shape and size are drawn, and during its actual enforcement, the kenel of each assembly, quantity and ratio can be a kind of random change, and its assembly Layout kenel is likely to increasingly complex.
Embodiment one
As it is shown in figure 1, the present invention provides the structural representation of a kind of image temperature prognoses system.Described image temperature is predicted System includes the software and hardware being arranged in computer equipment.Wherein, the hardware in described computer equipment comprises: input is single Unit, processing unit, memory element, caching and display unit etc., wherein, described processing unit can comprise and be exclusively used in convolution The chip of neutral net or integrated circuit and include the computer program of convolutional neural networks algorithm.Described processing unit leads to The sequential crossing program setting distributes the operation of each hardware, to perform the function of following each device.Wherein, described computer equipment bag Include but be not limited to: single server, multiple server coordinate the server cluster etc. run.
Described image temperature prognoses system 1 includes: image received device 11, downsampling device based on convolutional network 12, Up-sampling device 13 and temperature prediction means 14.
Described image received device 11 is for receiving along presetting several stacking images that dimension gathers.Here, described image Receive device 11 and can comprise the interface that processing unit, caching and the image library with storage stacking image are connected.Described image The processing unit received in device 11 indicates according to the sequential of program, is read in advance along presetting dimension by interface from image library Several stacking images gathered.Wherein, several stacking images described are from along captured by the time dimension preset or Spatial Dimension , can overlap and to embody the image set of captured dimensional characteristics relatedness read.
Such as, in CT image library, described image received device 11 reads several CT of continuous print from described CT image library Stacking image, wherein, the quantity of described CT stacking image is odd number.For the ease of follow-up several stacking images to being received Carry out temperature prediction, the consistent size of several stacking images described.
In a concrete example, the lung CT figure of described several stacking images for example, h × 512 × 512 size Sheet, wherein h is the scanning number of plies of CT.
Described downsampling device based on convolutional network 12 is for according to the same window and identical movement rule, by each Described 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 each characteristic image block is carried out down-sampling.
Convolutional network in described downsampling device based on convolutional network 12 is convolutional neural networks (Convolutional Neural Network is called for short CNN), described convolutional neural networks includes convolutional layer (alternating Convolutional layer) and pond layer (pooling layer).Wherein, described convolutional layer can be considered follow-up and will retouch in detail The each filter stated, described pond layer can be considered the follow-up downsampling unit that will describe in detail.
Here, described downsampling device based on convolutional network 12 includes: the process list of volume machine network algorithm can be processed Unit and the caching matched.This processing unit can share with the processing unit in image received device 11, it is also possible to refers to bag Containing being exclusively used in the chip of convolutional neural networks or the unit of integrated circuit.If described downsampling device based on convolutional network 12 In comprise said chip or integrated circuit.In order to improve image processing speed, described chip or integrated circuit fill with image-receptive Put and be connected by the image-receptive passage of hardware between 11.Each reception image received device 11 receiving channel parallel is provided Each stacking image.
Specifically, described downsampling device 12 based on convolutional network utilizes above-mentioned each hardware, composition comprise filter and The version of downsampling unit.Wherein, the quantity of described filter can be one, it is also possible to is multiple.Each filter and Downsampling unit can be integrated in described chip/integrated circuit.Described image received device 11 by each reception passage with at least One described filter is connected.Wherein, described filter include following at least one: filter based on color, based on lines Filter and filter based on gray scale.Wherein, the window size of described window is less than the picture size of stacking image.Described Window is unified at the movement rule of each stacking image.Processing unit in described downsampling device based on convolutional network 12 First the same starting pixels point (such as [0,0]) from each width stacking image starts, according to same movement rule (if span is 1 Movement rule), each width stacking image is carried out piecemeal.
Such as, stacking image A is divided into image block a11, a12 by described processing unit ..., a1n, a21, a22 ..., 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 according to actual temperature forecast demand, described window can be 1 pixel column/pixel according to span The movement rule of row carries out the piecemeal of blanket type and processes each width stacked graph picture, it is also possible to carry out point for default image-region Block processes.Existing or from now in technical scheme if any the partitioned mode carried out based on the present embodiment enlightenment all of the present invention Within the scope of.Then, the image block of same footmark in each stacking image (i.e. footmark is all [m, n]) is sent into by described processing unit Same filter merges convolution.Wherein, described merging convolution refer to by each image block all with the convolution kernel in filter Carry out convolution and be merged into a characteristic image block, so that the feature letter that in the characteristic image block after Guo Lving, each pixel is reacted Breath can the feature of each image block that inputted of integrating representation.
Such as, as in figure 2 it is shown, filter receives footmark in each stacking image is all the image block of [m, n], and by each image Block carries out convolution respectively with the convolution kernel in filter, and the result after each pixel convolution is taken and, obtaining can be comprehensive Embody the pixel characteristic information of each image block characteristics.
Or, described blocking process can be the part in filter.Described downsampling device based on convolutional network Convolutional network in 12 is folded by some layer stack and is formed, and the output of next general layer is the input of higher.Picture is by Bottom inputs, and top output is final result.The more specifically structure of neutral net can be a directed graph, is used for Complete some special tasks.The input of each filter layer is all three-dimensional array h, w, d, and wherein h and w is input Size, d represents the special of characteristic pattern or passage.More specifically, for input picture, h and w is 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, and gray-scale map is 1).When each stacking image is according to signal Input and output path enter filter time, the convolution kernel size in filter is window size.In the present embodiment, according to entirely The filter type of convolution, by each stacking image input filter, by filter according to window movement rule to piecemeal out Each image block of corresponding the same window position merges convolution.
In a kind of alternative, described downsampling device 12 based on convolutional network includes: normalization unit, and being used for will Several stacking images described are normalized, according still further to the same window and identical movement rule, by the institute after normalization State several stacking images and carry out piecemeal.
Specifically, the pixel value of each stacking image is normalized between [0,1] by described normalization unit.Again by normalization After each stacking image input filter.
In alternative dispensing means be, described filter for according to Each image block of the same window position corresponding in each stacking image is merged convolution;Wherein, WkFor kth convolution kernel, bkFor The side-play amount that corresponding kth convolution kernel adds on each passage, c is the port number of stacking image, and (u v) is WkSize, (i, j) For pixel position in image block.
In order to preferably extract can the characteristic information of each image block of integrating representation, described down-sampling based on convolutional network Filter in device 12 is multiple, and the cascade of each filter.As shown in Figure 3.
Such as, 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 reception passage.Filter Device Filter12 and Filter13 is after level is associated in Filter11 successively, and the characteristic image block being exported Filter11 is carried out Dependency for lines and gray scale filters.
Then, described downsampling unit chooses the mode such as extreme value or meansigma methods according to from default pixel packet, right Each characteristic image block of filter output carries out down-sampling.Wherein, in every level structure the convolution kernel size of filter be not more than In downsampling unit, the size of pixel packet is preferred, i.e. 0≤u≤s;0≤v≤s.Wherein, (u, v) is convolution kernel size, (s, S) it is the size of pixel packet in downsampling unit.
It appeared that after by window size being the down-sampling layer of s, the port number of data keeps constant, and length and width are respectively Taper to original 1/s, so make total data scale fall below original 1/s2.Down-sampling layer it is also an advantage that, just Being if being added without down-sampling layer, convolutional neural networks can be especially sensitive to the position of object, even movement the most by a small margin There is great deviation in the intermediate object program that can allow network.After adding down-sampling layer, convolutional neural networks can better adapt to The impact that displacement brings, keeps intermediate object program not have the biggest deviation.After data size reduces continuously, open because calculating Pin minimizing, we can suitably network deep layer increase characteristic pattern number so that convolutional neural networks can learn to arrive More picture feature.
It addition, in order to improve operation efficiency, each filter may utilize the multiple CPU (or GPU) in server to each window What the image block of position was parallel merges convolution.
Such as, described downsampling unit is grouped, in the characteristic image block received every 4 × 4 according to the pixel of 4 × 4 The eigenvalue of pixel is averaged, and the corresponding pixel points this meansigma methods being assigned to after down-sampling in image block.
In order to improve the processing speed that follow-up up-sampling processes, a kind of alternative is, described based under convolutional network Sampling apparatus 12 concentrates the characteristic information of each image block further.
Specifically, described downsampling device based on convolutional network 12 comprises: many groups by filter and downsampling unit group The structure become, and respectively group structure cascades with one another.
Wherein, the filter connected with described image received device 11 has the reception consistent with stacking amount of images and leads to Road, described reception passage transmits a width stacking image;Described filter is for by the same window position corresponding in each stacking image Each image block merge convolution, obtain should the characteristic image block of the window's position.In other groups, the reception of filter is led to Road and the output channel cascade of previous stage filter, be used for the characteristic image block received is carried out full convolution.Every grade of filtration The output channel of device is provided with described downsampling unit, for the characteristic image block received being carried out down-sampling, and by under adopt Characteristic image block after sample sends into next stage filter or the reception passage of up-sampling device 13.
