CN110246567A - A kind of medical image preprocess method - Google Patents
A kind of medical image preprocess method Download PDFInfo
- Publication number
- CN110246567A CN110246567A CN201810186353.3A CN201810186353A CN110246567A CN 110246567 A CN110246567 A CN 110246567A CN 201810186353 A CN201810186353 A CN 201810186353A CN 110246567 A CN110246567 A CN 110246567A
- Authority
- CN
- China
- Prior art keywords
- image
- medical image
- sample
- exposure mask
- information
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H30/00—ICT specially adapted for the handling or processing of medical images
- G16H30/20—ICT specially adapted for the handling or processing of medical images for handling medical images, e.g. DICOM, HL7 or PACS
Landscapes
- Health & Medical Sciences (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Radiology & Medical Imaging (AREA)
- Engineering & Computer Science (AREA)
- Epidemiology (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Primary Health Care (AREA)
- Public Health (AREA)
- Measuring And Recording Apparatus For Diagnosis (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of medical image preprocess methods, comprising the following steps: label information semantic image: pre-read digital medical image and its label information, and the text label information of tape format is converted to multilayer subseries mask image using distinguished number；Interest extracted region: reading digital medical image, removes transparency channel and obtains image, extracts the tissue regions profile in current slice image, divide an image into tissue regions and background area；More exposure mask sample classifications extract: utilizing multilayer subseries exposure mask generated, positive sample and negative sample are extracted in the tissue regions, and sample data information is packaged, form the structural data that can be applied to neural network model training and prediction.By the technical program, more efficient and more accurate data prediction may be implemented.
Description
Technical field
The present invention relates to technical field of medical image processing more particularly to a kind of medical image preprocess methods.
Background technique
The following contents is only the background introduction about present techniques of inventor's understanding, not necessarily constitutes existing skill
Art.
Computer-aided diagnosis (computer aided diagnosis, CAD) refers to by iconography, medical image
Reason technology and other possible physiology, biochemical apparatus, in conjunction with the analytical calculation of computer, auxiliary discovery lesion improves diagnosis
Accuracy rate method, the extensive use of " the third eye " of the doctor that is otherwise known as, CAD system helps to improve diagnosis
Sensibility and specificity.
In order to accurately and efficiently utilize these information, the computer-aided diagnosis research based on cancer medical image becomes
Industry hot spot, research and most widely used field are to manage medical image by machine learning and deep learning at present
Solution and identification, the computer-aided diagnosis based on machine learning mainly include the content of four aspects: (1) image preprocessing；(2) feel
The segmentation of region of interest (ROI)；(3) feature extraction, selection and classification；(4) identification (classification or segmentation) of tumor region.Its
In, the ultrahigh resolution in image preprocessing part, pathology medical image proposes huge challenge to pretreatment mode.
The existing Classification and Identification technology based on image segmentation is generally used for the Pathologic image analysis under small resolution ratio, there is no
Method effectively handles the super-resolution digital medical image of such huge data volume.In addition, in medical image recognition task, one
Sample used in aspect includes puncture sample used in postoperative large slice sample and early screening, on the other hand, each number
The form organized in word slice, area accounting is different, so that the calculation amount of sample extraction and accuracy also become a pair of
It is difficult to the contradiction balanced.Front end and input data source as artificial neural network, how efficiently and rapidly to medical image number
According to being pre-processed, have become one of the project needed to be studied in field of medical imaging.
Summary of the invention
For overcome the deficiencies in the prior art, technical problem solved by the invention be to provide one kind may be implemented efficiently,
The accurately medical image preprocess method of data prediction.
In order to solve the above technical problems, the technical solution adopted in the present invention content is specific as follows:
A kind of medical image preprocess method, comprising the following steps:
Label information semantic image: pre-read digital medical image and its label information, it will be with lattice using distinguished number
The text label information of formula is converted to multilayer subseries mask image；
Interest extracted region: reading digital medical image, removes transparency channel and obtains image, extracts current slice image
In tissue regions profile, divide an image into tissue regions and background area；
More exposure mask sample classifications extract: utilizing multilayer subseries exposure mask generated, sun is extracted in the tissue regions
Property sample and negative sample, and sample data information is packaged, formation can be applied to neural network model training and prediction
Structural data.
Efficiently and rapidly medical image is pre-processed to realize, inventor uses doctor in the technical scheme
The coordinate mapping policy between image pyramid multiresolution level is learned, the identification exposure mask of two kinds of granularities of thickness is constructed, passes through seat
Mark establishes connection, is respectively completed the target for carrying out quickly positioning to tissue regions and accurately dividing.
Compared to other modes, the quick positioning and accurate division that tissue regions are carried out in the technical program, Neng Goushi
Now efficiently and rapidly medical image is pre-processed.
