[0028] The endoscopic imaging system and image processing method of mathematical algorithms embodiment microscopic bladder Example
[0029] A mathematical algorithm microscopic bladder endoscopic imaging system, includes an endoscope handle, the outer sheath 3, the structure of the image processing module and a display module, the endoscope handle and the outer sheath 3 is a schematic view of such figure 1 , An endoscope handle is provided with a direction control means 1, one end of the light source module 2 and the image receiving module, and an end of the handle of the endoscope 3 is connected to the outer sheath, the outer sheath 3 and the other end is provided with a curved portion 4 , the outer sheath 3 and the connection end of the curved portion 4 is a composite braided tube, the bent tube portion 4 snake bone two-way, two-way can the deflection angle of 270 °, the direction of the control means 1 is connected to the curved portion 4, curved portion 4 of the control bending direction, an end portion bent away from the outer sheath 4 is provided with three tip portion 5, the tip section 5 is provided with an image sensor, the image sensor is a CMOS image sensor, a CMOS image sensor package distal end a set of optical microscope head, a light source module LED light source, the LED light source through the light pipe to couple light into an outer sheath within 3 fiber tube, to provide a light source for the image sensor, the image sensor and the image reception module 2 is connected to the image reception module 2 and the image processing module is connected, the image processing module The display module is connected.
[0030] The image processing method according to a mathematical algorithm microscopy bladder endoscopic imaging system embodiment of the present invention: a light source module providing light to the image sensor, an outer direction control means control the angle of the bending portion 3 of the sheath 4, the image sensor acquiring images of a patient, generating image data to the image reception module 2, the image reception module 2 receives the image data to the image processing module performs image processing, image processing module sends the data after the image processing to the display module, displaying image information in a patient .
[0031] The image processing method of the image processing module is specifically:
[0032] S1, an image denoising using fractional diffusion equation; specifically fractional diffusion equation: Wherein, u represents an ideal noise-free image, the equation solving output; F represents collected noisy original image; [lambda] is a weight parameter, for controlling the image of the original retention effect; [mu] is a weight parameter, go to control the fractional order the degree of noise; ▽ α Operator is the fractional order; G α Fractional represents the convolution kernel for controlling denoising content; ▽ to gradient operators;
[0033] S2, using the TV-Stokes model image deblurring, the depth and 3D image reconstruction technique learning algorithm constructed by fusing chromatography small organs and edges of the 3D medical imaging small organs, malignant tumors, for accurate segmentation;
[0034] S3, Layer 2 structural equation model the statistical model constructing medical image defect, the defect detecting medical image, a method using completely quaternion matrix color image robustness filling model, restore defects medical image; two-layer structure having a first layer of the model equation structural equation model used was model
[0035] U i = Λ k0 Ω i0 + ε ik0 , I = 1, ..., n, k = 1, ..., K,
[0036] v i = Λ t Ω ti + ε ti , I = 1, ..., n, t = 1, ..., T,
[0037] Among them, λ k0 Λ t Indicates the observable variable u i , V i Potential variable Ω i0 Ω ti Relationship load matrix, ε ik0 Ε ti Indicates error item; second layer model adopts factor analysis model; potential variable 一 ti = Γ t F (ω) 1,i , ..., Ω t-1,i ξ ti ) + ξ ti Where γ t Is unknown parameter matrix, f () is a function vector, ξ ti It is a misuse; Where Z ti Is an ordered classification data, W ti It is unordered classification data to introduce continuous type auxiliary variables when processing ordered classification data. And threshold α and z ti Corresponding, the information of the order classification data is passed And α portray, introduce continuous assisted bits for disorderly classification data ti Determine W by the size relationship between its components ti When Y ti There is a lack value, remember R t,ij For missing display, if Y t,ij Return, then R t,ij = 1, if r t,ij No missing, then R t,ij = 0; When the lack is not negligible, the missing mechanism is not negligible to be p (r t,ij = 1) = g (Y ti ), Where g (Y ti About Y ti Functions, different functions represent different missing mechanisms, using great likelihood and Bayes methods for parameter estimates and model tests.
[0038] S4, constructed evolution mathematical models small organ tumors; specific methods evolution mathematical model of a small organ tumors constructed: the local linear embedding, and other metrics mapping ISOMAP algorithm of the information data within the cell dimension reduction, extracting characteristics of cells and tissue contours, Construction data panel tumor prediction precursory information; re regularization produce sparse model by L1 selecting important feature, then a parallel FP-growth algorithm Spark based on frequent patterns between items Unicom weight matrix, load balancing policy data packets , Spark each computing node configured to speed up the rate conditions frequent pattern tree to identify a characteristic pattern in accordance with frequent itemset; dynamic Bayesian network, Bayesian networks applied on a time element, seen as Bayesian probability networks spread out in time series in the described state of the cell change over time, at different times of the Bayesian network, the type of feature, property, state the number of cells varies continuously changed, determining different states of distribution and the variations in weight, polyhydric - multidimensional - multilayer scenario corresponding to a combination of a set of risk, a scene is mapped to each of the risk value, determines the scene multidimensional multidimensional mappings, the mapping relationship to find high dimensional scene matrix and vector cancer risk, to achieve different scenarios of risk measurement, risk of tumors judged.
[0039] In the present embodiment the objective function L1 embodiments regularization is: in Called loss function J0, α || w || 1 Is L1 regularization term, α is the regularization factor, L1 regularization is the absolute value of the weights and; sparse model is added L1 after loss function J0 regularization terms; sparse model sparse grouplasso algorithm for feature selection, n th wherein L is divided into groups, the objective function is: in, , Impose constraints on the L groups is a first penalty term, such that the regression coefficients toward zero, Applying a second constraint penalty term for all coefficients, the coefficient group that tends to 0, the packet thinning using subsampling algorithm and selection combining method, repeated running features on different subsets of data and wherein the sub-set selection algorithm, the final summary feature selection result, an important feature of the score is 100%, characterized useless score of 0; dynamic Bayesian network constructed specifically as follows: first construct a topology of existing algorithms based on the existing data set, and then extension structure adjusted, then the probability distribution of computing nodes, re-use extension network structure learning algorithm for structure improvement, adjustment of the probability distribution of nodes, improve the Bayesian network; Bayesian network node joint probability distribution P (U) can be expressed as the product of each child node Xi conditional probability of its parent node with respect to the set of [pi] (Xi) or PI (Xi) of:
[0040] Image before processing as in the present embodiment example figure 2 , The processed image as image 3 Indicated.