Such as, described downsampling device based on convolutional network 12 comprises 4 grades of said structure Structure1, Structure2, Structure3 and Structure4.Wherein, every level structure comprises multiple filter and a down-sampling list Unit, the filter in each structure is also that cascade connects, and last filter of downsampling unit reception place group is exported Characteristic image block, and carry out down-sampling process.The filter being connected with described image received device 11 in structure Structure1 Each image block of the same window position corresponding in each stacking image is merged convolution, and by each characteristic pattern after merging convolution Send into next filter as block and carry out feature extraction, until last filter is by each characteristic pattern in structure Structure1 As block filters complete, each characteristic image block feeding downsampling unit is carried out down-sampling.Structure Structure1 is by after down-sampling The filter that is sequentially sent in structure Structure2 of each characteristic image block carry out feature extraction one by one, and in structure After in Structure2, each characteristic image block is filtered by last filter, by the downsampling unit in this structure to respectively Characteristic image block carries out down-sampling again.When each stacking image is through the feature of described downsampling device 12 based on convolutional network Extract and after down-sampling, the size of each characteristic image block is with the size much smaller than original picture block, but still remains original graph Characteristic information as block.
Then, each characteristic image block is admitted to up-sample device 13.Described up-sampling device 13 is for by after down-sampling Each characteristic image block carries out up-sampling process so that at least each characteristic image block after the up-sampling corresponding to adjacent window apertures position There is pixel overlapping.
Specifically, described up-sampling device 13 can according to the characteristic image block received and original picture block dimensionally Gap, according to the region in the original picture block corresponding to each pixel in each characteristic image block, the value of each pixel is multiple Make in respective regions.
Such as, described up-sampling device 13 utilizesUp-sample.Wherein, described region Size be (s, s).It can be seen that it is the rectangle of s*s that each input pixel has been duplicated into a size in the output, so Up-sampling while amplification data scale, maintain the space structure of former data.
In order to make the temperature prognostic chart finally given describe the common trait in each width stacking image, Yi Zhongke more accurately Selecting scheme is that the characteristic image block received, according to described window size, is carried out based on secondary-plug by described up-sampling device 13 The up-sampling of value.
Specifically, described up-sampling device 13 includes: up-sampling unit and smooth processing unit.
Wherein, described up-sampling unit, for based on each pixel value in the characteristic image block received, is filled in default Expansion window, with the characteristic image block once up-sampled.
Described up-sampling unit may utilize formula:Each in the characteristic image block that will be received What pixel repeated is filled in the corresponding region presetting expansion window.Wherein, described expansion window be smaller in size than aforementioned use In the window size that image is carried out piecemeal.Being probably real number in view of i/s and j/s, this example uses four nearest pixels Point row quadratic interpolation removes to approach the numerical value of each pixel.
Described smoothing processing unit is for based on the default convolution kernel consistent with window size, to through once up-sampling Characteristic image block carry out convolution, obtain corresponding described window size secondary up-sampling characteristic image block.
Use the up-sampling mode of described up-sampling unit and smooth processing unit, be an advantage over image intuitively directly Block is upsampled to, but it is very nearly the same to neutralize the former in the effect of actual experiment.If factor behind is exactly up-sampling layer It is more than s, then secondary above-mentioned after this filter followed by a convolutional layer and convolutional layer filter core size Interpolation can be obtained by the study of filter parameter.The most above-mentioned quadratic interpolation is included in later in some sense In convolutional layer, although not equivalent.The acceptance region of above-mentioned filter layer has intermeshing clearly, and each mistake Filter parameter is all shared in layer, and therefore propagated forward based on GPU and back propagation calculating can become the most efficient.
Here, described smoothing processing unit can only comprise the convolution kernel consistent with window size, it is possible at the base of convolution Carry out processing such as weighted average etc. to each characteristic image block on plinth.Each characteristic image block after described smoothing processing cell processing 1 span can be had according to the window's position order.
Here, the characteristic image block after up-sampling processes i.e. can be considered that each stacking image of reaction is at the respective window location drawing Temperature as block.
Then, described temperature prediction means 14 at least one characteristic image block after utilizing up-sampling is true or is false Probability draw the color of each pixel in temperature prognostic chart;Wherein, each pixel in described temperature prognostic chart belongs to At least one characteristic image block after up-sampling.
Here, described temperature prediction means 14 can share the hardware such as the processing unit in aforementioned each device, caching, and receive Each characteristic image block that up-sampling device 13 is exported.
Specifically, described temperature prediction means 14 utilizes each characteristic image block received consistent with true value or inconsistent Probability evaluate the temperature information of each pixel in temperature prognostic chart.
Judging to last input temperature figure y in view of convolutional neural networks, therefore, we will be to output y definition One loss function.Loss function can be output and true valueBetween quadratic sum.Such as, described loss function isDescribed loss function can also be arranged based on cross entropy.Such as, described loss function is
In actual applications, due to true value the most likely value 0 or 1, loss function just takes cross entropy in this case Can more show the degree of closeness of output and true value.A kind of optional mode is that described temperature prediction means 14 is according to default Loss function, carries out temperature evaluation one by one to the corresponding each characteristic image block received of each pixel of temperature prognostic chart, and Corresponding relation based on default temperature evaluation Yu temperature color, draws described temperature prognostic chart.Wherein, described loss function is used In the probability representing that in temperature prognostic chart, pixel place each characteristic image block is consistent or inconsistent with true value.Such as, described Loss function is characterized the mean square deviation between image block and true value, and the most described temperature prediction means 14 is the most right In temperature prognostic chart, the temperature of a pixel is evaluated:
At characteristic image block yi、yi+1... and yi+tIn all comprise pixel O in temperature prognostic chart (a, b), then described Temperature prediction means 14 utilizes formulaCalculate pixel O (a, temperature b).
It should be noted that above-mentioned loss function is only for example rather than limitation of the present invention.It is true that loss function is also Can be cross entropy or unknown losses function.
In order to predict the temperature of each stacking image more accurately, described temperature prediction means 14 obtains based on through study in advance To comprise true value weight and the loss function of falsity weight, each feature that each pixel of temperature prognostic chart correspondence is received Image block carries out temperature evaluation one by one, and draws described temperature prognostic chart based on obtained temperature evaluation.
Wherein, the loss function obtained through machine learning in advance can be the letter improved based on above-mentioned mean square deviation formula Number.In the present embodiment, preferred loss function obtains based on learning in advance and comprises true value weight and the friendship of falsity weight Fork entropy function:
l o s s = Σ i ( y i log ( y i ^ ) b l a c k _ r a t i o + ( 1 - y i ) log ( 1 - y i ^ ) w h i t e _ r a t i o ) .
Wherein, black_ratio represents the true value weight determined through machine learning in advance, and white_ratio represents warp Machine learning in advance and the falsity weight that determines.
Utilize above-mentioned loss function, described temperature prediction means 14 temperature of pixel each in temperature prognostic chart is carried out by The temperature evaluation of one, and then according to the corresponding relation of default temperature evaluation Yu color, draw temperature prognostic chart.
The true value of black matrix and the true value of white background can be carried out equalization, so that neutral net exists by balanced, crossover entropy well Training is just easier to restrain towards globally optimal solution.
Embodiment two
The basis of each alternative in above-described embodiment one, the present embodiment provides a kind of based on convolutional neural networks Lung tumor identification system.Wherein, described lung tumor identification system based on convolutional neural networks applies the CT in hospital to scheme As in scanner or the image processing equipment (such as server) that is connected with CT image scanning instrument.For CT image reading apparatus For device, at least include CT scan device.Described CT scan device is by empty for the edge, position to be checked (such as lung, brain or whole body) of live body Between axially carry out the scanning of cross section type, obtain several CT stacking images corresponding.Corresponding, described based on convolutional neural networks Lung tumor identification system 3 include: image library 31, the image combined such as any one alternative in above-described embodiment one Temperature prognoses system 32 and control device 33.As shown in Figure 4.
In the present embodiment, temporarily with lung tumor identification system diagram based on convolutional neural networks as temperature prognoses system 32 As a example by illustrate, described lung tumor identification system based on convolutional neural networks 3 is for identifying the tumor of pulmonary.
Wherein, described image temperature prognoses system 32 and CT scan device 2 share an image library 31.CT scan device 2 will The CT stacking image scanned is stored in image library 31 according to spatial order.Described image temperature prognoses system 32 is controlling dress Put under the control of 33, read several CT stacking images therein, and according to any one scheme described in embodiment one, carry out many The temperature prediction of width CT stacking image, and according to the operation of doctor, obtained temperature prognostic chart is supplied to doctor.
Wherein, described control device 33 is for judging whether described image temperature prognoses system 32 is extracted stored institute There is CT stacking image, if it is not, then control described image temperature prognoses system 32 to extract several CT stackings by several times according to spatial order Image, till determining that described image acquiring apparatus is extracted all CT stacking images stored.
Such as, described control device 33 is according to the order of stacking image each in image library 31, with 5 width CT stacking images for one Group, by several times provide several stacking images to image temperature prognoses system 32.When described image temperature prognoses system 32 is according to institute During the 5 width CT stacking images output temperature prognostic chart received, described control device 33 judges whether to be extracted stored all CT stacking image, the most then exit circulation, if otherwise continuing to extract next group CT stacking image, and inputs described image temperature Prognoses system 32, until having extracted all CT stacking images.
Wherein, the temperature prognostic chart that described image temperature prognoses system 32 is exported can be considered that this group CT stacking image shares Temperature prognostic chart.Preferably, the temperature prognostic chart that described image temperature prognoses system 32 is exported is considered this group CT stacking Middle temperature prognostic chart corresponding to a width CT stacking image in image.
In order to improve the temperature precision of prediction of CT stacking image, in a kind of alternative, travel through CT layer judging to be not fully complete In the case of folded image, described control device 33 span based on piece image extracts the CT stacking image of difference group, 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 images.
Such as, described control device 33 is first by { x1,x2,x3,x4,x55 width CT stacking images send into image temperatures prediction system System 32 is carried out for image x3Temperature prediction.When judging x5Not in image library 31 during last of CT stacking image, described Control device 33 to continue to select { x2,x3,x4,x5,x65 width CT stacking images send into image temperature prognoses systems 32 carry out for figure As x4Temperature prediction.So repeat, until judging the one group of CT stacking image extracted comprises last, exiting and following Ring.
