CN114494496A - Automatic intracranial hemorrhage delineation method and device based on head CT flat scanning image - Google Patents

Automatic intracranial hemorrhage delineation method and device based on head CT flat scanning image Download PDF

Info

Publication number
CN114494496A
CN114494496A CN202210100541.6A CN202210100541A CN114494496A CN 114494496 A CN114494496 A CN 114494496A CN 202210100541 A CN202210100541 A CN 202210100541A CN 114494496 A CN114494496 A CN 114494496A
Authority
CN
China
Prior art keywords
value
index
determining
matching layer
delineation
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
Application number
CN202210100541.6A
Other languages
Chinese (zh)
Other versions
CN114494496B (en
Inventor
王思伦
刘涵川
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen Yiwei Medical Technology Co Ltd
Original Assignee
Shenzhen Yiwei Medical Technology Co Ltd
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 Shenzhen Yiwei Medical Technology Co Ltd filed Critical Shenzhen Yiwei Medical Technology Co Ltd
Priority to CN202210100541.6A priority Critical patent/CN114494496B/en
Publication of CN114494496A publication Critical patent/CN114494496A/en
Application granted granted Critical
Publication of CN114494496B publication Critical patent/CN114494496B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T11/00—2D [Two Dimensional] image generation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/149—Segmentation; Edge detection involving deformable models, e.g. active contour models
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10081—Computed x-ray tomography [CT]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30016—Brain

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Apparatus For Radiation Diagnosis (AREA)
  • Measuring And Recording Apparatus For Diagnosis (AREA)

Abstract

The invention provides a method and a device for automatically delineating intracranial hemorrhage based on a head CT flat scanning image, wherein the method comprises the following steps: step 1: processing the original image of the head, reading the processed CT value, adjusting the required window width level, preprocessing the CT value in the window width level, and converting the preprocessed CT value into a single-channel matrix; step 2: inputting the processed original image and the preprocessing result into a preset delineation model; and 3, step 3: predicting the sketching data based on a preset sketching model and outputting a predicted value; and 4, step 4: based on the predicted value, a lesion delineation image outputting a predicted image is obtained. The intracranial hemorrhage is accurately segmented and sketched, so that the time required by diagnosis is reduced, and the detection efficiency and the stability of accurate diagnosis are improved.