In one or more embodiments, in the label information semantic image step:
More specifically, the transparency channel is the channel Alpha；
More specifically, the format of the text label is XML format；
It should be noted that the format of text label can be using a variety of, in a kind of embodiment, which is
XML format, remaining embodiment can be according to actual needs using other corresponding different formats.
More specifically, the distinguished number is closed polygon coordinate distinguished number.
It should be noted that in the present embodiment, the distinguished number uses closed polygon coordinate distinguished number, phase
Compared with other distinguished numbers, directly differentiated between pixel and arest neighbors polygon vertex due to having based on similar triangle theory
Positional relationship the characteristics of, therefore can quickly determine each pixel whether be closed polygon encirclement, to quickly be converted
For multilayer subseries mask image, the operation efficiency under the step is improved, and then improves the efficiency of data prediction.
In one or more more specific embodiments, the label information semantic image step is specifically included:
The resolution information for loading the digital medical image constructs the zero moment of same size according to the resolution information
Battle array is as blank exposure mask of equal value；
Load the label information of XML format corresponding to the digital medical image, the label information point of the XML format
The coordinate information of the closed polygon of several tab areas is not had recorded；
Determine whether pixel is in any one closed polygon using closed polygon coordinate distinguished number, and according to
The text label data of XML format are converted to mask image by this, and obtaining includes multi-level point for marking exposure mask and rejecting exposure mask
Class exposure mask.
It is described in the closed polygon coordinate distinguished number as one or more more specific embodiments, institute
The positional relationship for stating pixel and closed polygon is obtained by following formula:
In_ploy=(E1y-Py)(E1x-E0x)-(E1x-Px)(E1y-E0y)
Wherein E0、E1Indicate two endpoints of the closed polygon a line, x, y indicate its transverse and longitudinal coordinate；P is indicated
The pixel for needing to judge.
It should be noted that the formula need to only utilize polygon with the closing of the pixel arest neighbors for a certain pixel
The coordinate on two vertex of a line of shape, can directly determine whether it is located at this polygonal internal.Compared to other modes,
This formula, which can be realized more rapidly, is converted into multilayer subseries mask image, is improved under the step to a greater extent
Operation efficiency, and then the efficiency of data prediction is improved to a greater extent, to realize the goal of the invention of this programme.
In one or more embodiments, in the label information semantic image step further include: the unified number
The mark of word medical image and the corresponding mask image, and Correctness checking is carried out by preloading.
In a kind of specific Application Example, for each digital medical image, it is based on the matched side of filename
Method re-scales its multilayer subseries exposure mask, such as the label exposure mask being mentioned below and the path for rejecting exposure mask, automatically to all
Source digital medical image and its subsidiary mask image are preloaded, and the digital medical image file that can not be read is marked,
If it exists, it attempts to regenerate mask image, it will slice file and the removal of XML tag file if failure.By this step into
Row Correctness checking avoids subsequent place so as to reach the quick self-checking effect to all original pending data correctness
It is interrupted during reason because data can not load.
In one or more embodiments, in the interest region extraction step:
More specifically, the method for reading the digital medical image is read using Openslide；
More specifically, described image is RGB image；
It should be noted that the image in the technical program can be RGB image but it is also possible to be the figure of extended formatting
Picture, depending on depending on specific embodiment.
More specifically, further including using color gamut space transfer, corrosion and expansion, to extract tissue regions after obtaining image
Profile.
It should be noted that can quickly be defined in medical image by preset color threshold by this technical characteristic
Tissue regions, large stretch of tissue regions in image further can be directly oriented under large scale level, to improve emerging
The efficiency of interesting region extraction step, and then the technical program is improved for the efficiency of data prediction.
In one or more more specific embodiments, the interest region extraction step is specifically included:
The digital medical image is read using Openslide, the channel Alpha is removed and obtains RGB image, arrived using RGB
Space transfer, corrosion core and the processing for expanding core of HSV colour gamut, demarcate the tissue outer profile conduct in the digital medical image
Background area, while the interest identification region ROI using the external limitting casing of tissue regions as the digital medical image.
In a kind of specific Application Example, RGB is obtained using Openslide load source digital medical image S and is schemed
Picture removes transparency channel, obtains RGB image, is then converted into HSV space from rgb space, obtains the two of digital medical image
Value image M0, the scattered cavity in large stretch of tissue area is filled up using expansion core, recycles corrosion core to eliminate noise, is partitioned into solely
Vertical tissue regions element, final extract obtain the profile of tissue regions and tissue mask image M in medical image1, background
It is distinguished with tissue regions.Meanwhile in M1On its external limitting casing is generated to each white block of separation.
In one or more embodiments, in more exposure mask sample classification extraction steps: by sample number it is believed that
It is by being realized using TFrecords that breath, which is packaged,.