Flow process as shown in Figure 7, the work process of described lung tumor identification system 3 based on convolutional neural networks is illustrated such as Under:
After CT image scanning instrument obtains several CT stacking figures of patient, by patient data and these several CT stacking figures Correspondence is saved in image library 31.As the scanned instruction based on CT image scanning instrument of described control device 33 or doctor Give birth to by operation terminal when control device 33 place equipment have sent the instruction obtaining temperature prognostic chart, described control device 33 Find several CT stacking images deposited according to scanning sequency according to patient data correspondence, and inform in image temperature prognoses system 32 Image acquiring apparatus extract 1-5 and open CT stacking image.
Described image acquiring apparatus extracts several CT stacking images corresponding from image library 31 and utilizes corresponding each CT The passage that receives of stacking image passes to downsampling device based on convolutional network.Described downsampling device based on convolutional network Filter F ilter11 in middle first order structure is according to unified window size and window movement rule, by each CT stacking image Each image block of middle the same window position merges convolution, so obtains the characteristic image block of corresponding each the window's position.? Primary structure also includes and at least one filter (Filter12 ..., Filter1n) of Filter11 cascade.In the first order In structure, the convolution kernel size of each filter can be consistent with above-mentioned window size, in order to reduces amount of calculation and optimizes convolution The structure of neutral net.After downsampling unit in first order structure is positioned at filter F ilter1n, it is by by feature The mode of image block packet down-sampling, carries out down-sampling by the characteristic image block of corresponding each the window's position.Again by after down-sampling Characteristic image block sends at least one filter (Filter21 ..., Filter2n) in the structure of the second level.In the structure of the second level Each filter by each characteristic image block received is carried out full process of convolution respectively, obtain entering of corresponding each the window's position The characteristic image block of one step, wherein, the size being smaller in size than received characteristic image block of the convolution kernel in the structure of the second level. Again by the downsampling unit in the structure of the second level, reduce the size of each characteristic image block further.By m cascade structure, Characteristic image block corresponding to each the window's position is significantly reduced.
Then, up-sampling device utilizing the mode of quadratic interpolation, each characteristic image block after reducing recovers, with Obtain the temperature image block of corresponding each the window's position.
Then, temperature prediction means each temperature image block is utilized to have the feature of overlapping region, by temperature prognostic chart The temperature evaluation of each pixel transfers to the temperature image block covering this pixel to determine.Specifically, described temperature prediction dress Put use based on through learn in advance and obtain comprise true value weight and the loss function of falsity weight, each to temperature prognostic chart Each characteristic image block (i.e. temperature image block) that pixel correspondence is received carries out temperature evaluation one by one, and based on obtained Described temperature prognostic chart is drawn in temperature evaluation, and the temperature prognostic chart drawn is centre one width in corresponding 5 width CT stacking images The temperature prognostic chart of CT stacking image.
When described control dress determine image temperature prognoses system 32 complete the 5 width CT stacking images controlled in about in Between the temperature prognostic chart of a width CT stacking image time, it is judged that the 5 width CT stacking images completed have comprised corresponding same disease the most Last of people, the most then exit circulation, if it is not, 2-6 then opens CT stacking image transfer to image received device, lay equal stress on New execution image temperature prognoses system 32, to obtain the temperature prognostic chart of corresponding 4th width CT stacking image.So iterative cycles, Described lung tumor identification system based on convolutional neural networks can obtain in addition to each two width CT stacking images of head and the tail The temperature prognostic chart of remaining each CT stacking image.
Embodiment three
As it is shown in figure 5, the present invention provides the flow chart of a kind of image temperature Forecasting Methodology.Described image temperature Forecasting Methodology Mainly performed by prognoses system.Wherein said prognoses system includes the software and hardware being arranged in computer equipment.Wherein, Hardware in described computer equipment comprises: input block, processing unit, memory element, caching and display unit etc., wherein, Described processing unit can comprise and be exclusively used in chip or the integrated circuit of convolutional neural networks and include convolutional Neural net The computer program of network algorithm.Described processing unit distributes the operation of each hardware by the sequential of program setting, following to perform The function of each device.Wherein, described computer equipment includes but not limited to: single server, multiple server coordinate operation Server cluster etc..
Described image temperature Forecasting Methodology is by performing following steps, it is achieved the prediction to image temperature.
Step S110, reception are along presetting several stacking images that dimension gathers.Here, described prognoses system can comprise place The interface that reason unit, caching and the image library with storage stacking image are connected.Described prognoses system refers to according to the sequential of program Show, read from image library in advance along presetting several stacking images that dimension gathers by interface.Wherein, several stacking figures described Seem from along captured by the time dimension preset or Spatial Dimension, can overlap to embody the association of captured dimensional characteristics Property image set in read.
Such as, in CT image library, described prognoses system reads several CT stacking figures of continuous print from described CT image library Picture, wherein, the quantity of described CT stacking image is odd number.Heat is carried out for the ease of follow-up several stacking images to being received Degree prediction, the consistent size of several stacking images described.
In a concrete example, the lung CT figure of described several stacking images for example, h × 512 × 512 size Sheet, wherein h is the scanning number of plies of CT.
In step S120, described prognoses system, downsampling device based on convolutional network is according to the same window and identical Movement rule, carries out piecemeal by each described stacking image, and by the image block in each stacking image of corresponding the same window position Merge convolution, to obtain characteristic image block, and each characteristic image block is carried out down-sampling.
Convolutional network in described downsampling device based on convolutional network is convolutional neural networks (Convolutional Neural Network, is called for short CNN), described convolutional neural networks includes convolutional layer (alternating convolutional And pond layer (pooling layer) layer).Wherein, described convolutional layer can be considered the follow-up each filter that will describe in detail, Described pond layer can be considered the follow-up downsampling unit that will describe in detail.
Convolutional network in described downsampling device based on convolutional network is folded by some layer stack and is formed, general under The output of one layer is the input of higher.Picture is inputted by the bottom, and top output is final result.More specifically god Can be a directed graph through the structure of network, be used for 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 represents the special of characteristic pattern or passage.More specifically, right In input picture, h and w is exactly the height and width of picture, and d is input picture color channel number (the Color Channel number of conventional RGB figure Being 3, gray-scale map is 1).
Here, described downsampling device based on convolutional network includes: can process volume machine network algorithm processing unit, And the caching matched.This processing unit can share with the processing unit in image received device, it is also possible to refers to include It is exclusively used in the chip of convolutional neural networks or the unit of integrated circuit.If described downsampling device based on convolutional network comprises Said chip or integrated circuit.In order to improve image processing speed, between described chip or integrated circuit and image received device It is connected by the image-receptive passage of hardware.Each stacking figure that each reception image received device receiving channel parallel is provided Picture.
Specifically, described downsampling device based on convolutional network utilizes above-mentioned each hardware, composition comprise filter and under The version of sampling unit.Wherein, the quantity of described filter can be one, it is also possible to is multiple.Each filter and under Sampling unit can be integrated in described chip/integrated circuit.Described image received device by each reception passage and at least one Described filter is connected.Wherein, described filter include following at least one: filter based on color, mistake based on lines Filter and filter based on gray scale.Wherein, the window size of described window is less than the picture size of stacking image.Described window Movement rule at each stacking image is unified.Processing unit in described downsampling device based on convolutional network first from Same starting pixels point (such as [0,0]) in each width stacking image starts, according to same movement rule (if span is the movement of 1 Rule), each width stacking image is carried out piecemeal.
Such as, stacking image A is divided into image block a11, a12 by described processing unit ..., a1n, a21, a22 ..., 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 according to actual temperature forecast demand, described window can be 1 pixel column/pixel according to span The movement rule of row carries out the piecemeal of blanket type and processes each width stacked graph picture, it is also possible to carry out point for default image-region Block processes.Existing or from now in technical scheme if any the partitioned mode carried out based on the present embodiment enlightenment all of the present invention Within the scope of.
Then, the image block of same footmark in each stacking image (i.e. footmark is all [m, n]) is sent into by described processing unit Same filter merges convolution.Wherein, described merging convolution refer to by each image block all with the convolution kernel in filter Carry out convolution and be merged into a characteristic image block, so that the feature letter that in the characteristic image block after Guo Lving, each pixel is reacted Breath can the feature of each image block that inputted of integrating representation.
Such as, as in figure 2 it is shown, filter receives footmark in each stacking image is all the image block of [m, n], and by each image Block carries out convolution respectively with the convolution kernel in filter, and the result after each pixel convolution is taken and, obtaining can be comprehensive Embody the pixel characteristic information of each image block characteristics.
Or, described blocking process can be the part in filter.Described downsampling device based on convolutional network Convolutional network in 12 is folded by some layer stack and is formed, and the output of next general layer is the input of higher.Picture is by Bottom inputs, and top output is final result.The more specifically structure of neutral net can be a directed graph, is used for Complete some special tasks.The input of each filter layer is all three-dimensional array h, w, d, and wherein h and w is input Size, d represents the special of characteristic pattern or passage.More specifically, for input picture, h and w is 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, and gray-scale map is 1).When each stacking image is according to signal Input and output path enter filter time, the convolution kernel size in filter is window size.In the present embodiment, according to entirely The filter type of convolution, by each stacking image input filter, by filter according to window movement rule to piecemeal out Each image block of corresponding the same window position merges convolution.