Description

Automatic intracranial hemorrhage delineation method and device based on head CT flat scanning image
Technical Field
The invention relates to the technical field of AI medical image auxiliary diagnosis, in particular to a method and a device for automatically delineating intracranial hemorrhage based on a head CT (computed tomography) flat scan image.
Background
According to statistics, from 2017, stroke is a main death cause in China, and the incidence rate is high and the first in the world. Of the many types of stroke, cerebral hemorrhage, i.e., non-traumatic hemorrhage of the brain parenchyma, is the most lethal type of stroke. The diagnosis of cerebral hemorrhage usually adopts flat scan CT, and doctors diagnose the bleeding cause by the way of bleeding position on the CT image, intracranial bleeding volume, etc. Intracranial hemorrhage generally includes a number of subcategories in addition to bleeding in the brain parenchyma. These include subdural bleeding, and epidural bleeding, ventricular bleeding, subarachnoid bleeding mainly caused by rupture of an aneurysm. Accurate bleeding segmentation and quantification becomes critical, and both rehabilitation and therapy planning rely on the diagnostic results of CT images. The manual clinical delineation is often time-consuming and has the potential for error. The ABC/2 multi-field formula is a method for estimating bleeding volume in clinic, however, when multiple bleeding focus areas (multiple bleeding sites) occur, a good calculation effect cannot be achieved. The calculation of the formula is also influenced by the lesion shape, and this method is not applicable to bleeding subtypes other than cerebral hemorrhage. The measurement method based on the convolutional neural network segmentation can achieve higher compatibility, so that a segmentation method of the convolutional neural network is needed.
In the current neural network based segmentation method, the intracranial hemorrhage segmentation is mainly based on the segmentation of the cerebral parenchymal hemorrhage. The bleeding partition other than the brain parenchyma is not ideal, and the subarachnoid bleeding partition is also difficult to be embodied. The limited data is the main reason, or the characteristic extraction of the subdural hemorrhage near the skull part is different from the general cerebral parenchyma hemorrhage, and the complete segmentation is difficult by the common method. Detection of subtypes is also important clinically. Therefore, the invention provides a more comprehensive intracranial hemorrhage segmentation method, which can automatically and accurately outline focus areas of various hemorrhage types such as cerebral hemorrhage and the like.
Disclosure of Invention
The invention provides an automatic intracranial hemorrhage delineation method and device based on a head CT flat-scan image, which are used for accurately segmenting and delineating intracranial hemorrhage, reducing the time required by diagnosis and improving the detection efficiency and the stability of accurate diagnosis.
The invention provides an intracranial hemorrhage automatic delineation method based on a head CT flat scanning image, which comprises the following steps:
step 1: processing the original image of the head, reading the processed CT value, adjusting the required window width level, preprocessing the CT value in the window width level, and converting the preprocessed CT value into a single-channel matrix;
step 2: inputting the processed original image and the preprocessing result into a preset delineation model;
and step 3: predicting the sketching data based on a preset sketching model and outputting a predicted value;
and 4, step 4: based on the predicted value, a lesion delineation image outputting a predicted image is obtained.
In one possible implementation, step 1: processing the original image of the head, reading the processed CT value, adjusting the required window width and window level, preprocessing the CT value in the window width and window level, converting the preprocessed CT value into a single-channel matrix, and comprising the following steps:
reading the CT value of the original image of the head, screening the values which accord with a first condition in the CT values for reservation, and removing the residual values to obtain the coordinate information of the skull;
setting a window level of a required window width, and further extracting a first numerical value from a CT value of a head original image;
and normalizing the extracted first numerical value, masking the coordinate information of the skull in a matrix corresponding to the set window width window level, converting the coordinate information into 0, and constructing to obtain a single-channel matrix.
In a possible implementation manner, in step 2, before inputting into the preset delineation model, the method further includes:
constructing a basic intracranial hemorrhage segmentation model, and determining a segmentation network evaluation index set;
based on the evaluation index set, evaluating the basic intracranial hemorrhage segmentation model;
determining the basic grade of the basic intracranial hemorrhage segmentation model according to the evaluation result;
when the basic grade is a conventional grade, optimizing the basic intracranial hemorrhage segmentation model to obtain a preset delineation model;
and when the basic grade is a special grade, storing the basic intracranial hemorrhage segmentation model, wherein the stored model is the preset delineation model.
In a possible implementation manner, optimizing a basic intracranial hemorrhage segmentation model to obtain a preset delineation model includes:
extracting a set to be optimized of a basic intracranial hemorrhage segmentation model from a model optimization database;
determining a special part which is not finely identified by the basic intracranial hemorrhage segmentation model, determining the delineation difficulty of the special part, and further determining a corresponding fine delineation index according to the delineation difficulty;
performing architecture analysis on the basic intracranial hemorrhage segmentation model, and determining basic information of each layer and related information between adjacent layers in the basic intracranial hemorrhage segmentation model according to an analysis result;
determining a matching layer related to the set to be optimized and the fine delineation indexes based on the basic information and the related information;
determining indexes to be optimized in a set to be optimized and involved fine delineation indexes involved in each matching layer, and constructing an improved relation between the involved indexes in each matching layer and the matching layer;
judging whether related indexes in the matching layer are unique and can not be discarded according to the improvement relation;
determining the comprehensive optimization level corresponding to the remaining related indexes in the corresponding matching layer according to the judgment result, and analyzing the improvement degree of each matching layer and the improvement effect of the improvement degree on the basic intracranial hemorrhage segmentation model based on the improvement degree model;
determining a sequential execution scheme according to the execution attribute of the matching layer, and evaluating the overall improvement of the basic intracranial hemorrhage segmentation model according to the corresponding sequential execution scheme and the improvement degree and the improvement effect of the corresponding matching layer;
according to the evaluation result, calibrating the sequential execution scheme with execution pertinence, matching the corresponding sequential execution scheme to optimize the basic intracranial hemorrhage segmentation model according to the delineation requirement, and performing first storage;
performing effect screening on the execution sequence scheme without execution pertinence to obtain an optimal execution scheme, optimizing a basic intracranial hemorrhage segmentation model according to the optimal execution scheme, and performing second storage;
the drawing difficulty is related to the identification definition and the drawing definition of the original head image.
In a possible implementation manner, determining, according to the determination result, a comprehensive optimization level corresponding to the remaining related indexes in the corresponding matching layer includes:
determining a first execution content before optimization and a second execution content after optimization of the matching layer and a first improvement content of the matching layer, which relates to indexes, based on the improvement relation;
determining the difference degree Y1 between the first execution content and the first improved content and the second execution content;
Figure BDA0003492244560000041
wherein a denotes a first execution content, B denotes a first improvement content, and C denotes a second execution content;
Figure BDA0003492244560000042
representing a similarity ratio of the first execution content and the first modified content to the second execution content;
when the difference degree Y1 is smaller than the preset degree, judging that all related indexes in the matching layer are unique and can not be discarded, reserving all related indexes of the matching layer, and determining a comprehensive optimization level X1 corresponding to the reserved indexes;
Figure BDA0003492244560000043
wherein n represents the number of all related indexes; deltaiIndicating the index weight of the ith index related to the index at the matching layer; siIndicating the index conversion parameter value of the ith index related to the index at the matching layer; gamma rayiThe adjustment parameter of the matching layer to the ith related index is represented, and the value range is [0.9,1.1 ]]；
Figure BDA0003492244560000044
A standard parameter value representing a corresponding i-th index-related index in the matching layer;
Figure BDA0003492244560000045
a standard parameter value representing the corresponding i +1 th index-related item in the matching layer; si+1Indicating the (i + 1) th index conversion parameter value related to the index at the matching layer;
otherwise, judging whether the related indexes in the matching layer are unique and can not be discarded, searching the discardable indexes, reserving the discardable indexes, and meanwhile, establishing a calling index for the discardable indexes;
determining a discard value F of the discardable indicator;
Figure BDA0003492244560000046
where n1 represents the number of discardable pointers; djIndicating the jth discardable index r in the matching layerjThe number of associated indexes;
Figure BDA0003492244560000047
indicating the jth discardable index r in the matching layerjAssociated set of metrics
Figure BDA0003492244560000051
The correlation coefficient of (a);
Figure BDA0003492244560000052
indicating the jth discardable index r in the matching layerjAssociated set of metrics
Figure BDA0003492244560000053
A corresponding discard coefficient;