It should be noted that the information such as sample metadata, label are directly encapsulated into a record by TFrecords, and
A large amount of record can be packaged into an independent file, reduce code redundancy, the I/O load of system is greatly reduced, to reach
To load resource is saved, data loading efficiency is improved, and then makes the promotion of data-handling efficiency indirectly.
In one or more embodiments, in more exposure mask sample classification extraction steps: the sample number it is believed that
Breath includes one of the characteristic information of sample, source-information, position coordinates, image information, label data or a variety of；
More specifically, the characteristic information includes one of sample file name, sample store path or a variety of；It is described next
Source information includes the filename of samples sources digital medical image；The position coordinates include on original image Level-0 coordinate system
Centre coordinate；Described image information includes sample image data file；The label data includes sample label.
It should be noted that characteristic information is to determine specific path position of the sample in data set library；Source-information
To determine the corresponding relationship of sample and source medical image, both the above information is used to trace to the source to the subsequent sample that leaves a question open；
Position coordinates are to trace specific location of the sample in corresponding medical image that leave a question open；Image information is sample image, as base
This content；Label data has determined the positive negativity of the sample.
In one or more more specific embodiments, more exposure mask sample classification extraction steps are specifically included:
The result that the label exposure mask and the rejecting exposure mask subtract each other identifies exposure mask as diseased region, scans institute using sliding window
When stating the limitting casing of interest identification region ROI, first identify that exposure mask removes the rejecting exposure mask according to tissue regions, then according to disease
Become area's identification exposure mask and tissue regions are divided into positive and feminine gender, extracts positive sample and negative sample respectively.
As one or more more specific embodiments, during demarcating tissue outer profile, retain colour gamut transformation
Resulting first binary map identifies exposure mask as the tissue regions under small scale after operation, is scanning the limitting casing using sliding window
When, while the white pixel accounting that the tissue regions under same position under small scale identify exposure mask is calculated by coordinate transformation, lead to
Empty background area that may be present in preset threshold removal tissue is crossed, tissue regions sample is obtained.
It should be noted that due to original image enormous size, identification exposure mask (being binary map) obtained in basic scheme is
The lower binary map of fineness under large scale, and when extracting sample, sliding window is slided in original image, can be potentially encountered large scale two
It is worth unrecognized tiny white space under figure.So through the above scheme, the essence being achieved that under more tiny scale
True tissue regions sample.
Compared with prior art, the beneficial effects of the present invention are:
1, medical image preprocess method of the invention, using the seat between medical image pyramid multiresolution level
Mapping policy is marked, the identification exposure mask of two kinds of granularities of thickness is constructed, is established and is contacted by coordinate, is respectively completed and tissue regions is carried out
The target for quickly positioning and accurately dividing, efficiently and rapidly pre-processes medical image so as to realize, in turn
Reach and data prediction efficiently, accurately is carried out to medical image.
2, medical image preprocess method of the invention, using closed polygon coordinate distinguished number by the text of tape format
Label information be converted to multilayer subseries mask image due to have based on similar triangle theory directly differentiates pixel with it is nearest
The characteristics of positional relationship between adjacent polygon vertex, therefore can quickly determine whether each pixel is closed polygon encirclement,
To quickly be converted into multilayer subseries mask image, the operation efficiency under the step is improved, and then improve data
Pretreated efficiency.
3, medical image preprocess method of the invention, in label information semantic image step further include: in unification
The mark of digital medical image and the corresponding mask image is stated, and carries out Correctness checking by preloading；Pass through this
Step carries out Correctness checking and avoids so as to reach the quick self-checking effect to all original pending data correctness
In subsequent processes because data can not load interrupt caused by processing pause, improve data processing operation fluency and effect
Rate.
4, medical image preprocess method of the invention, obtain image after, further include using color gamut space transfer, corrosion and
Expansion, to extract tissue regions profile；Medical image can quickly be defined by preset color threshold by this technology feature
In tissue regions, large stretch of tissue regions in image further can be directly oriented under large scale level, to improve
The efficiency of interest region extraction step, and then the technical program is improved for the efficiency of data prediction.
5, medical image preprocess method of the invention, by sample data in more exposure mask sample classification extraction steps
It is by being realized using TFrecords that information, which is packaged,；Due to meeting in encapsulation step during Medical Image Processing
A large amount of code redundancies are generated, are handled in this scheme using TFrecords, the I/O load of system can be greatly reduced,
To reach saving load resource, data loading efficiency is improved, and then makes the promotion of data-handling efficiency indirectly.
6, medical image preprocess method of the invention retains colour gamut transformation behaviour during demarcating tissue outer profile
Resulting first binary map identifies exposure mask as the tissue regions under small scale after work, obtains the tissue regions sample of different finenesses
This；By this programme, the accurate tissue regions sample being achieved that under more tiny scale, so that identification is more smart
Really, the accuracy of image procossing is improved.