In a kind of alternative, described downsampling device based on convolutional network includes: normalization unit, for by institute State 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 pixel value of each stacking image is normalized between [0,1] by described normalization unit.Again by normalization After each stacking image input filter.
In alternative dispensing means be, described filter for according to Each image block of the same window position corresponding in each stacking image is merged convolution;Wherein, WkFor kth convolution kernel, bkFor The side-play amount that corresponding kth convolution kernel adds on each passage, c is the port number of stacking image, and (u v) is WkSize, (i, j) For pixel position in image block.
In order to preferably extract can the characteristic information of each image block of integrating representation, described down-sampling based on convolutional network Filter in device is multiple, and the cascade of each filter.As shown in Figure 3.
Such as, 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 reception passage.Filter Filter12 and Filter13 is after level is associated in Filter11 successively, and the characteristic image block being exported Filter11 carries out pin The dependency of lines and gray scale is filtered.
Then, described downsampling unit chooses the mode such as extreme value or meansigma methods according to from default pixel packet, right Each characteristic image block of filter output carries out down-sampling.Wherein, in every level structure the convolution kernel size of filter be not more than In downsampling unit, the size of pixel packet is preferred, i.e. 0≤u≤s;0≤v≤s.Wherein, (u, v) is convolution kernel size, (s, S) it is the size of pixel packet in downsampling unit.
It appeared that after by window size being the down-sampling layer of s, the port number of data keeps constant, and length and width are respectively Taper to original 1/s, so make total data scale fall below original 1/s2.Down-sampling layer it is also an advantage that, just Being if being added without down-sampling layer, convolutional neural networks can be especially sensitive to the position of object, even movement the most by a small margin There is great deviation in the intermediate object program that can allow network.After adding down-sampling layer, convolutional neural networks can better adapt to The impact that displacement brings, keeps intermediate object program not have the biggest deviation.After data size reduces continuously, open because calculating Pin minimizing, we can suitably network deep layer increase characteristic pattern number so that convolutional neural networks can learn to arrive More picture feature.
It addition, in order to improve operation efficiency, each filter may utilize the multiple CPU (or GPU) in server to each window What the image block of position was parallel merges convolution.
Such as, described downsampling unit is grouped, in the characteristic image block received every 4 × 4 according to the pixel of 4 × 4 The eigenvalue of pixel is averaged, and the corresponding pixel points this meansigma methods being assigned to after down-sampling in image block.
In order to improve the processing speed that follow-up up-sampling processes, a kind of alternative is, described based under convolutional network Sampling apparatus concentrates the characteristic information of each image block further.
Specifically, described downsampling device based on convolutional network comprises: many groups are made up of filter and downsampling unit Structure, and each group structure cascades with one another.
Wherein, there is with the filter that described image received device connects the reception passage consistent with stacking amount of images, Described reception passage transmits a width stacking image;Each for by the same window position corresponding in each stacking image of described filter Image block merges convolution, obtains should the characteristic image block of the window's position.In other groups the reception passage of filter with The output channel cascade of previous stage filter, for carrying out full convolution by the characteristic image block received.At every grade of filter Output channel is provided with described downsampling unit, for the characteristic image block received carries out down-sampling, and by after down-sampling Characteristic image block send into next stage filter or up-sampling device reception passage.
Such as, described downsampling device based on convolutional network comprises 4 grades of said structure Structure1, Structure2, Structure3 and Structure4.Wherein, every level structure comprises multiple filter and a down-sampling list Unit, the filter in each structure is also that cascade connects, and last filter of downsampling unit reception place group is exported Characteristic image block, and carry out down-sampling process.The filter being connected with described image received device in structure Structure1 will In each stacking image, each image block of corresponding the same window position merges convolution, and by each characteristic image after merging convolution Block is sent into next filter and is carried out feature extraction, until last filter is by each characteristic image in structure Structure1 Block filters complete, and each characteristic image block feeding downsampling unit is carried out down-sampling.Structure Structure1 is by after down-sampling The filter that each characteristic image block is sequentially sent in structure Structure2 carries out feature extraction one by one, and in structure After in Structure2, each characteristic image block is filtered by last filter, by the downsampling unit in this structure to respectively Characteristic image block carries out down-sampling again.
When each stacking image is after the feature extraction and down-sampling of described downsampling device based on convolutional network, each spy Levy the size of image block with the size much smaller than original picture block, but still remain the characteristic information of original picture block.
The up-sampling device that step S130, each characteristic image block are admitted in described prognoses system.Described up-sampling device Each characteristic image block after down-sampling is carried out up-sampling process so that at least after the up-sampling corresponding to adjacent window apertures position It is overlapping that each characteristic image block has pixel.
Specifically, described up-sampling device can be according to the characteristic image block received and original picture block difference dimensionally Away from, according to the region in the original picture block corresponding to each pixel in each characteristic image block, the value of each pixel is replicated In respective regions.
Such as, described up-sampling device utilizesUp-sample.Wherein, described region A size of (s, s).It can be seen that it is the rectangle of s*s that each input pixel has been duplicated into a size in the output, such Up-sampling maintains the space structure of former data while amplification data scale.
In order to make the temperature prognostic chart finally given describe the common trait in each width stacking image, Yi Zhongke more accurately Selecting scheme is that the characteristic image block received, according to described window size, is carried out based on quadratic interpolation by described up-sampling device Up-sampling.
Specifically, described step S130 includes: S131, S132 (are all unillustrated).
Step S131, based on each pixel value in the characteristic image block received, be filled in default expansion window, with To the characteristic image block once up-sampled.
Specifically, formula is utilized:Each pixel weight in the characteristic image block that will be received Multiple being filled into is preset in the corresponding region expanding window.Wherein, described being smaller in size than of expansion window is previously used for image Carry out the window size of piecemeal.Being probably real number in view of i/s and j/s, this example uses four nearest pixel row secondaries Interpolation removes to approach the numerical value of each pixel.
Step S132, based on the default convolution kernel consistent with window size, to the characteristic image through once up-sampling Block carries out convolution, obtains the characteristic image block of the secondary up-sampling of corresponding described window size.
Use step S131 and the up-sampling mode of S132, be an advantage over intuitively being directly upsampled to by image block, But it is very nearly the same that the effect at actual experiment neutralizes the former.If factor behind is exactly, up-sampling layer is followed by a volume Lamination and convolutional layer filter core size are more than s, then quadratic interpolation above-mentioned after this filter can be passed through The study of filter parameter obtains.The most above-mentioned quadratic interpolation is included in later convolutional layer in some sense, though The most not equivalent.The acceptance region of above-mentioned filter layer has intermeshing clearly, and each filter parameter is at layer Being the most all shared, therefore propagated forward based on GPU and back propagation calculating can become the most efficient.
Here, the convolution kernel consistent with window size can only be comprised, it is possible to each characteristic image block on the basis of convolution Carry out processing such as weighted average etc..Each characteristic image block after described smoothing processing cell processing can according to the window's position order There is 1 span.
Here, the characteristic image block after up-sampling processes i.e. can be considered that each stacking image of reaction is at the respective window location drawing Temperature as block.
Then, step S140, utilize up-sampling after at least one characteristic image block be true or be false probability draw The color of each pixel in temperature prognostic chart;Wherein, after each pixel in described temperature prognostic chart belongs to up-sampling extremely A few characteristic image block.
Specifically, described prognoses system utilizes the probability that each characteristic image block received is consistent or inconsistent with true value Evaluate the temperature information of each pixel in temperature prognostic chart.
Judging to last input temperature figure y in view of convolutional neural networks, therefore, we will be to output y definition One loss function.Loss function can be output and true valueBetween quadratic sum.Such as, described loss function isDescribed loss function can also be arranged based on cross entropy.Such as, described loss function is
In actual applications, due to true value the most likely value 0 or 1, loss function just takes cross entropy in this case Can more show the degree of closeness of output and true value.
A kind of optional mode is that described prognoses system is according to default loss function, each pixel to temperature prognostic chart The corresponding each characteristic image block received carries out temperature evaluation one by one, and based on default temperature evaluation and temperature color right Should be related to, draw described temperature prognostic chart.Wherein, described loss function is used for representing each spy in pixel place in temperature prognostic chart Levy the probability that image block is consistent or inconsistent with true value.Such as, described loss function is characterized between image block and true value Mean square deviation, the temperature of in temperature prognostic chart pixel is commented by the most described temperature prediction means the most successively Valency:
At characteristic image block yi、yi+1... and yi+tIn all comprise pixel O in temperature prognostic chart (a, b), then described Prognoses system utilizes formulaCalculate pixel O (a, temperature b).
It should be noted that above-mentioned loss function is only for example rather than limitation of the present invention.It is true that loss function is also Can be cross entropy or unknown losses function.
In order to predict the temperature of each stacking image more accurately, described prognoses system is based on the bag obtained through learning in advance Containing true value weight and the loss function of falsity weight, the corresponding each characteristic image block received of each pixel to temperature prognostic chart Carry out temperature evaluation one by one, and draw described temperature prognostic chart based on obtained temperature evaluation.
Wherein, the loss function obtained through machine learning in advance can be the letter improved based on above-mentioned mean square deviation formula Number.In the present embodiment, preferred loss function obtains based on learning in advance and comprises true value weight and the friendship of falsity weight Fork entropy function:
l o s s = Σ i ( y i log ( y i ^ ) b l a c k _ r a t i o + ( 1 - y i ) log ( 1 - y i ^ ) w h i t e _ r a t i o ) .