setting a calling possibility label for the calling index according to the discarding value F, wherein the calling possibility label is related to the sketching part;
determining whether the calling possibility label is related to the calling possibility label or not according to the drawn part bias;
if the index exists, determining a first index weight of the non-discardable index and a second index weight of the discardable index in the corresponding matching layer, and further obtaining a corresponding comprehensive optimization grade;
and if not, determining the comprehensive optimization level corresponding to the residual related indexes in the corresponding matching layer.
In a possible implementation manner, in step 1, during the process of processing the head original image, the method further includes:
determining an initial format of the head original image;
when the initial format does not meet the identification condition, converting the initial format into a standard format and identifying;
when the recognition condition is satisfied, the recognition is continued.
In a possible implementation manner, step 3, predicting the delineation data based on a preset delineation model, and outputting a predicted value, includes:
identifying the processed original image and the preprocessing result based on a preset sketching model, automatically identifying a bleeding area in the processed original image and segmenting if a bleeding phenomenon exists;
generating a corresponding predicted value based on the recognition result and the segmentation result;
wherein, if the bleeding phenomenon exists, the bleeding area in the processed original image is automatically identified and segmented, and the method comprises the following steps:
roughly determining a first area with a bleeding phenomenon and a second area without the bleeding phenomenon, judging the positions of the first area and the second area at the head and the position relation between the first area and the second area, and further determining the current identification precision of the first area;
determining an identification circulation mode based on the preset identification precision of the preset delineation model and the current identification precision, and finely identifying the processed original image according to the identification circulation mode;
and segmenting the processed original image according to multiple fine identification results.
In a possible implementation manner, after obtaining the lesion delineation image outputting the predicted image based on the predicted value, step 4 further includes:
carrying out data conversion on an image value in a focus sketching image;
and according to the output format, carrying out format adjustment on the converted data to obtain a view result and outputting and displaying the view result.
In one possible implementation, step 4: obtaining a lesion delineation image outputting a predicted image based on the predicted value, comprising:
carrying out numerical value unified classification on the predicted values to obtain focus blocks corresponding to different numerical values, and obtaining a block matrix of each focus block;
according to the intracranial construction standard value, obtaining edge contour matrixes of the corresponding parts of different focus blocks in a normal state;
performing edge registration on the edge contour matrix and the block matrix, judging whether the edge contour matrix is completely consistent with edge elements in the block matrix, and if so, reserving the block matrix;
otherwise, determining the element position and the dispersion degree of the unmatched elements existing in the edge contour matrix, planning a first offset value, simultaneously determining the element position and the dispersion degree of the unmatched elements existing in the block matrix, and planning a second offset value;
determining first line segments which do not match elements in the block matrix, tracking initial elements and final elements of each first line segment, and determining curve changes corresponding to the first line segments;
determining the positions of the primary element and the final element, and when the primary element and the final element are on an edge line segment corresponding to the edge contour matrix, intercepting a second line segment corresponding to the edge contour matrix, and determining offset subsegments of the first line segment and the second line segment;
performing a first adjustment to the lesion block based on the offset subsegment, the first offset value, and the second offset value;
when any element of the initial element and the final element is not on the edge line segment corresponding to the edge outline matrix, determining a second position of the element not on the edge line segment, the element orientation of the element not on the edge line segment and the edge standard element, and the curve change results of all the first line segments, establishing an extension direction, and extending the corresponding first line segment according to the extension direction to obtain an extension sub-segment;
performing a second adjustment to the lesion block based on the extended sub-segment, the first offset value, and the second offset value;
and outputting to obtain a focus drawing according to the first adjustment result and the second adjustment result.
The invention provides an automatic intracranial hemorrhage delineation device based on a head CT flat scanning image, which comprises:
the preprocessing module is used for processing the head original image based on the window width and window level to obtain a single-channel matrix, and meanwhile, reading an identification standard value obtained by normalizing and converting the CT value in the single-channel matrix and preprocessing the identification standard value;
the input module is used for inputting the processed original image and the preprocessing result into a preset delineation model;
the output module is used for predicting the sketching data based on a preset sketching model and outputting a predicted value;
an obtaining module for obtaining a lesion delineation image outputting a predicted image based on the predicted value
Compared with the prior art, the invention has the following beneficial effects:
1. through setting up the window width window level that corresponds, and through carrying out the preliminary treatment to the CT value, can guarantee the degree of recognition of model on small characteristic to combine to predetermine the model of drawing, carry out accurate segmentation drawing to intracranial hemorrhage, reduce the diagnosis required time, improve detection efficiency and accurate diagnosis.
2. The basic model is analyzed, the basic information and the relevance of the layer are determined, the matching layer is determined by determining the index to be optimized and the fine delineation index, the improvement effect of the matching layer is convenient to determine, the improvement relation between the matching layer and different indexes is determined, whether the index is unique and can not be discarded is convenient to determine, the execution flow and the identification flow are effectively saved, the identification efficiency is improved, the optimization grade of each layer is determined and the improved model is combined, the improvement effect can be effectively determined, finally, different screening and storage are performed according to the execution pertinence through the sequential execution scheme, the optimization feasibility of the basic model is ensured, and the delineation accuracy is ensured.
3. The method comprises the steps of judging whether the index is unique or not through calculating the difference, and further determining comprehensive optimization levels under different judgment results, wherein the comprehensive optimization levels are determined through three modes, one mode is to calculate the comprehensive optimization levels corresponding to all related indexes, the other mode is to set a possibility label through determining a discard value, the first mode is to reduce redundant calculated amount brought by unnecessary indexes, but to effectively and reasonably use the discard indexes when determining a deviation part so as to fully ensure the rationality of the obtained comprehensive optimization levels, the other mode is to directly determine the comprehensive optimization levels of the remaining indexes, a basis is provided for obtaining an improvement effect and an improvement degree, and the sketching accuracy is improved.
4. Through classifying the predicted value in unison, and then follow-up edge registration that carries on, obtain skew field and extension subsection, adjust the focus piece, it is mainly in order to adjust the edge lines in fact, guarantee the precision of the focus piece that obtains.
Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and drawings.
The technical solution of the present invention is further described in detail by the accompanying drawings and embodiments.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention. In the drawings:
FIG. 1 is a flowchart of an automatic intracranial hemorrhage delineation method based on a head CT flat scan image according to an embodiment of the present invention;
FIG. 2 is a diagram of a blood focus attached to the skull in an embodiment of the present invention;
FIG. 3 is a system configuration of a segmentation model for intracranial hemorrhage according to an embodiment of the present invention;
FIG. 4 is a sketch of a doctor in an embodiment of the invention;
FIG. 5 is a prediction graph associated with a physician's sketch in accordance with an embodiment of the present invention;
FIG. 6 is a comparison of a physician's sketching associated with an embodiment of the present invention;
FIG. 7 is an original drawing of a hemorrhage breaking into the ventricle in an embodiment of the present invention;
FIG. 8 is a graphical representation of a predicted stroke of a bleeding into the ventricle in an embodiment of the present invention;
FIG. 9 is an artwork of a subarachnoid hemorrhage in accordance with an embodiment of the present invention;
FIG. 10 is a graphical representation of a predicted subarachnoid hemorrhage in accordance with an embodiment of the present invention;
FIG. 11 is a block diagram of an offset subsegment in accordance with an embodiment of the present invention;
fig. 12 is a structural diagram of an automatic intracranial hemorrhage delineation device based on a head CT flat scan image according to an embodiment of the present invention.
Detailed Description
The preferred embodiments of the present invention will be described in conjunction with the accompanying drawings, and it will be understood that they are described herein for the purpose of illustration and explanation and not limitation.
Example 1:
the invention provides an automatic intracranial hemorrhage delineation method based on a head CT flat scanning image, which comprises the following steps as shown in figure 1:
step 1: processing the original image of the head, reading the processed CT value, adjusting the required window width level, preprocessing the CT value in the window width level, and converting the preprocessed CT value into a single-channel matrix;
step 2: inputting the processed original image and the preprocessing result into a preset delineation model;
and step 3: predicting the sketching data based on a preset sketching model and outputting a predicted value;
and 4, step 4: based on the predicted value, a lesion delineation image outputting a predicted image is obtained.
In this embodiment, when a doctor diagnoses intracranial hemorrhage, the window width window level mainly used may be a cerebral window [40,80], when diagnosing subdural hemorrhage, the subdural window level may be used, and may be [80,200], when diagnosing subdural hemorrhage, the doctor diagnoses hemorrhage at different parts of the head, and the window width window level at the corresponding part may be used.
In this embodiment, since a general model can distinguish general bone and bleeding features, but a bleeding focus linked to the skull, as shown in fig. 2, the model has difficulty in detecting edges (both bleeding and skull are highlighted) and is recognized as negative, the implementation of the above steps 1-4 is proposed.
In this embodiment, the preset delineation model is obtained based on neural network model training.
In this embodiment, the predicted value is, for example, a value of each pixel in the head image is predicted, and the segmentation is performed according to the predicted value.
The beneficial effects of the above technical scheme are: through setting up the window width window level that corresponds, and through carrying out the preliminary treatment to the CT value, can guarantee the degree of recognition of model on small characteristic to combine to predetermine the model of drawing, carry out accurate segmentation drawing to intracranial hemorrhage, reduce the diagnosis required time, improve detection efficiency and accurate diagnosis.
Example 2:
based on example 1, step 1: processing the original image of the head, reading the processed CT value, adjusting the required window width and window level, preprocessing the CT value in the window width and window level, converting the preprocessed CT value into a single-channel matrix, and comprising the following steps:
reading the CT value of the original image of the head, screening the values which accord with a first condition in the CT values for reservation, and removing the residual values to obtain the coordinate information of the skull;
setting a window level of a required window width, and further extracting a first numerical value from a CT value of a head original image;
and normalizing the extracted first numerical value, masking the coordinate information of the skull in a matrix corresponding to the set window width window level, converting the coordinate information into 0, and constructing to obtain a single-channel matrix.
In this embodiment, the extracted numerical value is normalized to convert the value to 0 to 1, the previously stored skull coordinate information is masked in the brain window matrix, and the data of this portion is converted to 0. Finally, the 256 × 256 matrix of a single channel and the labeled data of the same size are stored as model training input.
In this embodiment, the stored CT value of the Dicom raw image of the head flat scan CT is read, and the window width and window level information is adjusted to be unified into a brain window (brain window) [40,80], which is the window width and window level mainly used by the doctor for diagnosing intracranial hemorrhage. Sometimes, a physician may use subdural window (80, 200) to diagnose subdural bleeding, better separating the bleeding of the skull and the near-to-skull, and the method will remove the skull, so selecting a more distinctive window will not cause the skull to connect with the bleeding highlight.
In this example, each input is a complete serial image of a diagnosed bleeding lesion, with a thickness varying from 1.5mm to 10 mm. The single sample without the bleeding lesion was regarded as a negative sample, and the other sample was regarded as a positive sample.
In this embodiment, the first condition, such as obtaining coordinate values of the skull portion after normalizing the image from the CT value, is to retain the portion with the CT value greater than 400.
The beneficial effects of the above technical scheme are: by adopting a sample enhancement preprocessing mode, the richness of the sample is improved by methods of random rotation, scaling, turning, shearing and the like on the image, the generalization capability of the model is enhanced, and the recognition degree and the delineation accuracy are indirectly improved.
Example 3:
based on embodiment 1, before being input into the preset delineation model in step 2, the method further includes:
constructing a basic intracranial hemorrhage segmentation model, and determining a segmentation network evaluation index set;
based on the evaluation index set, evaluating the basic intracranial hemorrhage segmentation model;
determining the basic grade of the basic intracranial hemorrhage segmentation model according to the evaluation result;
when the basic grade is a conventional grade, optimizing the basic intracranial hemorrhage segmentation model to obtain a preset delineation model;
and when the basic grade is a special grade, storing the basic intracranial hemorrhage segmentation model, wherein the stored model is the preset delineation model.
In this embodiment, the construction of the basic intracranial hemorrhage segmentation model, as shown in fig. 3, can be implemented as follows:
the convolutional neural network adopting the Unet structure consists of 2 parts of down sampling and up sampling;
the downsampling part has 5 convolution blocks in total, the features are extracted through 2d convolution, each block completes two times of convolution by using the convolution kernel size of 3x3, feature selection is carried out through maximum pooling, an activation function ReLU is used for each layer, and a generated feature map is used as the input of the next convolution layer. The fifth block is not pooled;
the up-sampling part performs transposition convolution by using a 2d layer with convolution kernel of 2x2, and returns the image matrix to the input size by connecting the information of the corresponding down-sampling layer;
and the last layer is a 2d convolution layer with a sigmoid activation function and 1, 1 convolution kernel, and finally a trained model is output, and the model also uses an Adam optimization algorithm.
In this embodiment, segmenting the network evaluation index set includes: the Dice coefficient and IoU coefficient, etc. are included as evaluation measurement indexes.
In this embodiment, for example, when the model of the head is conventionally recognized according to the basic level of the model determined by the evaluation index, the model is optimized, and if the model itself is trained to recognize the tiny features of the head, the model can be regarded as the preset delineation model, as shown in fig. 4 to 10.
The beneficial effects of the above technical scheme are: the basic intracranial hemorrhage segmentation model is constructed, so that the basic model is conveniently constructed, and the basic model is evaluated to determine the grade, further determine whether optimization is needed or not, and guarantee the accuracy of prediction and delineation.
Example 4:
based on the embodiment 3, the basic intracranial hemorrhage segmentation model is optimized to obtain a preset delineation model, which comprises the following steps:
extracting a set to be optimized of a basic intracranial hemorrhage segmentation model from a model optimization database;
determining a special part which is not finely identified by the basic intracranial hemorrhage segmentation model, determining the delineation difficulty of the special part, and further determining a corresponding fine delineation index according to the delineation difficulty;
performing architecture analysis on the basic intracranial hemorrhage segmentation model, and determining basic information of each layer and related information between adjacent layers in the basic intracranial hemorrhage segmentation model according to an analysis result;
determining a matching layer related to the set to be optimized and the fine delineation indexes based on the basic information and the related information;
determining indexes to be optimized in a set to be optimized and involved fine delineation indexes involved in each matching layer, and constructing an improved relation between the involved indexes in each matching layer and the matching layer;
judging whether related indexes in the matching layer are unique and can not be discarded according to the improvement relation;
determining the comprehensive optimization level corresponding to the remaining related indexes in the corresponding matching layer according to the judgment result, and analyzing the improvement degree of each matching layer and the improvement effect of the improvement degree on the basic intracranial hemorrhage segmentation model based on the improvement degree model;
determining a sequential execution scheme according to the execution attribute of the matching layer, and evaluating the overall improvement of the basic intracranial hemorrhage segmentation model according to the corresponding sequential execution scheme, the improvement degree and the improvement effect of the corresponding matching layer;
according to the evaluation result, calibrating the sequential execution scheme with execution pertinence, matching the corresponding sequential execution scheme to optimize the basic intracranial hemorrhage segmentation model according to the delineation requirement, and performing first storage;
performing effect screening on the execution sequence scheme without execution pertinence to obtain an optimal execution scheme, optimizing a basic intracranial hemorrhage segmentation model according to the optimal execution scheme, and performing second storage;
the drawing difficulty is related to the identification definition and the drawing definition of the original head image.
In this embodiment, specific sites such as ventricular hemorrhage, subdural hemorrhage, epidural hemorrhage, and subarachnoid space are used.
In this embodiment, the model optimization database is preset, and includes optimization indexes related to ventricular hemorrhage, subdural hemorrhage, epidural hemorrhage, subarachnoid space and other parts, so as to facilitate optimization of the basic model, for example, the basic model has an unsatisfactory recognition effect on ventricular hemorrhage and subdural hemorrhage, and at this time, the optimization indexes capable of recognizing ventricular hemorrhage and subdural hemorrhage need to be extracted, and then a set to be optimized is obtained.
And the optimization index is related to the part structure of the part to be optimally identified.