The above description is only an overview of the technical scheme of the present invention, in order to better understand the technical means of the present invention,
And it can be implemented in accordance with the contents of the specification, and in order to allow above and other objects, features and advantages of the invention can
It is clearer and more comprehensible, it is special below to lift preferred embodiment, and cooperate attached drawing, detailed description are as follows.
Detailed description of the invention
Fig. 1 is color gamut space transfer in interest region extraction step of the present invention, burn into expansion and outer profile generation phase
Schematic diagram；
Fig. 2 is diseased region exposure mask in the more exposure mask sample classification extraction steps of the present invention, rejects exposure mask, diseased region identification exposure mask
And normal area identifies exposure mask generation phase schematic diagram；
Fig. 3 is sample of the present invention multi-subarea extracting schematic diagram；
Fig. 4 is closed polygon coordinate diagnostic method calculating process signal in label information semantic image step of the present invention
Figure；
Fig. 5 is a kind of frame flow diagram of more preferably embodiment of the present invention.
Specific embodiment
It is of the invention to reach the technical means and efficacy that predetermined goal of the invention is taken further to illustrate, below in conjunction with
Attached drawing and preferred embodiment, to specific embodiment, structure, feature and its effect according to the present invention, detailed description are as follows:
Embodiment 1
The present embodiment provides medical image preprocess methods of the present invention comprising following steps:
Label information semantic image: pre-read digital medical image and its label information, it will be with lattice using distinguished number
The text label information of formula is converted to multilayer subseries mask image；
Interest extracted region: reading digital medical image, removes transparency channel and obtains image, extracts current slice image
In tissue regions profile, divide an image into tissue regions and background area, as shown in Figure 1；
More exposure mask sample classifications extract: utilizing multilayer subseries exposure mask generated, sun is extracted in the tissue regions
Property sample and negative sample, as shown in Figures 2 and 3；And be packaged sample data information, formation can be applied to neural network
The structural data of model training and prediction.
It is the basic embodiment of the technical program one of which above.In present embodiment, inventor is in this skill
The coordinate mapping policy between medical image multiresolution level is used in art scheme, the identification of building two kinds of granularities of thickness is covered
Film is established by coordinate and is contacted, and the target for carrying out quickly positioning to tissue regions and accurately dividing is respectively completed, by tissue
The quick positioning and accurate division that region carries out, can be realized and efficiently and rapidly pre-process to medical image.
Embodiment 2
The present embodiment is a kind of preferred embodiment on the basis of above-described embodiment 1, the present embodiment 2 and above-mentioned reality
The difference for applying example is: having in the present embodiment, in the label information semantic image step following one of or more
Kind of preferred embodiment, these embodiments individually or can combine and implemented:
In some embodiments, the transparency channel is the channel Alpha.
In some embodiments, the format of the text label is XML format.The format of text label can use
A variety of, in a kind of embodiment, which is XML format, remaining embodiment can use other phases according to actual needs
Answer different formats.
In some embodiments, the distinguished number is closed polygon coordinate distinguished number.In present embodiment
In, the distinguished number uses closed polygon coordinate distinguished number, compared to other distinguished numbers, due to having based on similar
Triangle Principle directly differentiates the characteristics of positional relationship between pixel and arest neighbors polygon vertex, therefore can quickly determine
Whether each pixel is closed polygon encirclement, to quickly be converted into multilayer subseries mask image, improves the step
Under operation efficiency, and then improve the efficiency of data prediction.
In one or more more specific embodiments, the label information semantic image step is specifically included:
The resolution information for loading the digital medical image constructs the zero moment of same size according to the resolution information
Battle array is as blank exposure mask of equal value；
Load the label information of XML format corresponding to the digital medical image, the label information point of the XML format
The coordinate information of the closed polygon of several tab areas is not had recorded；
Determine whether pixel is in any one closed polygon using closed polygon coordinate distinguished number, and according to
The text label data of XML format are converted to mask image by this, and obtaining includes multi-level point for marking exposure mask and rejecting exposure mask
Class exposure mask.
It is described in the closed polygon coordinate distinguished number as one or more more specific embodiments, such as
Shown in Fig. 4, the positional relationship of the pixel and closed polygon is obtained by following formula:
In_ploy=(E1y-Py)(E1x-E0x)-(E1x-Px)(E1y-E0y)
Wherein E0、E1Indicate two endpoints of the closed polygon a line, x, y indicate its transverse and longitudinal coordinate；P is indicated
The pixel for needing to judge.
For a certain pixel, which need to only utilize two with a line of the closed polygon of the pixel arest neighbors
The coordinate on a vertex, can directly determine whether it is located at this polygonal internal.Compared to other modes, this formula can be more
It rapidly realizes and is converted into multilayer subseries mask image, improve the operation efficiency under the step to a greater extent, into
And the efficiency of data prediction is improved to a greater extent, to realize the goal of the invention of this programme.