Wherein, black_ratio represents the true value weight determined through machine learning in advance, and white_ratio represents warp Machine learning in advance and the falsity weight that determines.
Utilizing above-mentioned loss function, described prognoses system carries out heat one by one to the temperature of pixel each in temperature prognostic chart Degree is evaluated, and then according to the corresponding relation of default temperature evaluation Yu color, draws temperature prognostic chart.
The true value of black matrix and the true value of white background can be carried out equalization, so that neutral net exists by balanced, crossover entropy well Training is just easier to restrain towards globally optimal solution.
Embodiment four
The basis of each alternative in above-described embodiment three, as shown in Figure 6, the present embodiment provides a kind of based on convolution The lung tumor identification method of neutral net.Wherein, described lung tumor identification method based on convolutional neural networks mainly by Lung tumor identification system based on convolutional neural networks performs.Described lung tumor identification system based on convolutional neural networks System is applied at the CT image scanning instrument of hospital or the image processing equipment (such as server) that is connected with CT image scanning instrument In.For CT image scanning instrument, at least include CT scan device.Described CT scan device is by the position to be checked of live body (such as lung, brain or whole body) axially carries out the scanning of cross section type along space, obtains several CT stacking images corresponding.
In the present embodiment, described lung tumor identification method based on convolutional neural networks is utilized to identify the swollen of pulmonary Tumor.Described lung tumor identification method based on convolutional neural networks is used the figure described by each alternative in embodiment three As temperature Forecasting Methodology.It is to say, described lung tumor identification system based on convolutional neural networks comprises image temperature Prognoses system.
Wherein, described image temperature prognoses system and CT scan device share an image library.
Specifically, the CT stacking image scanned is stored in image library by CT scan device according to spatial order.
Then, described lung tumor identification system based on convolutional neural networks controls rule according to described, starts image Temperature prognoses system performs step S220: read several CT stacking images therein, and arbitrary according to described in embodiment three The scheme of kind, carries out the temperature prediction of several CT stacking images, and according to the operation of doctor, is provided by obtained temperature prognostic chart To doctor.Idiographic flow is shown in Fig. 7.Wherein, Fig. 8 is lung CT stacking figure gray-scale map, artificial cognition lung tumors position and utilization The schematic diagram of the lung tumors position of the present embodiment identification.
Followed by, described lung tumor identification system based on convolutional neural networks controls rule according to described, performs step Rapid S230: judge whether described image temperature prognoses system is extracted all CT stacking images stored, if it is not, then control institute State image temperature prognoses system and extract several CT stacking images, until determining described image acquiring apparatus by several times according to spatial order Till being extracted all CT stacking images stored.
Such as, suitable according to stacking image each in image library of described lung tumor identification system based on convolutional neural networks Sequence, is one group with 5 width CT stacking images, from the beginning of the first width CT stacking image, and providing to image temperature prognoses system of packet CT stacking image.When described image temperature prognoses system is according to the 5 width CT stacking image output temperature prognostic chart received, institute State control device to judge whether to be extracted all CT stacking images stored, the most then exit circulation, if otherwise continuing to extract Next organizes CT stacking image, and inputs described image temperature prognoses system, until having extracted all CT stacking images.
Wherein, the temperature prognostic chart that described image temperature prognoses system is exported can be considered what this group CT stacking image shared Temperature prognostic chart.Preferably, the temperature prognostic chart that described image temperature prognoses system is exported is considered this group CT stacking image In the temperature prognostic chart corresponding to the width CT stacking image of centre one.
In order to improve the temperature precision of prediction of CT stacking image, in a kind of alternative, travel through CT layer judging to be not fully complete In the case of folded image, described lung tumor identification system based on convolutional neural networks in advance with the span of piece image to CT Stacking image is grouped, and so can obtain the temperature prediction of all CT stacking images in addition to head and the tail two width CT stacking images Figure.
Such as, first by { x1,x2,x3,x4,x55 width CT stacking images send into image temperature prognoses systems carry out for image x3Temperature prediction.When judging x5Not in image library during last of CT stacking image, described control device continues to select {x2,x3,x4,x5,x65 width CT stacking images send into image temperature prognoses systems carry out for image x4Temperature prediction.So Repeating, until judging the one group of CT stacking image extracted comprises last, exiting circulation.
As it is shown in fig. 7, the work process of described lung tumor identification system based on convolutional neural networks is exemplified below:
After CT image scanning instrument obtains several CT stacking figures of patient, by patient data and these several CT stacking figures Correspondence is saved in image library.When described lung tumor identification system based on convolutional neural networks is based on CT image scanning instrument Scanned instruction or doctor have sent by operation terminal and obtain the finger of temperature prognostic chart to controlling device place equipment When making, described control device finds several CT stacking images deposited according to scanning sequency according to patient data correspondence, and informs figure CT stacking image is opened as the image acquiring apparatus in temperature prognoses system extracts 1-5.
Image acquiring apparatus in described image temperature prognoses system extracts several CT stacking figures corresponding from image library Picture also utilizes the passage that receives of corresponding each CT stacking image to pass to downsampling device based on convolutional network.Described based on volume In the downsampling device of long-pending network, filter F ilter11 in first order structure moves according to unified window size and window Rule, merges convolution by each image block of the same window position in each CT stacking image, so obtains corresponding each window position The characteristic image block put.First order structure also includes and at least one filter of Filter11 cascade (Filter12、…、Filter1n).In first order structure, the convolution kernel size of each filter can be with above-mentioned window size phase Unanimously, in order to reduce amount of calculation and optimize the structure of convolutional neural networks.Downsampling unit in first order structure is positioned at After filter F ilter1n, it is by the way of by characteristic image block packet down-sampling, by the feature of corresponding each the window's position Image block carries out down-sampling.Again the characteristic image block after down-sampling is sent at least one filter in the structure of the second level (Filter21、…、Filter2n).Each filter in the structure of the second level is by entering each characteristic image block received respectively The full process of convolution of row, obtains the further feature image block of corresponding each the window's position, wherein, the convolution kernel in the structure of the second level The size being smaller in size than received characteristic image block.Again by the downsampling unit in the structure of the second level, reduce further The size of each characteristic image block.By repeatedly cascade structure, the characteristic image block corresponding to each the window's position is significantly reduced.
Then, the up-sampling device in described image temperature prognoses system the mode of quadratic interpolation is utilized, after reducing Each characteristic image block recover, to obtain the temperature image block of corresponding each the window's position.
Then, the temperature prediction means in described image temperature prognoses system each temperature image block is utilized to have overlay region The feature in territory, transfers to the temperature image block covering this pixel to come really by the temperature evaluation of pixel each in temperature prognostic chart Fixed.Specifically, described temperature prediction means use based on through learn in advance and obtain comprise true value weight and falsity weight Loss function, is carried out one by one the corresponding each characteristic image block (i.e. temperature image block) received of each pixel of temperature prognostic chart Temperature evaluation, and draw described temperature prognostic chart based on obtained temperature evaluation, the temperature prognostic chart drawn is corresponding 5 The temperature prognostic chart of the centre one width CT stacking image in width CT stacking image.Fig. 8 is followed successively by lung CT stacking figure from left to right As gray-scale map, the tumor figure of handmarking and the lung tumors temperature utilizing described image temperature prognoses system to be identified are predicted Comparison diagram between figure three.
When described control device determine during image temperature prognoses system completes the 5 width CT stacking images controlled about in Between the temperature prognostic chart of a width CT stacking image time, it is judged that the 5 width CT stacking images completed have comprised corresponding same disease the most Last of people, the most then exit circulation, if it is not, 2-6 then opens CT stacking image transfer to image received device, and in advance Survey the temperature prognostic chart of the 4th CT stacking image.So iterative cycles, described lung tumor identification based on convolutional neural networks System can obtain the temperature prognostic chart of remaining each CT stacking image in addition to each two width CT stacking images of head and the tail.
In sum, various embodiments of the present invention are extracted by convolutional neural networks and can be represented the common of several stacking images Each characteristic image block of feature, has so obtained each image and has included the characteristic information of dimension outside place plane and plane, entered And by each characteristic image block being up-sampled and the prediction of temperature figure, be effectively increased for the stacking image with spatial correlation The discrimination of regional area.Meanwhile, convolutional neural networks arranges downsampling unit, advantageously reduces convolutional neural networks Amount of calculation, be effectively improved the efficiency of tumor identification.
It addition, use the mode of quadratic interpolation when up-sampling, it is possible to the color transition of smooth each image block, improve and identify Accuracy.
It addition, use odd number continuous print stacking image, not only contribute to the characteristic information of retaining space dimension, moreover it is possible to protect Demonstrate,prove the Balance Evaluation of image between forward and backward image pair.
It addition, it is the most optimum to embody model training because of loss function, it is impossible to as last accurate result index. Calculate through the later stage of experiment and contrast the position of each tumor and calculate False Positive and False Negative.This Invention has descended to close to 5% on False Negative, have decreased to 20% on False Positive.Thus may be used Seeing, compared with prior art, the present invention has been achieved for breakthrough.
So, the present invention effectively overcomes various shortcoming of the prior art and has high industrial utilization.
The principle of above-described embodiment only illustrative present invention and effect thereof, not for limiting the present invention.Any ripe Above-described embodiment all can be modified under the spirit and the scope of the present invention or change by the personage knowing this technology.Cause This, have usually intellectual such as complete with institute under technological thought without departing from disclosed spirit in art All equivalences become are modified or change, and must be contained by the claim of the present invention.