In this embodiment, the drawing difficulty is related to the position structure of the drawing part in addition to the identification definition and the drawing definition, so as to determine corresponding fine drawing indexes, such as drawing chromaticity, drawing thickness, identification drawing definition, and the like.
In this embodiment, the architecture analysis may be determined according to the model building manner in embodiment 3, so as to obtain the architecture of the model, and the basic information may include the functions executed by the layer, the operation parameters and the execution flow related to the layer, which are preset, and the relevant relationships, such as that adjacent layers supplement each other in the process of executing some functions, or are sequentially executed, and the like.
In this embodiment, for example, there are layers 1, 2, 3, and 4 in the basic model, there are indexes to be optimized 11, 12, 13, and 14, and there are fine delineation indexes 21 and 22, in this case, the determined matching layers are layer 1 and layer 3, in this case, the indexes related to layer 1 include: 11. 12, 21, and the indexes related to the layer 3 comprise 13, 14 and 22, so that the improved relation of the layer 1 and the layer 2 of the layers 11, 12 and 21 is determined, and the improved relation of the layer 2 and the layers 13, 14 and 22 is determined.
In this embodiment, the number of remaining related indicators is determined according to the unique non-discardable result, and the number of remaining related indicators is less than or equal to the number of all related indicators of the corresponding layer.
In this embodiment, the improvement degree model is also trained in advance, and is obtained by training different indexes, different index combinations in different layers, and the improvement effect, the improvement degree and the like of the corresponding combination indexes.
In this embodiment, the improvement degree refers to the improvement degree of the matching layer, for example, the layer 1, for example, the ventricular hemorrhage can be recognized originally, but the recognition is unclear, the accurate hemorrhage range can not be determined, in this case, after the improvement, the recognition can be clear, and the improvement degree is related to the process result that the recognition is unclear.
In this embodiment, according to the execution attribute, the determined layer 1 and layer 2 are the layer 1 and layer 2 or the layer 2 and layer 1 in the sequential execution scheme, and at this time, the improvement degree and the improvement effect can be determined according to different execution schemes, so as to perform improvement evaluation.
In this embodiment, the execution targeting refers to the targeted site, for example, the site is executed in the order of layer 1 and layer 2, and the ventricular hemorrhage is targeted, and in this case, the sequential execution scheme is regarded as having the execution targeting.
Otherwise, it is deemed as having no execution pertinence.
The beneficial effects of the above technical scheme are: the basic model is analyzed, the basic information and the relevance of the layer are determined, the matching layer is determined by determining the index to be optimized and the fine delineation index, the improvement effect of the matching layer is convenient to determine, the improvement relation between the matching layer and different indexes is determined, whether the index is unique and can not be discarded is convenient to determine, the execution flow and the identification flow are effectively saved, the identification efficiency is improved, the optimization grade of each layer is determined and the improved model is combined, the improvement effect can be effectively determined, finally, different screening and storage are performed according to the execution pertinence through the sequential execution scheme, the optimization feasibility of the basic model is ensured, and the delineation accuracy is ensured.
Example 5:
based on the embodiment 4, determining the comprehensive optimization level corresponding to the remaining related indexes in the corresponding matching layer according to the judgment result, including:
determining a first execution content before optimization and a second execution content after optimization of the matching layer and a first improvement content of the matching layer, which relates to indexes, based on the improvement relation;
determining the difference degree Y1 between the first execution content and the first improved content and the second execution content;
Figure BDA0003492244560000141
wherein a denotes a first execution content, B denotes a first improvement content, and C denotes a second execution content;
Figure BDA0003492244560000142
representing a similarity ratio of the first execution content and the first modified content to the second execution content;
when the difference degree Y1 is smaller than the preset degree, judging that all related indexes in the matching layer are unique and can not be discarded, reserving all related indexes of the matching layer, and determining a comprehensive optimization level X1 corresponding to the reserved indexes;
Figure BDA0003492244560000151
wherein n represents the number of all related indexes; deltaiIndicating the index weight of the ith index related to the index at the matching layer; siIndicating the index conversion parameter value of the ith index related to the index at the matching layer; gamma rayiThe adjustment parameter of the matching layer to the ith related index is represented, and the value range is [0.9,1.1 ]]；
Figure BDA0003492244560000152
A standard parameter value representing a corresponding i-th index-related index in the matching layer;
Figure BDA0003492244560000153
indicating the corresponding ith in the matching layer+1 standard parameter values relating to the index; si+1Indicating the (i + 1) th index conversion parameter value related to the index at the matching layer;
otherwise, judging whether the related indexes in the matching layer are unique and can not be discarded, searching the discardable indexes, reserving the discardable indexes, and meanwhile, establishing a calling index for the discardable indexes;
determining a discard value F of the discardable indicator;
Figure BDA0003492244560000154
where n1 represents the number of discardable pointers; djRepresents the jth discardable index r in the matching layerjThe number of associated indexes;
Figure BDA0003492244560000155
indicating the jth discardable index r in the matching layerjAssociated set of metrics
Figure BDA0003492244560000156
The correlation coefficient of (a);
Figure BDA0003492244560000157
represents the jth discardable index r in the matching layerjAssociated set of metrics
Figure BDA0003492244560000158
A corresponding discard coefficient;
setting a calling possibility label for the calling index according to the discarding value F, wherein the calling possibility label is related to the sketching part;
determining whether the calling possibility label is associated with the calling possibility label or not according to the drawn part bias;
if yes, determining a first index weight of a non-discardable index and a second index weight of a discardable index in a corresponding matching layer, and further obtaining a corresponding comprehensive optimization level X2;
if not, determining the comprehensive optimization level X3 corresponding to the residual related indexes in the corresponding matching layer.
In this embodiment, the calculation of the composite optimization level associated with the discarded index is similar to the calculation of all remaining related indices.
Figure BDA0003492244560000161
Figure BDA0003492244560000162
Wherein, deltai1A first index weight representing the i1 th discard index at the matching layer; si1Index conversion parameter values representing the i1 th discarded index at the matching layer; gamma rayi1The adjustment parameter of the matching layer to the i1 th discarding index is represented, and the value range is [0.9,1.1 ]]；δi2A second index weight representing the i2 th remaining index at the matching layer; si2Index conversion parameter values representing the i2 th remaining index at the matching layer; gamma rayi2The adjustment parameters of the matching layer to the i2 th residual indexes are represented, and the value range is [0.9,1.1 ]]；
In this embodiment, the first execution content and the first modified content are consistent with the second execution content in principle, but a certain error always exists in the operation process, so that the difference between the first execution content and the second modified content is calculated, the first execution content refers to the executable process before the matching layer is not optimized, the second execution content refers to the executable process after the matching layer is optimized, and the first modified content refers to the content optimizing the matching layer which is not optimized.
In this embodiment, the preset degree is preset, and is generally 0.1;
in this embodiment, the only non-discardable means that the index must remain at the matching layer.
In this example, when
Figure BDA0003492244560000163
The larger the correlation coefficient is, the more the influence on the improvement effect is, and further the corresponding discarding coefficient is smaller, and different discarding values, the set possibility label is different, and the label is related to not only the sketched part but also the discarding value.
In this embodiment, the location is biased, for example, toward the ventricle, and at this time, the probability label related to the ventricle may be called, so that the accuracy of the delineation is improved by a plurality of indexes.
The beneficial effects of the above technical scheme are: the method comprises the steps of judging whether the index is unique or not through calculating the difference, and further determining comprehensive optimization levels under different judgment results, wherein the comprehensive optimization levels are determined through three modes, one mode is to calculate the comprehensive optimization levels corresponding to all related indexes, and the other mode is to set a probability label through determining a discarded value.
Example 6:
based on embodiment 1, step 1, in the process of processing the head original image, further includes:
determining an initial format of a head original image;
when the initial format does not meet the identification condition, converting the initial format into a standard format and identifying;
when the recognition condition is satisfied, the recognition is continued.
Such as: the initial format may be DICOM or NII medical images with Hounsfield Unit number, which the system will automatically read for conversion. When the initial format system cannot automatically recognize the format, the initial format is converted into a recognizable format for continuous recognition.