In a kind of specific Application Example, when P point is located at E in the vertical direction0、E1Between when (E0y≤Py≤
E1yOr E0y≥Py≥E1y), the value of in_ploy is calculated by above formula.Work as in_ploy=0, indicates pixel on side；Work as in_
Ploy < 0 indicates that pixel is located at the left side on side；Work as in_ploy > 0, indicates that pixel is located at the right on side.
For each of null matrix element, all closed polygons and its all sides are traversed, determines square according to the following formula
The positional relationship of pixel and each closed polygon represented by array element element:
Wherein, i indicates i-th side of certain closed polygon, and N indicates the number of edges of closed polygon, and in_ploy indicates i-th
Side the differentiation of pixel is contributed.
Work as IN_PLOY=0, then the pixel is the interior point of closed polygon, and it is white that pixel value is updated to (255,255,255)
Color.
According to the above method, the text label data of XML format are converted into mask image, obtain label exposure mask and rejecting
Exposure mask (background area).
By above-mentioned technical approach, it can be quickly and accurately positioned out identified tab area in XML tag data, it will
Text label data are converted to region recognition mask image.
In one or more embodiments, in the label information semantic image step further include: unified above-mentioned number
The mark of word medical image and the corresponding mask image, and Correctness checking is carried out by preloading.
In a kind of specific Application Example, for each digital medical image, it is based on the matched side of filename
Method re-scales its multilayer subseries exposure mask, such as the label exposure mask being mentioned below and the path for rejecting exposure mask, automatically to all
Source digital medical image and its subsidiary mask image are preloaded, and the digital medical image file that can not be read is marked,
If it exists, it attempts to regenerate mask image, it will slice file and the removal of XML tag file if failure.By this step into
Row Correctness checking avoids subsequent place so as to reach the quick self-checking effect to all original pending data correctness
It is interrupted during reason because data can not load.
Remaining embodiment of the present embodiment is same as the previously described embodiments, and all embodiments cited by the present embodiment are equal
It can be combined implementation individually or with above-described embodiment 1, constitute different embodiments, be not repeated herein.
Embodiment 3
The present embodiment is a kind of preferred embodiment on the basis of above-described embodiment 1, the present embodiment 3 and above-mentioned reality
The difference for applying example is: in the present embodiment, having following one or more of them preferred in the interest region extraction step
Embodiment, these embodiments can individually and also combine implemented:
In some embodiments, the method for reading the digital medical image is read using Openslide.
In some embodiments, described image is RGB image.Image in the technical program can be RGB image,
It may also be the image of extended formatting, depending on depending on specific embodiment.
It further include using color gamut space transfer, corrosion and expansion, to mention after obtaining image in some embodiments
Take tissue regions profile.By this technical characteristic, the tissue in medical image can be quickly defined by preset color threshold
Region further can directly orient large stretch of tissue regions in image, to improve interest region under large scale level
The efficiency of extraction step, and then the technical program is improved for the efficiency of data prediction.
In some embodiments, the interest region extraction step is specifically included:
The digital medical image is read using Openslide, the channel Alpha is removed and obtains RGB image, arrived using RGB
Space transfer, corrosion core and the processing for expanding core of HSV colour gamut, demarcate the tissue outer profile conduct in the digital medical image
Background area, while the interest identification region ROI using the external limitting casing of tissue regions as the digital medical image.
In a kind of specific Application Example, RGB is obtained using Openslide load source digital medical image S and is schemed
Picture removes transparency channel, obtains RGB image, is then converted into HSV space from rgb space, obtains the two of digital medical image
Value image M0, the scattered cavity in large stretch of tissue area is filled up using expansion core, recycles corrosion core to eliminate noise, is partitioned into solely
Vertical tissue regions element, final extract obtain the profile of tissue regions and tissue mask image M in medical image1, background
It is distinguished with tissue regions.Meanwhile in M1On its external limitting casing is generated to each white block of separation.
Remaining embodiment of the present embodiment is same as the previously described embodiments, and all embodiments cited by the present embodiment are equal
It can be combined implementation individually or with above-described embodiment 1 or 2, constitute different embodiments, be not repeated herein.
Embodiment 4
The present embodiment is a kind of preferred embodiment on the basis of above-described embodiment 1, the present embodiment 4 and above-mentioned reality
The difference for applying example is: having in the present embodiment, in more exposure mask sample classification extraction steps following one of or more
Kind of preferred embodiment, these embodiments individually or can combine and implemented:
In some embodiments, it is by being realized using TFrecords that sample data information, which is packaged,.
The information such as sample metadata, label are directly encapsulated into a record by TFRecords, and a large amount of record can be packaged into list
An only file reduces code redundancy, greatly reduces the I/O load of system, to reach saving load resource, improves data
Loading efficiency, and then make the promotion of data-handling efficiency indirectly.
