Fatty dividing method in a kind of abdomen based on deep learning
Technical field
The present invention relates to medical domains, more particularly, to dividing method fatty in a kind of abdomen based on deep learning.
Background technique
Fat in abdomen, which is divided into different parts in the field of medicine, very big value.It is intuitive compared to more
For body-mass index (BMI), the volume of visceral adipose tissue (VAT) and subcutaneus adipose tissue (SAT) can be more accurate
Effectively prediction disease relevant to analysis obesity.
Fatty method is mainly segmentation by hand in existing most of segmentation abdomen, i.e., by doctor or has professional experiences
Personnel carry out manual processing to original image, identify the range of fat and interior fat in abdomen.Currently, having already appeared on the market
The software of semi-automatic segmentation, but what its groundwork was still manually performed.Although this kind of software can be to a certain extent
The efficiency of segmentation is promoted, but doctor completed the segmentation of single picture there is still a need for 6-8 minutes time.In addition, these softwares
Price is usually expensive, and such as the fatty segmentation software of existing a SliceOmatic needs 4000 dollars.As it can be seen that these softwares
Practicability there are also to be hoisted.
Certainly, Part Full Automatic partitioning algorithm is suggested now, for example figure cuts method (Graph cut), driving wheel
Wide modelling (Active contour), threshold method (Thresholding) etc., these methods also have universal defect.First
It is that these algorithms picture new for each requires to re-start training, takes a substantial amount of time;Followed by these algorithms are wanted
SAT and VAT is split respectively, needs to handle picture twice, which increase the probability of extra false occur.
Summary of the invention
The purpose of the present invention is provide fatty dividing method in a kind of abdomen based on deep learning regarding to the issue above.
The purpose of the present invention can be achieved through the following technical solutions:
Fatty dividing method in a kind of abdomen based on deep learning, for dividing visceral adipose tissue and subcutaneous fat simultaneously
Tissue, the method includes the following steps:
1) stomach fat pixel is identified;
2) deep learning is carried out to the stomach fat pixel identified in step 1), obtains the sheet of stomach fat pixel
Matter feature;
3) substantive characteristics of stomach fat pixel obtained in step 2) is input in sorting algorithm, is tentatively divided
Cut result;
4) primary segmentation result obtained in step 3) is gone in polar coordinates, the stomach fat that further amendment is accidentally divided
Pixel obtains stomach fat segmentation figure;
5) all types of fat ratios in stomach fat segmentation figure obtained in step 4) are calculated according to gradient height,
And calculated result is intuitively shown by Volume Rendering Techniques.
The identification stomach fat pixel is realized by Active contour algorithm.
The step 1) is specially:
11) edge of abdomen is indicated by full curve;
12) energy functional is defined, the independent variable of the energy functional includes the full curve in step 11);
13) solution procedure 12) in energy functional minimum value, independent variable corresponding with energy functional minimum value be identify
Stomach fat pixel out.
The step 2) is specially:
21) it is inputted using the multiple dimensioned piece of feature as the stomach fat pixel identified in step 1), is successively constructed
Neural network;
22) all layers of neural network is trained in sequence, carries out tuning using wake-sleep algorithm, obtains
The substantive characteristics of stomach fat pixel.
The step 22) is specially:
221) weight modification between each layer neural network is two-way weight by the neural network for retaining top, described
Two-way weight includes recognizing weight upwards and generating weight downwards;
222) according to extraneous feature and cognition weight upwards, using gradient decline modify between each layer neural network to
Lower generation weight;
223) weight is generated according to the neural network of top and downwards, generates the state of bottom neural network, and is repaired
Change the upward cognition weight between each layer neural network;
224) using each layer neural network modified in step 223) as the substantive characteristics of stomach fat pixel.
The step 4) is specially:
41) primary segmentation result obtained in step 3) is gone in polar coordinates;
42) subcutaneus adipose tissue is used as image base under polar coordinates;
43) subcutaneus adipose tissue and visceral adipose tissue are compared, the stomach fat pixel that amendment is accidentally divided obtains abdomen
Fatty segmentation figure.
Compared with prior art, the invention has the advantages that:
(1) once training process is completed, it can be applied to new picture at once, without training again, greatly promote speed.
(2) SAT and VAT are distinguished simultaneously, without being split again, improves the efficiency of fat segmentation in abdomen.
(3) since the segmentation for SAT and VAT carries out simultaneously, the segmentation result of two sides can be carried out pair
It is modified than after, improves the accuracy of segmentation, reduce error generation rate.
(4) full-automatic dividing is realized, only just needs manually to carry out the simple later period by software under necessary condition
Amendment, high degree of automation.
(5) since the present invention is using improved deep neural network algorithm, the learning effect of the algorithm is good, improves
The accuracy of fat segmentation in abdomen.