The beneficial effects of the above technical scheme are: through the unification of the identification format, the image is convenient to effectively identify, and the accuracy of the sketching is indirectly guaranteed.
Example 7:
based on the embodiment 1, step 3 predicts the delineation data based on the preset delineation model, and outputs a predicted value, including:
identifying the processed original image and the preprocessing result based on a preset sketching model, automatically identifying a bleeding area in the processed original image and segmenting if a bleeding phenomenon exists;
generating a corresponding predicted value based on the recognition result and the segmentation result;
wherein, if the bleeding phenomenon exists, the bleeding area in the processed original image is automatically identified and segmented, and the method comprises the following steps:
roughly determining a first area with a bleeding phenomenon and a second area without the bleeding phenomenon, judging the positions of the first area and the second area at the head and the position relation between the first area and the second area, and further determining the current identification precision of the first area;
determining an identification circulation mode based on the preset identification precision of the preset delineation model and the current identification precision, and finely identifying the processed original image according to the identification circulation mode;
and segmenting the processed original image according to multiple fine recognition results.
The beneficial effects of the above technical scheme are: through discerning the segmentation to the image, obtain the predicted value, be convenient for obtain the picture of delineating, simultaneously, also can guarantee the precision of prediction picture of delineating, confirm the discernment complexity in this region through confirming position and position relation, more complicated correspond the discernment precision need be higher, earlier through rough determination, and then meticulous determination, can effectively cut apart the image, and discernment circulation mode, for example, obtain a plurality of sub-images to the image cutting, confirm discernment entry point from the sub-image and discern, realize effectively cutting apart.
Example 8:
based on the embodiment 1, after obtaining the lesion delineation image for outputting the predicted image based on the predicted value in step 4, the method further includes:
carrying out data conversion on an image value in a focus sketching image;
and according to the output format, carrying out format adjustment on the converted data to obtain a view result and outputting and displaying the view result.
In this embodiment, the image values generated by calculation may be subjected to data conversion, and an output format is selected, or an image number is directly selected in the system, and a result view returned by the AI algorithm is generated. The brain image view format may also be generated nii on demand.
The beneficial effects of the above technical scheme are: by adjusting the output format, the effectiveness of output display is convenient to guarantee, and diagnosis is convenient.
Example 9:
based on the embodiment 1, the method comprises the following steps: obtaining a lesion delineation image outputting a predicted image based on the predicted value, comprising:
carrying out numerical value unified classification on the predicted values to obtain focus blocks corresponding to different numerical values, and obtaining a block matrix of each focus block;
according to the intracranial construction standard value, obtaining edge contour matrixes of the corresponding parts of different focus blocks in a normal state;
performing edge registration on the edge contour matrix and the block matrix, judging whether the edge contour matrix is completely consistent with edge elements in the block matrix, and if so, reserving the block matrix;
otherwise, determining the element position and the dispersion degree of the unmatched elements existing in the edge contour matrix, planning a first offset value, simultaneously determining the element position and the dispersion degree of the unmatched elements existing in the block matrix, and planning a second offset value;
determining first line segments which do not match elements in the block matrix, tracking initial elements and final elements of each first line segment, and determining curve changes corresponding to the first line segments;
determining the positions of the primary element and the final element, and when the primary element and the final element are on an edge line segment corresponding to the edge contour matrix, intercepting a second line segment corresponding to the edge contour matrix, and determining offset subsegments of the first line segment and the second line segment;
performing a first adjustment to the lesion block based on the offset subsegment, the first offset value, and the second offset value;
when any element of the initial element and the final element is not on the edge line segment corresponding to the edge outline matrix, determining a second position of the element not on the edge line segment, the element orientation of the element not on the edge line segment and the edge standard element, and the curve change results of all the first line segments, establishing an extension direction, and extending the corresponding first line segment according to the extension direction to obtain an extension sub-segment;
performing a second adjustment to the lesion block based on the extended sub-segment, the first offset value, and the second offset value;
and outputting to obtain a focus drawing according to the first adjustment result and the second adjustment result.
In this embodiment, the first line segment is inclusive of the non-matching element.
In this embodiment, the predicted value is, for example, a value between 0 and 1, and each pixel has its assigned predicted value, and according to the predicted value, different focal blocks are obtained, and further a block matrix is obtained, and the block matrix is related to the value and the position.
In this embodiment, the edge profile matrix is standard, is pre-defined, and is registered, and is illuminated to determine whether to retain the block matrix.
In this embodiment, the first offset value is offset based on the standard of the edge contour matrix, and the second offset value is offset based on the block matrix, for example, the first offset value is 0.1, and the second offset value is 0.12, so that when the positions of the unmatched elements are adjusted, the focus mass adjustment accuracy is not high due to excessive offset or too large deviation of the offset from the standard, and the focus mass adjustment accuracy is also matched with the actual bleeding situation.
In this embodiment, as shown in fig. 11, a1 represents a first line segment, a11 represents a primary element, a22 represents a final element, the offset sub-segment is b1, and a2 represents a second line segment. a3 represents a mismatch element.
In this embodiment, the element directions are extension arrows in the arrow direction toward the standard outline direction, and the corresponding last extension sub-segment is c1, where the element directions of the element not on the edge line segment and the edge standard element are mainly used as the main basis of the extension sub-segment, and the remaining indexes are used as the auxiliary basis, so as to finally obtain the extension sub-segment.
In this embodiment, each lesion mass may relate to either or both of the offset sub-segment, the extended sub-segment.
The beneficial effects of the above technical scheme are: the predicted values are classified in a unified mode, then edge registration is carried out subsequently, the offset fields and the extension subsections are obtained, the focus blocks are adjusted, and the accuracy of the obtained focus blocks is guaranteed mainly for adjusting edge lines.
Example 10:
the invention provides an automatic intracranial hemorrhage delineation device based on a head CT flat scan image, which comprises the following components as shown in figure 12:
the preprocessing module is used for processing the original image of the head, reading the processed CT value, adjusting the required window width level, preprocessing the CT value in the window width level and converting the preprocessed CT value into a single-channel matrix;
the input module is used for inputting the processed original image and the preprocessing result into a preset delineation model;
the output module is used for predicting the sketching data based on a preset sketching model and outputting a predicted value;
and the acquisition module is used for acquiring a focus delineation image of the output prediction image based on the prediction value.
The beneficial effects of the above technical scheme are: through setting up the window width window level that corresponds, and through carrying out the preliminary treatment to the CT value, can guarantee the degree of recognition of model on small characteristic to combine to predetermine the model of drawing, carry out accurate segmentation drawing to intracranial hemorrhage, reduce the diagnosis required time, improve detection efficiency and accurate diagnosis.
Example 11:
based on any of the embodiments 1-9 above, the invention further provides a readable storage medium, on which instructions are stored, and the instructions, when executed by a computer, implement the functions of any of the method embodiments above.
The invention also provides a computer program product which, when executed by a computer, implements the functionality of any of the method embodiments described above.
In the above embodiments, the implementation may be wholly or partially realized by software, hardware, firmware, or any combination thereof. When implemented in software, may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer programs. The procedures or functions according to the embodiments of the present disclosure are wholly or partially generated when the computer program is loaded and executed on a computer. The computer may be a general purpose computer, a special purpose computer, a network of computers, or other programmable device. The computer program can be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer program can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wire (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device, such as a server, a data center, etc., that incorporates one or more of the available media. The usable medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a magnetic tape), an optical medium (e.g., a Digital Video Disk (DVD)), or a semiconductor medium (e.g., a Solid State Disk (SSD)), among others.
It will be apparent to those skilled in the art that various changes and modifications may be made in the present invention without departing from the spirit and scope of the invention. Thus, if such modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include such modifications and variations.