In some embodiments, in more exposure mask sample classification extraction steps: the sample data information includes
One of the characteristic information of sample, source-information, position coordinates, image information, label data are a variety of；
More specifically, the characteristic information includes one of sample file name, sample store path or a variety of；It is described next
Source information includes the filename of samples sources digital medical image；The position coordinates include on original image Level-0 coordinate system
Centre coordinate；Described image information includes sample image data file；The label data includes sample label.Preferred side herein
In case, characteristic information is to determine specific path position of the sample in data set library；Source-information is to determine sample and come
The corresponding relationship of source medical image, both the above information are used to trace to the source to the subsequent sample that leaves a question open；Position coordinates are to trace
Leave a question open specific location of the sample in corresponding medical image；Image information is sample image, as basic content；Label data is true
The positive negativity of the sample is determined.
In some embodiments, more exposure mask sample classification extraction steps are specifically included:
The result that the label exposure mask and the rejecting exposure mask subtract each other identifies exposure mask as diseased region, scans institute using sliding window
When stating the limitting casing of interest identification region ROI, first identify that exposure mask removes the rejecting exposure mask according to tissue regions, then according to disease
Become area's identification exposure mask and tissue regions are divided into positive and feminine gender, extracts positive sample and negative sample respectively.
As one or more more specific embodiments, during demarcating tissue outer profile, retain colour gamut transformation
Resulting first binary map identifies exposure mask as the tissue regions under small scale after operation, when sliding window scans the limitting casing,
The white pixel accounting that the tissue regions under same position under small scale identify exposure mask is calculated by coordinate transformation simultaneously, by pre-
If threshold value removes empty background area that may be present in tissue, the tissue regions sample of different finenesses is obtained.Herein preferably
In scheme, due to original image enormous size, identification exposure mask (being binary map) obtained in basic scheme is fine under large scale
When spending lower binary map, and extracting sample, sliding window is slided in original image, and can be potentially encountered can not identify under large scale binary map
Tiny white space.So through the above scheme, the accurate tissue regions sample being achieved that under more tiny scale
This.
In one or more more specific embodiments, diseased region is generated respectively using two kinds of XML mark files and is covered
Film MobjWith rejecting exposure mask Mexc, Mobj-MexcResult as diseased region identify exposure mask Mpos, M0-(Mobj-Mexc) result
Exposure mask M is identified as normal areaneg.Then with 256 × 256 sliding window, do not scan M overlappingly1On limitting casing, while sit
Mark is mapped to diseased region identification exposure mask MposExposure mask M is identified with normal areanegOn.For positive sample, work as MposWhite pixel in window
Accounting is greater than 70%, and M0White pixel accounting is greater than 40% in the upper same size area of same position, extracts sliding window inner tissue
Block is as positive sample；For negative sample, work as MnegWhite pixel accounting is greater than 25%, and M in windowposUpper same position is same big
For white pixel accounting less than 20%, extraction sliding window inner tissue block is negative sample in zonule.In this preferred embodiment, M1 is coarse grain
More coarse tissue regions identify exposure mask, M under degree/large scaleposWith MnegFor identification more fine under fine granularity/small scale
Exposure mask is converted by the coordinate before the two and determines corresponding relationship, wherein can reach first with the former to tissue regions
Quickly positioning；Further, it is mapped on fine-grained mask location, white pixel accounting, energy is calculated on fine granularity exposure mask
Enough achieve the effect that accurately identify tissue block.
Remaining embodiment of the present embodiment is same as the previously described embodiments, and all embodiments cited by the present embodiment are equal
It can be combined implementation individually or with above-described embodiment 1 or 2 or 3, constitute different embodiments, be it as shown in Figure 5
One of.It is not repeated herein.
The above embodiment is only the preferred embodiment of the present invention, and the scope of protection of the present invention is not limited thereto,
The variation and replacement for any unsubstantiality that those skilled in the art is done on the basis of the present invention belong to institute of the present invention
Claimed range.
Claims (10)
1. a kind of medical image preprocess method, which comprises the following steps:
Label information semantic image: pre-read digital medical image and its label information, using distinguished number by tape format
Text label information is converted to multilayer subseries mask image；
Interest extracted region: reading digital medical image, removes transparency channel and obtains image, extracts in current slice image
Tissue regions profile, divides an image into tissue regions and background area；
More exposure mask sample classifications extract: utilizing multilayer subseries exposure mask generated, positive sample is extracted in the tissue regions
Sheet and negative sample, and sample data information is packaged, form the knot that can be applied to neural network model training and prediction
Structure data.
2. medical image preprocess method as described in claim 1, which is characterized in that the label information semantic imageization step
In rapid:
Preferably, the transparency channel is the channel Alpha；
Preferably, the format of the text label is XML format；
Preferably, the distinguished number is closed polygon coordinate distinguished number.