Detailed description of the invention
Fig. 1 is flow chart of the method for the present invention;
Specific embodiment
The present invention is described in detail with specific embodiment below in conjunction with the accompanying drawings.The present embodiment is with technical solution of the present invention
Premised on implemented, the detailed implementation method and specific operation process are given, but protection scope of the present invention is not limited to
Following embodiments.
As shown in Figure 1, being specifically included following the present invention provides dividing method fatty in a kind of abdomen based on deep learning
Step:
1) identification of stomach fat pixel.Due to the special shape feature of abdomen, movable contour model (Active
Contour model) many algorithms are developed on its basis, basic thought is that target side is expressed using full curve
Edge, as soon as and define an energy functional its independent variable made to include boundary curve, cutting procedure, which is changed into, solves energy functional
The process of minimum value can generally realize that energy reaches curve position when minimum by solving the corresponding Eulerian equation of function
Set be exactly target profile where.Active contour model be one it is top-down positioning characteristics of image mechanism, user or its
He automatically processes process by being placed around an initial profile line in interesting target in advance, in internal energy (internal force) and outside
External energy is deformed under the action of portion's energy (external force) attracts active contour to move towards object edge, and internal energy holding activity
The slickness and topological of profile, when energy reaches minimum, active contour converges to object edge to be detected.Therefore,
Active contour algorithm can identify abdomen area well.
2) to the pixel in each abdomen, we use the multiple dimensioned piece of feature as the pixel for input, pass through one
The new deep neural network algorithm of kind learns the abstract substantive characteristics of layering automatically.These features are input to sorting algorithm
In, so that it may obtain a preliminary segmentation result.
The algorithm of deep neural network is divided into supervised learning and unsupervised learning, and thought is as follows:
In supervised learning, the problem of pervious multilayer neural network is to be easily trapped into Local Extremum.If training sample
This sufficiently covers following sample enough, then the multilayer weight acquired can be very good for predicting new test sample.But
That many tasks are difficult to obtain enough marker samples, in this case, simple model, for example, linear regression or certainly
Plan tree tends to obtain result more better than multilayer neural network (better generalization, worse training error).
In unsupervised learning, in the past without effective method construct multitiered network.The top layer of multilayer neural network is bottom
The advanced expression of feature, for example bottom is pixel, upper one layer of node may indicate horizontal line, triangle;And top layer may have one
A ode table is leted others have a look at face.One successful algorithm should be able to allow the top-level feature of generation maximumlly to represent the sample of bottom.Such as
For fruit to the training simultaneously of all layers, time complexity can be too high;If one layer of training every time, deviation will be transmitted successively.This meets
Face the problem opposite in supervised learning above, it can serious poor fitting.
The specific method is as follows:
Successively building monolayer neuronal is first first, is one single layer network of training every time in this way.
After having trained for all layers, hinton carries out tuning using wake-sleep algorithm.Other layers of top will be removed
Between weight become two-way, such top is still a monolayer neural networks, and other layers have then become graph model.To
On weight be used for " cognition ", downward weight is used for " generation ".Then all weights are adjusted using Wake-Sleep algorithm.
It allows cognition and generation to reach an agreement, that is, guarantees that the top generated indicates correctly restore the knot of bottom as far as possible
Point.For example an ode table of top layer is leted others have a look at face, then the image of all faces should activate this node, and this result
The image generated downwards should be able to show as a general facial image.Wake-Sleep algorithm is divided into awake (wake) and sleeps
(sleep) two parts.
Wake stage, cognitive process generate each layer of pumping by extraneous feature and upward weight (cognition weight)
As expression (node state), and use the downlink weight (generating weight) of gradient decline modification interlayer.Namely " if reality
Imagine different with me, change the thing that my weight imagines me be exactly as ".
2sleep stage, generating process indicate (concept to learn when waking up) and downward weight by top layer, generate bottom
State, while modifying the upward weight of interlayer.Namely " if the scene in dream is not the corresponding concepts in my brain, change me
Cognition weight to make this scene in my view be exactly this concept.
Although 3) deep learning can learn to more essential feature, due to explicitly in view of SAT and
The spatial distribution characteristic of VAT will cause the erroneous judgement of certain pixels.More accurate segmentation result in order to obtain, we will be original
Segmentation result proceed in polar coordinates.SAT becomes the bottom of image under polar coordinates, can be further according to this feature
The pixel that amendment is accidentally divided.
4) obtained according to plurality of pictures as a result, being counted in the way of gradient height to all types of fat ratios
It calculates, and this group of data is carried out with the Volume Rendering Techniques for directly generating two dimensional image on screen by 3 d data field, it is intuitive aobvious
Show segmentation effect.
According to above-mentioned steps, fat slice in collected totally 60 abdomens handle and has counted dependency number
According to.Running environment is 8 system of Win, and 4G memory, GPU is Nvidia 740m.Statistical result is as follows:Average time:20.4s quasi-
True rate:SAT:0.937, VAT:0.886.As can be seen that the splitting speed of this method has compared with conventional method in from the above
It is obviously improved, and also ensures higher accuracy rate.