Claims (10)

1. An automatic intracranial hemorrhage delineation method based on a head CT flat scanning image is characterized by comprising the following steps:
step 1: processing the original image of the head, reading the processed CT value, adjusting the required window width level, preprocessing the CT value in the window width level, and converting the preprocessed CT value into a single-channel matrix;
step 2: inputting the processed original image and the preprocessing result into a preset delineation model;
and step 3: predicting the sketching data based on a preset sketching model and outputting a predicted value;
and 4, step 4: and obtaining a focus delineation image for outputting a prediction image based on the prediction value.
2. The method for automatically delineating intracranial hemorrhage according to claim 1, wherein step 1: processing the original image of the head, reading the processed CT value, adjusting the required window width and window level, preprocessing the CT value in the window width and window level, converting the preprocessed CT value into a single-channel matrix, and comprising the following steps:
reading the CT value of the original image of the head, screening values which accord with a first condition in the CT values for reservation, and removing the residual values to obtain the coordinate information of the skull;
setting a required window width window level, and further extracting a first numerical value from the CT value of the head original image;
and normalizing the extracted first numerical value, masking the coordinate information of the skull in a matrix corresponding to the set window width and window level, converting the coordinate information into 0, and constructing to obtain a single-channel matrix.
3. The method for automatically delineating intracranial hemorrhage according to claim 1, wherein the step 2, before inputting into the preset delineation model, further comprises:
constructing a basic intracranial hemorrhage segmentation model, and determining a segmentation network evaluation index set;
evaluating the basic intracranial hemorrhage segmentation model based on the evaluation index set;
determining a basic grade of the basic intracranial hemorrhage segmentation model according to the evaluation result;
when the basic grade is a conventional grade, optimizing the basic intracranial hemorrhage segmentation model to obtain a preset delineation model;
and when the basic grade is a special grade, storing the basic intracranial hemorrhage segmentation model, wherein the stored model is the preset delineation model.
4. The method according to claim 3, wherein the optimizing the basic intracranial hemorrhage segmentation model to obtain a preset delineation model comprises:
extracting a set to be optimized of the basic intracranial hemorrhage segmentation model from a model optimization database;
determining a special part which is not finely identified by the basic intracranial hemorrhage segmentation model, determining the delineation difficulty of the special part, and further determining a corresponding fine delineation index according to the delineation difficulty;
performing architecture analysis on the basic intracranial hemorrhage segmentation model, and determining basic information of each layer and related information between adjacent layers in the basic intracranial hemorrhage segmentation model according to an analysis result;
determining a matching layer related to the set to be optimized and the fine delineation index based on basic information and related information;
determining indexes to be optimized in a set to be optimized and involved fine delineation indexes involved in each matching layer, and constructing an improved relation between the involved indexes in each matching layer and the matching layer;
judging whether related indexes in the matching layer are unique and can not be discarded according to the improvement relation;
determining the comprehensive optimization level corresponding to the remaining related indexes in the corresponding matching layer according to the judgment result, and analyzing the improvement degree of each matching layer and the improvement effect of the improvement degree on the basic intracranial hemorrhage segmentation model based on an improvement degree model;
determining a sequential execution scheme according to the execution attribute of the matching layer, and evaluating the overall improvement of the basic intracranial hemorrhage segmentation model according to the sequential execution scheme and the improvement degree and the improvement effect of the corresponding matching layer;
according to the evaluation result, calibrating the sequential execution scheme with execution pertinence, matching the corresponding sequential execution scheme to optimize the basic intracranial hemorrhage segmentation model according to the delineation requirement, and performing first storage;
performing effect screening on the scheme without execution pertinence execution sequence to obtain an optimal execution scheme, optimizing the basic intracranial hemorrhage segmentation model according to the optimal execution scheme, and performing second storage;
the drawing difficulty is related to the identification definition and the drawing definition of the original head image.
5. The method for automatically delineating intracranial hemorrhage according to claim 4, wherein determining the comprehensive optimization level corresponding to the remaining related indexes in the corresponding matching layer according to the determination result comprises:
determining a first execution content before optimization and a second execution content after optimization of the matching layer and a first improvement content of indexes related to the matching layer on the basis of an improvement relation;
determining the difference degree Y1 between the first execution content and the first improved content and the second execution content;
Figure FDA0003492244550000031
wherein a denotes a first execution content, B denotes a first improvement content, and C denotes a second execution content;
Figure FDA0003492244550000032
indicating first execution content and first modificationA similarity ratio of the incoming content to the second execution content;
when the difference degree Y1 is smaller than a preset degree, judging that all related indexes in the matching layer are unique and can not be discarded, reserving all related indexes of the matching layer, and determining a comprehensive optimization level X1 corresponding to the reserved indexes;
Figure FDA0003492244550000033
wherein n represents the number of all related indexes; deltaiIndicating the index weight of the ith index related to the index at the matching layer; siIndicating the index conversion parameter value of the ith index related to the index at the matching layer; gamma rayiThe adjustment parameter of the matching layer to the ith related index is represented, and the value range is [0.9,1.1 ]]；
Figure FDA0003492244550000034
A standard parameter value representing a corresponding i-th index-related index in the matching layer;
Figure FDA0003492244550000035
a standard parameter value representing the corresponding i +1 th index-related item in the matching layer; si+1Indicating the (i + 1) th index conversion parameter value related to the index at the matching layer;
otherwise, judging whether the related indexes in the matching layer are unique and can not be discarded, searching for a discardable index, leaving the discardable index, and meanwhile, establishing a calling index for the discardable index;
determining a discard value F of the discardable indicator;
Figure FDA0003492244550000036
where n1 represents the number of discardable pointers; djIndicating the jth discardable index r in the matching layerjThe number of associated indexes;
Figure FDA0003492244550000037
indicating the jth discardable index r in the matching layerjAssociated set of metrics
Figure FDA0003492244550000038
The correlation coefficient of (a);
Figure FDA0003492244550000039
represents the jth discardable index r in the matching layerjAssociated set of metrics
Figure FDA0003492244550000041
A corresponding discard coefficient;
setting a calling possibility label for the calling index according to a discarded value F, wherein the calling possibility label is related to a sketching part;
determining whether the calling possibility label is associated with the calling possibility label or not according to the drawn part bias;
if the index exists, determining a first index weight of the non-discardable index and a second index weight of the discardable index in the corresponding matching layer, and further obtaining a corresponding comprehensive optimization grade;
and if not, determining the comprehensive optimization level corresponding to the residual related indexes in the corresponding matching layer.
6. The method for automatically delineating intracranial hemorrhage according to claim 1, wherein the step 1 of processing the original image of the head further comprises:
determining an initial format of the head original image;
when the initial format does not meet the identification condition, converting the initial format into a standard format and identifying;
when the recognition condition is satisfied, the recognition is continued.
7. The method for automatically delineating intracranial hemorrhage according to claim 1, wherein the step 3 of predicting delineation data based on a preset delineation model and outputting a predicted value comprises:
identifying the processed original image and the preprocessing result based on a preset sketching model, automatically identifying a bleeding area in the processed original image and segmenting if a bleeding phenomenon exists;
generating a corresponding predicted value based on the recognition result and the segmentation result;
wherein, if the bleeding phenomenon exists, the bleeding area in the processed original image is automatically identified and segmented, and the method comprises the following steps:
roughly determining a first area with a bleeding phenomenon and a second area without the bleeding phenomenon, judging the positions of the first area and the second area at the head and the position relation between the first area and the second area, and further determining the current identification precision of the first area;
determining an identification circulation mode based on the preset identification precision of the preset delineation model and the current identification precision, and finely identifying the processed original image according to the identification circulation mode;
and segmenting the processed original image according to multiple fine identification results.
8. The intracranial hemorrhage automatic delineation method according to claim 1, wherein the step 4, after obtaining the focus delineation image outputting a prediction image based on the prediction value, further comprises:
carrying out data conversion on an image value in a focus sketching image;
and according to the output format, carrying out format adjustment on the converted data to obtain a view result and outputting and displaying the view result.
9. The method for automatically delineating intracranial hemorrhage according to claim 1, wherein step 4: obtaining a lesion delineation image outputting a predicted image based on the predicted value, comprising:
carrying out numerical value unified classification on the predicted values to obtain focus blocks corresponding to different numerical values, and obtaining a block matrix of each focus block;
according to the intracranial construction standard value, obtaining edge contour matrixes of the corresponding parts of different focus blocks in a normal state;
performing edge registration on the edge contour matrix and the block matrix, judging whether the edge contour matrix is completely consistent with edge elements in the block matrix, and if so, reserving the block matrix;
otherwise, determining the element position and the dispersion degree of the unmatched elements existing in the edge contour matrix, planning a first offset value, simultaneously determining the element position and the dispersion degree of the unmatched elements existing in the block matrix, and planning a second offset value;
determining first line segments which do not match elements in the block matrix, tracking initial elements and final elements of each first line segment, and determining curve changes corresponding to the first line segments;
determining the positions of the primary element and the final element, and when the primary element and the final element are on an edge line segment corresponding to the edge contour matrix, intercepting a second line segment corresponding to the edge contour matrix, and determining offset subsegments of the first line segment and the second line segment;
performing a first adjustment to the lesion block based on the offset subsegment, the first offset value, and the second offset value;
when any element of the initial element and the final element is not on the edge line segment corresponding to the edge outline matrix, determining a second position of the element not on the edge line segment, the element orientation of the element not on the edge line segment and the edge standard element, and curve change results of all the first line segments, establishing an extension direction, and extending the corresponding first line segments according to the extension direction to obtain extension sub-segments;
performing a second adjustment to the lesion block based on the extended sub-segment, the first offset value, and the second offset value;
and outputting to obtain a focus drawing according to the first adjustment result and the second adjustment result.
10. The utility model provides an automatic device that sketches of intracranial hemorrhage based on head CT flat scanning image which characterized in that includes:
the preprocessing module is used for processing the head original image based on the window width and window level to obtain a single-channel matrix, and meanwhile, reading an identification standard value obtained by normalizing and converting the CT value in the single-channel matrix and preprocessing the identification standard value;
the input module is used for inputting the processed original image and the preprocessing result into a preset delineation model;
the output module is used for predicting the sketching data based on a preset sketching model and outputting a predicted value;
and the acquisition module is used for acquiring a focus delineation image of the output prediction image based on the prediction value.
CN202210100541.6A 2022-01-27 2022-01-27 Automatic intracranial hemorrhage delineation method and device based on head CT flat scanning image Active CN114494496B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210100541.6A CN114494496B (en) 2022-01-27 2022-01-27 Automatic intracranial hemorrhage delineation method and device based on head CT flat scanning image