3. medical image preprocess method as claimed in claim 2, which is characterized in that the label information semantic imageization step
Suddenly it specifically includes:
The resolution information for loading the digital medical image is made according to the null matrix that the resolution information constructs same size
For blank exposure mask of equal value；
The label information of XML format corresponding to the digital medical image is loaded, the label information of the XML format is remembered respectively
The coordinate information of the closed polygon of several tab areas is recorded；
It determines whether pixel is in any one closed polygon using closed polygon coordinate distinguished number, and accordingly will
The text label data of XML format are converted to mask image, and obtaining includes that label exposure mask is covered with the multilayer subseries for rejecting exposure mask
Film.
4. medical image preprocess method as claimed in claim 3, which is characterized in that described in the closed polygon coordinate
In distinguished number, the positional relationship of the pixel and closed polygon is obtained by following formula:
In_ploy=(E1y-Py)(E1x-E0x)-(E1x-Px)(E1y-E0y)
Wherein E0、E1Indicate two endpoints of the closed polygon a line, x, y indicate its transverse and longitudinal coordinate；P indicates to need
The pixel of judgement.
5. medical image preprocess method as described in claim 1, which is characterized in that the label information semantic imageization step
In rapid further include: the mark of the unified digital medical image and the corresponding mask image, and carried out by preloading
Correctness checking.
6. medical image preprocess method as described in claim 1, which is characterized in that in the interest region extraction step:
Preferably, the method for reading the digital medical image is read using Openslide；
Preferably, described image is RGB image；
It preferably, further include using color gamut space transfer, corrosion and expansion, to extract tissue regions profile after obtaining image.
7. medical image preprocess method as claimed in claim 6, which is characterized in that the interest region extraction step is specific
Include:
The digital medical image is read using Openslide, the channel Alpha is removed and obtains RGB image, utilize RGB to HSV color
Space transfer, corrosion core and the processing for expanding core in domain, demarcate the tissue outer profile in the digital medical image as background
Region, while the interest identification region ROI using the external limitting casing of tissue regions as the digital medical image.
8. medical image preprocess method as described in claim 1, which is characterized in that more exposure mask sample classifications extract step
In rapid:
Preferably, sample data information is packaged is by being realized using TFrecords；
Preferably, the sample data information includes characteristic information, source-information, position coordinates, image information, the label of sample
One of data are a variety of；
It is highly preferred that the characteristic information includes one of sample file name, sample store path or a variety of；The source letter
Breath includes the filename of samples sources digital medical image；The position coordinates include the center on original image Level-0 coordinate system
Coordinate；Described image information includes sample image data file；The label data includes sample label.
9. such as the described in any item medical image preprocess methods of claim 3 or 8, which is characterized in that more exposure mask samples
Classification extraction step specifically includes:
The result that the label exposure mask and the rejecting exposure mask subtract each other identifies exposure mask as diseased region, is scanned using sliding window described emerging
When the limitting casing of interesting identification region ROI, first identify that exposure mask removes the background area and rejecting exposure mask institute really according to tissue regions
Then fixed rejecting region identifies that tissue regions are divided into positive and feminine gender by exposure mask according to diseased region, extracts positive sample respectively
Sheet and negative sample.
10. medical image preprocess method as claimed in claim 9, which is characterized in that
During demarcating tissue outer profile, resulting first binary map is as under small scale after retaining colour gamut transition operation
Tissue regions identify exposure mask, calculate small scale under same position when sliding window scans the limitting casing, while through coordinate transformation
Under tissue regions identification exposure mask white pixel accounting, pass through preset threshold and remove empty background area that may be present in tissue
Domain obtains tissue regions sample.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810186353.3A CN110246567B (en) | 2018-03-07 | 2018-03-07 | Medical image preprocessing method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810186353.3A CN110246567B (en) | 2018-03-07 | 2018-03-07 | Medical image preprocessing method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110246567A true CN110246567A (en) | 2019-09-17 |
| CN110246567B CN110246567B (en) | 2023-07-25 |
Family
ID=67882102
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810186353.3A Active CN110246567B (en) | 2018-03-07 | 2018-03-07 | Medical image preprocessing method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110246567B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110717908A (en) * | 2019-10-08 | 2020-01-21 | 中山大学附属第三医院 | Pathological picture processing method |