Claims (24)

1. an image temperature prognoses system, it is characterised in that including:
Image received device, for receiving along presetting several stacking images that dimension gathers;
Downsampling device based on convolutional network, for according to the same window and identical movement rule, by each described stacking figure As carrying out piecemeal process, and the image block of the same window position corresponding in each stacking image is merged convolution, obtain characteristic image Block, and each characteristic image block obtained is carried out down-sampling;
Up-sampling device, for up-sampling each characteristic image block after down-sampling;
Temperature prediction means, at least one the characteristic image block after utilizing up-sampling is true or is that false probability is to draw heat The color of each pixel in degree prognostic chart;Wherein, after each pixel in described temperature prognostic chart belongs to up-sampling at least One characteristic image block.
Image temperature prognoses system the most according to claim 1, it is characterised in that described down-sampling based on convolutional network Device also includes:
Normalization unit, for being normalized several stacking images described, according still further to the same window and identical shifting Several stacking images described in after normalization are carried out piecemeal by dynamic rule.
Image temperature prognoses system the most according to claim 1 and 2, it is characterised in that described based under convolutional network Sampling apparatus comprises: the structure that least one set is made up of filter and downsampling unit;Wherein, the filter number in described structure Amount is at least one, each filter cascade when the number of filters in described structure is multiple, in same group of structure;Work as institute Stating the quantity of structure when being multiple, each group structure cascades with one another;
Wherein, there is with the filter that described image received device connects the reception passage consistent with stacking amount of images, described Receive passage and transmit a width stacking image;Described filter is for each image by the same window position corresponding in each stacking image Block merges convolution, obtains should the characteristic image block of the window's position;
The output channel cascade receiving passage and previous stage filter of filter in other groups, for the characteristic pattern that will be received As block carries out full convolution;
The output channel of every grade of filter is provided with described downsampling unit, under the characteristic image block received is carried out Sampling, and the characteristic image block after down-sampling is sent into next stage filter or the reception passage of up-sampling device.
Image temperature prognoses system the most according to claim 3, it is characterised in that described be connected with image received device Filter for according toBy the same window corresponding in each stacking image Each image block of position merges convolution;Wherein, WkFor kth convolution kernel, bkFor corresponding kth convolution kernel at each passage On the side-play amount that adds, c is the port number of stacking image, and (u v) is WkSize, (i j) is pixel position in image block.
Image temperature prognoses system the most according to claim 3, it is characterised in that the filter in each structure includes following At least one: filter based on color, filter based on lines and filter based on gray scale.
Image temperature prognoses system the most according to claim 1, it is characterised in that described up-sampling device is for according to institute State window size, the characteristic image block received is carried out up-sampling based on quadratic interpolation.
Image temperature prognoses system the most according to claim 6, it is characterised in that described up-sampling device includes:
Up-sampling unit, for based on each pixel value in the characteristic image block received, is filled in default expansion window, To the characteristic image block once up-sampled;
Smoothing processing unit, for based on the default convolution kernel consistent with window size, to the described feature once up-sampled Image block carries out convolution, obtains the characteristic image block of the secondary up-sampling of corresponding described window size.
Image temperature prognoses system the most according to claim 1, it is characterised in that described temperature prediction means for based on Through learn in advance and obtain comprise true value weight and the loss function of falsity weight, corresponding to each pixel of temperature prognostic chart The each characteristic image block received carries out temperature evaluation one by one, and draws the prediction of described temperature based on obtained temperature evaluation Figure.
Image temperature prognoses system the most according to claim 8, it is characterised in that described loss function is based on learning in advance Practise and obtain and the cross entropy function that comprises true value weight and falsity weight.
Image temperature prognoses system the most according to claim 1, it is characterised in that several stacking view data described are Continuous print odd number image.
11. 1 kinds of lung tumor identification systems based on convolutional neural networks, it is characterised in that including:
Image library, for spatially several lung CT stacking images of sequential storage;
Image temperature prognoses system as described in arbitrary in claim 1-10;
Described image temperature prognoses system is according to corresponding several lung CT described of several lung CT stacking images output received The temperature prognostic chart of stacking image;
Control device, for judging whether described image temperature prognoses system is extracted all lung CT stacking figures stored Picture, if it is not, then control described image temperature prognoses system to extract several lung CT stacking images by several times according to spatial order, until Till determining that described image acquiring apparatus is extracted all lung CT stacking images stored.
12. lung tumor identification systems based on convolutional neural networks according to claim 11, it is characterised in that described Control device span based on piece image and extract several CT stacking images by several times.
13. 1 kinds of image temperature Forecasting Methodologies, it is characterised in that including:
Receive along presetting several stacking images that dimension gathers;
Described many good fortune stacking image is sent in downsampling device based on convolutional network, be described adopt based under convolutional network Each described stacking image, according to the same window and identical movement rule, is carried out piecemeal by sampling device, and by each stacking image The image block of corresponding the same window position merges convolution, to obtain characteristic image block, and enters each characteristic image block Row down-sampling;
Each characteristic image block after down-sampling is carried out up-sampling process;
At least one the characteristic image block after up-sampling is utilized to be true or be that false probability is to each picture drawing in temperature prognostic chart The color of vegetarian refreshments;Wherein, at least one the characteristic image block after each pixel in described temperature prognostic chart belongs to up-sampling.
14. image temperature Forecasting Methodologies according to claim 13, it is characterised in that receiving along presetting what dimension gathered After several stacking images, also include:
Several stacking images described are normalized;
According to the same window and identical movement rule, several stacking images described in after normalization are carried out piecemeal.
15. according to the image temperature Forecasting Methodology described in claim 13 or 14, it is characterised in that described based on convolutional network Downsampling device includes: the structure that least one set is made up of filter and downsampling unit;Wherein, the filter in described structure Quantity is at least one, each filter cascade when the number of filters in described structure is multiple, in same group of structure;When When the quantity of described structure is multiple, each group structure cascades with one another;
Wherein, there is with the filter that described image received device connects the reception passage consistent with stacking amount of images, described Receive passage and transmit a width stacking image;Each image block of the same window position corresponding in each stacking image is entered by described filter Row merges convolution, obtains should the characteristic image block of the window's position;
The output channel cascade receiving passage and previous stage filter of filter, the characteristic image block that will be received in other groups Carry out full convolution;
Described downsampling unit in the output channel of every grade of filter, carries out down-sampling to the characteristic image block received, And the characteristic image block after down-sampling is sent into next stage filter or carries out up-sampling operation.
16. image temperature Forecasting Methodologies according to claim 15, it is characterised in that described be connected with image received device Filter each image block of the same window position corresponding in each stacking image is merged convolution, obtain should window position The characteristic image block put, including: described filter according toBy each layer In folded image, each image block of corresponding the same window position merges convolution;Wherein, WkFor kth convolution kernel, bkFor corresponding the The side-play amount that k convolution kernel adds on each passage, c is the port number of stacking image, and (u v) is WkSize, (i j) is image Pixel position in block.
17. image temperature Forecasting Methodologies according to claim 15, it is characterised in that the filter in each structure include with Descend at least one: filter based on color and/or filter based on lines and filter based on gray scale.
18. image temperature Forecasting Methodologies according to claim 13, it is characterised in that described by each feature after down-sampling Image block carries out up-sampling process, including: according to described window size, the characteristic image block received is carried out based on secondary-plug The up-sampling of value.
19. image temperature Forecasting Methodologies according to claim 18, it is characterised in that described according to window size, to institute The characteristic image block received carries out up-sampling based on quadratic interpolation and includes:
Based on each pixel value in the characteristic image block received, it is filled in default expansion window, once to be up-sampled Characteristic image block;
Based on the default convolution kernel consistent with window size, the described characteristic image block once up-sampled is carried out convolution, The characteristic image block up-sampled to the secondary of corresponding described window size.
20. image temperature Forecasting Methodologies according to claim 13, it is characterised in that after described utilization up-sampling at least One characteristic image block be true or be false probability to draw the color of each pixel in temperature prognostic chart, including: based on warp Learn in advance and obtain comprise true value weight and the loss function of falsity weight, in each pixel correspondence of temperature prognostic chart After sampling each characteristic image block carry out temperature evaluation one by one, and it is pre-to draw described temperature based on obtained temperature evaluation Mapping.
21. image temperature Forecasting Methodologies according to claim 20, it is characterised in that described loss function is based in advance Study and obtain and the cross entropy function that comprises true value weight and falsity weight.
22. image temperature Forecasting Methodologies according to claim 13, it is characterised in that several stacking view data described are Continuous print odd number image.
23. 1 kinds of lung tumor identification methods based on convolutional neural networks, it is characterised in that including:
The most spatially several lung CT stacking images of sequential storage;
According to described image temperature Forecasting Methodology arbitrary in claim 13-22, according to several lung CT stackings received The temperature prognostic chart of corresponding several lung CT stacking images described of image output;
Judge whether described image temperature prognoses system is extracted all lung CT stacking images stored, if it is not, then control Described image temperature prognoses system extracts several lung CT stacking images, until determining that described image carries by several times according to spatial order Fetching is put till being extracted all lung CT stacking images stored.
24. lung tumor identification methods based on convolutional neural networks according to claim 23, it is characterised in that described Control device span based on piece image and extract several CT stacking images by several times.
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 true CN106203327A (en) 2016-12-07
CN106203327B 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)