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210100541.6A CN114494496B (en) 2022-01-27 2022-01-27 Automatic intracranial hemorrhage delineation method and device based on head CT flat scanning image

Publications (2)

Publication Number Publication Date
CN114494496A true CN114494496A (en) 2022-05-13
CN114494496B CN114494496B (en) 2022-09-20

Family

ID=81477516

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210100541.6A Active CN114494496B (en) 2022-01-27 2022-01-27 Automatic intracranial hemorrhage delineation method and device based on head CT flat scanning image

Country Status (1)

Country Link
CN (1) CN114494496B (en)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105719295A (en) * 2016-01-21 2016-06-29 浙江大学 Intracranial hemorrhage area segmentation method based on three-dimensional super voxel and system thereof
CN112582048A (en) * 2021-03-01 2021-03-30 四川大学 Radiotherapy target area automatic delineation system and method based on deep learning
CN112614144A (en) * 2020-12-30 2021-04-06 深圳市联影高端医疗装备创新研究院 Image segmentation method, device, equipment and storage medium
CN112634308A (en) * 2020-12-25 2021-04-09 南方医科大学 Nasopharyngeal carcinoma target area and endangered organ delineation method based on different receptive fields
CN112790782A (en) * 2021-02-02 2021-05-14 广州柏视医疗科技有限公司 Automatic delineation method of pelvic tumor CTV based on deep learning
CN113160245A (en) * 2021-04-01 2021-07-23 深圳市铱硙医疗科技有限公司 CT brain parenchyma segmentation system, method and device based on block region growing method

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105719295A (en) * 2016-01-21 2016-06-29 浙江大学 Intracranial hemorrhage area segmentation method based on three-dimensional super voxel and system thereof
CN112634308A (en) * 2020-12-25 2021-04-09 南方医科大学 Nasopharyngeal carcinoma target area and endangered organ delineation method based on different receptive fields
CN112614144A (en) * 2020-12-30 2021-04-06 深圳市联影高端医疗装备创新研究院 Image segmentation method, device, equipment and storage medium
CN112790782A (en) * 2021-02-02 2021-05-14 广州柏视医疗科技有限公司 Automatic delineation method of pelvic tumor CTV based on deep learning
CN112582048A (en) * 2021-03-01 2021-03-30 四川大学 Radiotherapy target area automatic delineation system and method based on deep learning
CN113160245A (en) * 2021-04-01 2021-07-23 深圳市铱硙医疗科技有限公司 CT brain parenchyma segmentation system, method and device based on block region growing method

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
HAI YE .ETC: "Precise diagnosis of intracranial hemorrhage and subtypes using a three-dimensional joint convolutional and recurrent neural network", 《EUROPEAN RADIOLOGY》 *
刘晓东等: "基于CT图像的肿瘤放射治疗感兴趣区勾画方法", 《医疗卫生装备》 *
徐利明等: "FDG-PET/CT图像SUV阈值在放疗计划设计时的应用", 《中国医学物理学杂志》 *
江东等: "基于先验知识的CT影像自动勾画研究及应用", 《电脑知识与技术》 *
邓金城等: "深度卷积神经网络在放射治疗计划图像分割中的应用", 《中国医学物理学杂志》 *

Also Published As

Publication number Publication date
CN114494496B (en) 2022-09-20

Similar Documents

Publication Publication Date Title
CN110060774B (en) Thyroid nodule identification method based on generative confrontation network
CN111798462B (en) Automatic delineation method of nasopharyngeal carcinoma radiotherapy target area based on CT image
WO2018120942A1 (en) System and method for automatically detecting lesions in medical image by means of multi-model fusion
CN103249358B (en) Medical image-processing apparatus
CN108133476B (en) Method and system for automatically detecting pulmonary nodules
CN111028206A (en) Prostate cancer automatic detection and classification system based on deep learning
CN109886933A (en) A kind of medical image recognition method, apparatus and storage medium
CN110974306B (en) System for discernment and location pancreas neuroendocrine tumour under ultrasonic endoscope
US20220383661A1 (en) Method and device for retinal image recognition, electronic equipment, and storage medium
CN113077479A (en) Automatic segmentation method, system, terminal and medium for acute ischemic stroke focus
WO2022110525A1 (en) Comprehensive detection apparatus and method for cancerous region
US9483705B2 (en) Image processing device, image processing method, and image processing program
CN114494215A (en) Transformer-based thyroid nodule detection method
CN111524188A (en) Lumbar positioning point acquisition method, equipment and medium
CN117036288A (en) Tumor subtype diagnosis method for full-slice pathological image
CN114494496B (en) Automatic intracranial hemorrhage delineation method and device based on head CT flat scanning image
CN116228731A (en) Multi-contrast learning coronary artery high-risk plaque detection method, system and terminal
CN116468923A (en) Image strengthening method and device based on weighted resampling clustering instability
CN116416221A (en) Ultrasonic image analysis method
CN113409273B (en) Image analysis method, device, equipment and medium
CN113379770B (en) Construction method of nasopharyngeal carcinoma MR image segmentation network, image segmentation method and device
CN115409812A (en) CT image automatic classification method based on fusion time attention mechanism
CN111210436B (en) Lens segmentation method, device and storage medium
CN114782448B (en) Brain glioma auxiliary grading system based on image processing
CN116682109B (en) Pathological microscopic image analysis method, device, equipment and storage medium

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