| CN111144305A (en) * | 2019-12-26 | 2020-05-12 | 京东数字科技控股有限公司 | Training data generation method and device, electronic equipment and medium |
| CN112016542A (en) * | 2020-05-08 | 2020-12-01 | 珠海欧比特宇航科技股份有限公司 | Urban waterlogging intelligent detection method and system |
| CN114757953A (en) * | 2022-06-15 | 2022-07-15 | 深圳瀚维智能医疗科技有限公司 | Medical ultrasonic image recognition method, equipment and storage medium |
| WO2023272495A1 (en) * | 2021-06-29 | 2023-01-05 | 京东方科技集团股份有限公司 | Badging method and apparatus, badge detection model update method and system, and storage medium |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPS63313279A (en) * | 1987-06-16 | 1988-12-21 | Dainippon Printing Co Ltd | Formation of mask of picture pattern |
| US20130094733A1 (en) * | 2010-07-23 | 2013-04-18 | Hirokazu Nosato | Region segmented image data creating system and feature extracting system for histopathological images |
| CN105574859A (en) * | 2015-12-14 | 2016-05-11 | 中国科学院深圳先进技术研究院 | Liver tumor segmentation method and device based on CT (Computed Tomography) image |
-
2018
- 2018-03-07 CN CN201810186353.3A patent/CN110246567B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPS63313279A (en) * | 1987-06-16 | 1988-12-21 | Dainippon Printing Co Ltd | Formation of mask of picture pattern |
| US20130094733A1 (en) * | 2010-07-23 | 2013-04-18 | Hirokazu Nosato | Region segmented image data creating system and feature extracting system for histopathological images |
| CN105574859A (en) * | 2015-12-14 | 2016-05-11 | 中国科学院深圳先进技术研究院 | Liver tumor segmentation method and device based on CT (Computed Tomography) image |
Non-Patent Citations (1)
| Title |
|---|
| RICHARD CHEN等: "Identifying Metastases in Sentinel Lymph Nodes with Deep Convolutional Neural Networks", 《ARXIV》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110717908A (en) * | 2019-10-08 | 2020-01-21 | 中山大学附属第三医院 | Pathological picture processing method |
| CN111144305A (en) * | 2019-12-26 | 2020-05-12 | 京东数字科技控股有限公司 | Training data generation method and device, electronic equipment and medium |
| CN112016542A (en) * | 2020-05-08 | 2020-12-01 | 珠海欧比特宇航科技股份有限公司 | Urban waterlogging intelligent detection method and system |
| WO2023272495A1 (en) * | 2021-06-29 | 2023-01-05 | 京东方科技集团股份有限公司 | Badging method and apparatus, badge detection model update method and system, and storage medium |
| CN114757953A (en) * | 2022-06-15 | 2022-07-15 | 深圳瀚维智能医疗科技有限公司 | Medical ultrasonic image recognition method, equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110246567B (en) | 2023-07-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US10482603B1 (en) | Medical image segmentation using an integrated edge guidance module and object segmentation network | |
| US10410353B2 (en) | Multi-label semantic boundary detection system | |
| CN110246567A (en) | A kind of medical image preprocess method | |
| Gertych et al. | Machine learning approaches to analyze histological images of tissues from radical prostatectomies | |
| Beevi et al. | Automatic mitosis detection in breast histopathology images using convolutional neural network based deep transfer learning | |
| Xie et al. | Beyond classification: structured regression for robust cell detection using convolutional neural network | |
| CN106570505B (en) | Method and system for analyzing histopathological images | |
| Mi et al. | Deep learning-based multi-class classification of breast digital pathology images | |
| Yonekura et al. | Automatic disease stage classification of glioblastoma multiforme histopathological images using deep convolutional neural network | |
| CN108230339A (en) | A kind of gastric cancer pathological section based on pseudo label iteration mark marks complementing method | |
| CN109389129A (en) | A kind of image processing method, electronic equipment and storage medium | |
| Xu et al. | Deep learning for histopathological image analysis: Towards computerized diagnosis on cancers | |
| CN104484886B (en) | A kind of dividing method and device of MR images | |
| Sigirci et al. | Detection of mitotic cells in breast cancer histopathological images using deep versus handcrafted features | |
| CN112150442A (en) | New crown diagnosis system based on deep convolutional neural network and multi-instance learning | |
| CN110991535A (en) | pCR prediction method based on multi-type medical data | |
| CN110598030A (en) | Oracle bone rubbing classification method based on local CNN framework | |
| WO2021057148A1 (en) | Brain tissue layering method and device based on neural network, and computer device | |
| Lv et al. | Nuclei R-CNN: improve mask R-CNN for nuclei segmentation | |
| Nateghi et al. | Automatic detection of mitosis cell in breast cancer histopathology images using genetic algorithm | |
| CN114550169A (en) | Training method, device, equipment and medium for cell classification model | |
| CN114600155A (en) | Weakly supervised multitask learning for cell detection and segmentation | |
| Razavi et al. | Minugan: Dual segmentation of mitoses and nuclei using conditional gans on multi-center breast h&e images | |
| Scheurer et al. | Semantic segmentation of histopathological slides for the classification of cutaneous lymphoma and eczema | |
| Pawełczyk et al. | Towards detecting high-uptake lesions from lung CT scans using deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |