CN109165668A

CN109165668A - A kind of processing method of brain anomaly classification

Info

Publication number: CN109165668A
Application number: CN201810739890.6A
Authority: CN
Inventors: 吴振洲
Original assignee: Beijing Ander Medical Intelligence Technology Co Ltd
Current assignee: Beijing Ander Medical Intelligence Technology Co Ltd
Priority date: 2018-07-06
Filing date: 2018-07-06
Publication date: 2019-01-08

Abstract

The present invention provides a kind of processing methods of brain anomaly classification, include the following steps: (1) in the multiple images comprising brain exception, mark the position of the brain exception respectively using three-dimensional array, and record the quantity of the brain exception；(2) according to the position, and according to the prior distribution of predetermined brain localization and particular types brain exception probability of occurrence, the probability of occurrence of the extremely corresponding one or more types of brain described in each image is determined respectively；And (3) calculate the average value of identified multiple probabilities of occurrence, the probability of occurrence as the extremely corresponding one or more types of brain described in described multiple images.The present processes can be only in accordance with the type of location estimating brain exception.

Description

A kind of processing method of brain anomaly classification

Technical field

The present invention relates to the Magnetic Resonance Image Segmentation technical field of brain exception more particularly to a kind of brain anomaly classifications Processing method.

Background technique

The Accurate Segmentation of multi-modal magnetic resonance image is for diagnosis, surgery planning, postoperative Analysis and chemotherapy/radiotherapy planning It is most important.

There are many researchers to propose the segmentation side of the multi-modal magnetic resonance image of various GBM both at home and abroad at present Method is broadly divided into the partitioning algorithm based on pixel or voxel and the algorithm etc. based on figure segmentation.Based on pixel or voxel The basic thought of partitioning algorithm is that luminance information, texture information etc. divide the pixel on multi-modality images according to each pixel Class is into corresponding classification.

The algorithm of classification includes unsupervised cluster and the study for having supervision.For example, being based on fuzzy C-mean algorithm (Fuzzy C- Means, referred to as FCM) fuzzy means clustering algorithm, using the gray scale of multi-modal magnetic resonance image as feature vector, benefit All tissue points are clustered with FCM to obtain initial classification, according to symmetry, the priori knowledges such as intensity profile are to initial point Class optimizes, and obtains final segmentation result.When due to FCM cluster, space neighborhood information is not accounted for, and GBM is organized Intensity profile can generate overlapping, therefore be easy to produce accidentally divide.Image is described based on the vertex of figure partitioning algorithm figure Pixel, the similitude of 2 pixels is described with the side of figure, a network is consequently formed, pass through solve energy minimization problem Figure is divided into sub-network figure, reaches the similitude inside difference and same subnet the network figure between different sub-network network figure most Greatly.This kind of algorithm usually requires to solve the problems, such as a solution generalized eigenvector, and when image is bigger, this kind of algorithm can be met with The big problem of computation complexity.

Summary of the invention

The present invention provides a kind of processing method of brain anomaly classification, method of the invention solves deposits in the prior art The problem of, provide a kind of more accurate classification method.

The purpose of the present invention is realized by following scheme:

A kind of processing method of brain anomaly classification, includes the following steps:

(1) in the multiple images comprising brain exception, the position of the brain exception is marked respectively using three-dimensional array, And record the quantity of the brain exception；

(2) according to the position, and according to predetermined brain localization and particular types brain exception probability of occurrence Prior distribution, determine the probability of occurrence of the extremely corresponding one or more type of brain described in each image respectively；And

(3) average value of multiple probabilities of occurrence determined by calculating, it is extremely right as brain described in described multiple images The probability of occurrence for the one or more types answered.

Preferably, the prior distribution is a four-dimensional array, and wherein three-dimensional array represents particular types brain and goes out extremely Existing position, fourth dimension array represent the abnormal probability of occurrence in the position of variety classes brain.

Preferably, further includes: indicate that brain is without exception with the presence or absence of exception, 0 with the 0 and 1 label brain, 1 indicates big Brain exists abnormal.

Method of the invention only passes through the three-dimensional array of the prior distribution of abnormal position and Brain slices in brain can be pre- The type for surveying brain exception, i.e., only go out the type of brain exception, without reference brain figure by the location estimating of brain exception Picture reduces the medical pressure of doctor, improves the accuracy and objectivity that brain judges extremely.

Detailed description of the invention

Fig. 1 is the basic flow chart of image partition method of the invention；

Fig. 2 is the T2 image of certain patient's brain epidermal cyst；

Fig. 3 is the T2 Image Segmentation figure of the patient.

Specific embodiment

With reference to the accompanying drawing, detailed explanation is carried out to the present invention.

As shown in Figure 1, method of the invention is a kind of processing method of brain anomaly classification, include the following steps: that (1) is right The possible position of known exception is labeled in brain, generates a kind of prior distribution of possible distributing position of big abnormalities of brain, the elder generation Testing distribution is one group of four-dimension array；(2) the segmentation sectioning image for obtaining big abnormalities of brain, handles the sectioning image and extraction figure Picture obtains a three-dimensional array, marks there are the position of brain exception, the quantity of brain exception is denoted as X；(3) step (2) are marked The abnormal position of note is compareed with the prior distribution in step (1), obtains the vector of X 1*N, and N is certain brain of the position Extremely the probability occurred；(4) vector of X 1*N is averaged, obtains the vector of 1*N ', which is that certain brain is different Normal probability.

Prior distribution in the step (1) is one group of four-dimension array, which refers to different location in brain There is the probability of different brain exceptions, wherein three-dimensional representation is coordinate in brain, and what fourth dimension represented is variety classes brain The probability of portion's exception.Three-dimensional array refers to the coordinate position in brain in the step (2), is indicated with 0 and 1, and 0 indicates brain Without exception, 1 indicates that brain has exception.

Fig. 2 is the T2 image of certain patient's brain epidermal cyst；Fig. 3 is the T2 Image Segmentation figure of the patient, from the image point It cuts figure and obtains one group of three-dimensional array, and then the three-dimensional array is carried out after compareing averagely with prior distribution, obtain the brain of the patient Portion is epidermal cyst extremely, and the type of brain exception, and the deduction and Fig. 2 are gone out merely by the location estimating of brain exception In T2 image be consistent.

Prior distribution obtains by the following method:

1, from infection from hospital first data accordingly；

2, for each patient, it is understood that the type of the brain exception of patient and position, the position are expressed as 24 × 320 × 320 binary segmentation masks；

3, abnormal for each brain, all segmentation masks that may be obtained are collected from database；

4, it is vector superposed that the progress of mask 24 × 320 × 320 will be divided, obtain 24 × 320 × 320 vectors, and by available Segmentation mask carried out scalar division；

5, above-mentioned steps are operated extremely for N number of brain to get N number of 24 × 320 × 320 vector is arrived；It is each to Amount is a probability distribution, and each pixel value has scalar value, and brain of the scalar value instruction based on the location information is abnormal The probability of generation；

6, N number of 24 × 320 × 320 vector is obtained into 24 × 320 × 320 × N vector, as first by axis connection Test distribution.

The use of prior distribution is as follows:

1, the segmentation mask of a patient, the binary vector which is one 24 × 320 × 320 are obtained；

2, with the segmentation mask of the patient, i.e. 24 × 320 × 320 vectors, with 24 × 320 × 320 × N in prior distribution to Amount is multiplied, and obtains 24 × 320 × 320 × N brain spatial abnormal feature vector of the patient；

3, summation is with averagely on first three axis of brain spatial abnormal feature vector, and to obtain 1 × N vector, which is indicated The location information for being based solely on brain exception obtains the probability of brain exception.Such as N=10, the vector are expressed as v_1,

V_1=(1,1,2,2,1,1,1,1,1,1).

V_1 vector shows that according only to location information, third and fourth kind of brain is that the appearance of other brains exception is general extremely Twice of rate.

Provide the convolutional neural networks of any brain anomaly classification, it is assumed that the convolutional neural networks generate 1 × 10 vector v _ 2 =(0.5,0.5,0.1,0.1,0.5,0.5,0.5,0.5,0.5,0.5)；

V_1 obtains (0.5,0.5,0.2,0.2,0.5,0.5,0.5,0.5,0.5,0.5) multiplied by v_2, it has been found that convolution The output data of neural network is determined by the position distribution vector of brain exception, i.e., relies on the position of brain exception merely To infer the type of brain exception.

Brain has regiospecificity extremely, and certain exceptions more readily occur in some regions of brain.Doctor is only capable of Location information, which is used alone, can list the possible brain exception of patient.By the way that v_1 is multiplied with v_2, we can be by position Confidence breath, which is added in any convolutional neural networks, to be judged.

The foregoing is only a preferred embodiment of the present invention, but scope of protection of the present invention is not limited thereto, Within the technical scope of the present disclosure, any changes or substitutions that can be easily thought of by anyone skilled in the art, It should all be included within the scope of the present invention.Therefore, protection scope of the present invention should be with the protection of claims Subject to range.

Claims

1. a kind of processing method of brain anomaly classification characterized by comprising

(2) according to the position, and according to the elder generation of predetermined brain localization and particular types brain exception probability of occurrence Distribution is tested, determines the probability of occurrence of the extremely corresponding one or more types of brain described in each image respectively；And

(3) average value of multiple probabilities of occurrence determined by calculating, it is extremely corresponding as brain described in described multiple images The probability of occurrence of one or more types.

2. the method according to claim 1, wherein the prior distribution is a four-dimensional array, wherein three-dimensional Array represents the position that particular types brain occurs extremely, and fourth dimension array represents abnormal the going out in the position of variety classes brain Existing probability.

3. the method according to claim 1, wherein further include: with the 0 and 1 label brain with the presence or absence of different Often.