Cited By (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106778785A (en) * 2016-12-23 2017-05-31 东软集团股份有限公司 Build the method for image characteristics extraction model and method, the device of image recognition
CN106778867A (en) * 2016-12-15 2017-05-31 北京旷视科技有限公司 Object detection method and device, neural network training method and device
CN106920227A (en) * 2016-12-27 2017-07-04 北京工业大学 Based on the Segmentation Method of Retinal Blood Vessels that deep learning is combined 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
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
CN107169485A (en) * 2017-03-28 2017-09-15 北京捷通华声科技股份有限公司 A kind of method for identifying mathematical formula and device
CN107292884A (en) * 2017-08-07 2017-10-24 北京深睿博联科技有限责任公司 The method and device of oedema and hemotoncus in a kind of identification MRI image
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
CN107437110A (en) * 2017-07-11 2017-12-05 中国科学院自动化研究所 The piecemeal convolution optimization method and device of convolutional neural networks
CN107665491A (en) * 2017-10-10 2018-02-06 清华大学 The recognition methods of pathological image and system
CN108022238A (en) * 2017-08-09 2018-05-11 深圳科亚医疗科技有限公司 Method, computer-readable storage medium and the system being detected to object in 3D rendering
CN108446724A (en) * 2018-03-12 2018-08-24 江苏中天科技软件技术有限公司 A kind of fusion feature sorting technique
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
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
CN109800873A (en) * 2019-01-29 2019-05-24 北京旷视科技有限公司 Image processing method and device
WO2020019739A1 (en) * 2018-07-24 2020-01-30 深圳先进技术研究院 Vascular wall plaque segmentation method and apparatus, and computer-readable storage medium

