Study On Several Key Techniques Of In Vivo Non-invasive Spectral Analysis For Medical Treatment

Spectrometric detection technology has become an advanced biomedicalresearch because of its non-invasiveness, convenience, efficiency, etc. With thecontinuous improvement of spectral resolution and wavelength range expansion，it isfound that the measured spectrum of biological tissues contains more extensivemicrostructure information which can reflect the physiological and pathologicalchanges of some tissue cells, which makes the vivo spectroscopy detection technologymore feasible in medical noninvasive domain. However, there is a strong nonlinearrelationship and complexity between the objective and spectrum information of thebody tissue due to the individual differences and a large number of overlapping ofspectrum in the diagnosis of each component, which brings severe challenges tohigh-dimensional, strongly correlated spectral data analysis and processing. It is reallyin need to propose some appropriate intelligent data mining algorithms to extract theoptical properties of the measured target, so as to reveal the implicit objective rulesbetween the physiological and pathological changes in biological tissues andspectroscopic.Tongue is a key window to observe the functional changes and diseases, which isone of the best measuring points of noninvasive detection of human. It‘s an importantproposition to diagnose disease in clinical diagnosis of noninvasive medical test basedon the feature information of the tongues. In this paper, based on the application ofspectroscopy in the objective study, according to the inspection of tongue and thequantity detection of serum protein, it is dedicated to research on some keytechnology of intelligent analysis of high dimensional and complex fuzzy nonlinearmapping of spectral data and levels of serum protein. It provides ideas, methods andtechnical supports to promote this type of non-invasive spectroscopy in medicalresearch for further exploration and development.1. In view of the complexity and ambiguous mapping relationship betweentongue and the physiological and pathological information, considering the currentlimit of tongue information collection and the defect of important connotation losesdue to processing mode discontinuously to extract the mixed information, designed anew way of thinking that can improve information access through collecting the tongue hyperspectral information with microstructure organizational changes, whichcan extract specific spectral index group after combining with a variety of linear andnon-linear data mining algorithms to a blackbox, associating with physiological andpathological information and considering the mixed and overlapping information ofspectrum as a whole. So it explores a practical new model for TCM noninvasivedetection and provides a guideline for subsequent research.2. A noninvasive detection method based on three kinds of biochemical markersof near-infrared spectroscopy of tongues, the fitting nonlinear mapping ability ofdifferent modeling algorithms was analyzed. Using a variety of data mining method toestablish a quantitative model for predicting protein content, experimental analysishas proved that tongue-based NIR noninvasive detection of serum protein content isfeasible, and it is possible to provide an easy, non-invasive advanced method forclinical protein detection. It also validated that SVM can effectively resist nonlinearfactors in vivo determination and quantitative analysis, enhance the robustness ofmodel, so it can be used as the basis for evaluation of optimal wavelength selection inadvance.3. To solve the weakness and uncertainty of spectrum signal due to the severeoverlapping of the characteristic spectrum peak of each component, an extraction,combination and selection method is proposed for the band which has better nonlinearmapping capability. By this method, the criteria of nonlinear recognition ability ofeach wavelength or combination of wave lengths are designed as SVM predictionaccuracy of cross-validation, nonlinear interval selection method and adaptive geneticoptimization algorithm respectively. The former can lock the range of characteristicwavelength after rough screening for hyper-spectral data; the latter carefully selectsthe optimal band combination in these intervals through global optimization searchstrategy. Experiments prove that it can decrease the complexity of detection model oflevels of3kinds of serum protein, effectively improve the prediction ability ofnonlinear model and can further overcome the nonlinear factor caused by overlappingof spectral peaks and individual difference by applying this method on NIRwavelength selection of tongue.4. To improve the generalization capability and universality of model, the datamining in depth is necessary for the sample set with large scale and large dynamicrange. Based on this premise, the cost-effective parallel GPU platform is used, afine-grained parallel algorithms is proposed and developed for SVM cross validation to solve the computational efficiency issues of SVM cross validation in the mosttime-consuming part of intelligent analysis algorithm. After the test of variety ofmiddle and large scale benchmark set, the algorithm can fully schedule the parallelresource of GPU, effectively and concurrently execute calculation tasks of crossvalidation and significantly improve the computing efficiency, meanwhile it can alsoensure the accuracy of calculation, especially for density data like high-dimensionalspectral data, its performance is more obvious. It has provided technical support topromote the intelligent analysis algorithm to medium and large-scale spectral data indepth data mining. Furthermore, to solve the problem of no significant accelerationfor small sample data set, a GPU-based parallel SVM grid search strategy is proposedin advance. Experiments proved that it gained36.02times speedup in parameterselection tests of58cases of near-infrared spectral data.