Citations (5)

* 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
US20150139485A1 (en) * 2013-11-15 2015-05-21 Facebook, Inc. Pose-aligned networks for deep attribute modeling
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

Patent Citations (5)

* 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
US20150139485A1 (en) * 2013-11-15 2015-05-21 Facebook, Inc. Pose-aligned networks for deep attribute modeling
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

Cited By (25)

* 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
CN106778867A (en) * 2016-12-15 2017-05-31 北京旷视科技有限公司 Object detection method and device, neural network training method and device
US10671919B2 (en) 2016-12-15 2020-06-02 Beijing Kuangshi Technology Co., Ltd. Target detection method and device, 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
CN106778785A (en) * 2016-12-23 2017-05-31 东软集团股份有限公司 Build the method for image characteristics extraction model and method, the device of image recognition
CN106920227A (en) * 2016-12-27 2017-07-04 北京工业大学 Based on the Segmentation Method of Retinal Blood Vessels that deep learning is combined with conventional method
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
CN107169485A (en) * 2017-03-28 2017-09-15 北京捷通华声科技股份有限公司 A kind of method for identifying mathematical formula 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
CN107437110A (en) * 2017-07-11 2017-12-05 中国科学院自动化研究所 The piecemeal convolution optimization method and device of convolutional neural networks
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
CN107292884A (en) * 2017-08-07 2017-10-24 北京深睿博联科技有限责任公司 The method and device of oedema and hemotoncus in a kind of identification MRI image
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
CN108022238A (en) * 2017-08-09 2018-05-11 深圳科亚医疗科技有限公司 Method, computer-readable storage medium and the system being detected to object in 3D rendering
CN107665491A (en) * 2017-10-10 2018-02-06 清华大学 The recognition methods of pathological image and system
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
CN108446724A (en) * 2018-03-12 2018-08-24 江苏中天科技软件技术有限公司 A kind of fusion feature sorting technique
WO2020019739A1 (en) * 2018-07-24 2020-01-30 深圳先进技术研究院 Vascular wall plaque segmentation method and apparatus, and computer-readable storage medium
WO2020019740A1 (en) * 2018-07-24 2020-01-30 深圳先进技术研究院 Left ventricle myocardium segmentation method and device, and computer readable storage medium
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
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
CN109800873A (en) * 2019-01-29 2019-05-24 北京旷视科技有限公司 Image processing method and device
CN109800873B (en) * 2019-01-29 2021-03-23 北京旷视科技有限公司 Image processing method and device

Also Published As

Publication number Publication date
CN106203327B (en) 2019-04-19

Similar Documents

Publication Publication Date Title
CN106940816B (en) CT image pulmonary nodule detection system based on 3D full convolution neural network
US20200175352A1 (en) Structure defect detection using machine learning algorithms
Nhat-Duc et al. Automatic recognition of asphalt pavement cracks using metaheuristic optimized edge detection algorithms and convolution neural network
Bayramoglu et al. Deep learning for magnification independent breast cancer histopathology image classification
CN104680508B (en) Convolutional neural networks and the target object detection method based on convolutional neural networks
Gecer et al. Detection and classification of cancer in whole slide breast histopathology images using deep convolutional networks
CN106096670B (en) Concatenated convolutional neural metwork training and image detecting method, apparatus and system
CN105849747B (en) Method and system for facial image identification
CN106570477B (en) Vehicle cab recognition model building method and model recognizing method based on deep learning
Yang et al. Building extraction at scale using convolutional neural network: Mapping of the united states
CN106358444B (en) Method and system for face verification
WO2017215622A1 (en) Object segmentation method and apparatus and computing device
CN105912990B (en) The method and device of Face datection
Suzuki et al. Linear-time connected-component labeling based on sequential local operations
CN108154508B (en) Method, apparatus, storage medium and the terminal device of product defects detection positioning
CN102308306B (en) A constraint generator for use in image segregation
CN102169587B (en) Device and method for image processing
CN106780612B (en) Object detecting method and device in a kind of image
CN104424634B (en) Object tracking method and device
Utans et al. Selecting neural network architectures via the prediction risk: Application to corporate bond rating prediction
CN107680678A (en) Based on multiple dimensioned convolutional neural networks Thyroid ultrasound image tubercle auto-check system
CN106548127A (en) Image-recognizing method
CN106156781B (en) Sort convolutional neural networks construction method and its image processing method and device
CN1894703B (en) Pattern recognition method and device
CN107704857A (en) A kind of lightweight licence plate recognition method and device end to end

Legal Events

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