The Research On The Application Of Micro-hyperspectral Imaging Technology In Neural Classification

Peripheral nerve injury is one of the common diseases frequently encountered during orthopedic clinical practice. In order to obtain optimum repairing effect, the most important issue to be addressed is how to identify motor nerve and sensory nerve accurately. In that the existing clinical methods have limitations, based on the microscopic hyperspectral imaging system developed by the laboratory, this thesis does some research on the feasibility of neural classification by combining spatial image and spectral information. With the help of interactive data language (IDL), lots of experiments have been done, and the main work is as follows:(1) To eliminate the system noise generated in the process of data acquisition, the thesis mainly introduces and compares two kinds of preprocessing algorithms, the joint correction algorithm and a preprocessing algorithm based on Beer-Lambert law. The results show that both of them perform well. However, combined with the analysis of k nearest neighbor (kNN) classifier, it is showed that the data processed by the joint correction algorithm has a higher classification accuracy.(2) In order to realize the feature analysis, the thesis proposes a pixel purity algorithm by a cross-comparison of the region of interest (ROI) and the pixel purity index (PPI), which can also be used in determining training data. Compared to N-finder, the new method is more applicable to the data containing multiple categories. For another, by analyzing the mean spectrum curve of pure pixels of different categories, the results show that whether it’s stained or unstained nerve sample, there are differences between motor nerve and sensory nerve, which provides basis for the feasibility of neural classification.(3) This thesis does research on neural classification mainly based on the distance metric learning algorithms and the hyperspectral statistical characteristics. Combined with kNN classifier, it also proposes a independent discrimi-native component analysis (I-DCA) method, which is based on the existing distance metric learning algorithms. The result shows that the new algorithm has good classification effect of peripheral nerve by comparing with other distance metrics. But on the whole, N classifier is time-consuming, and the performance for neural classification is not superior as well. Therefore, The thesis also studies support vector machine (SVM), maximum likelihood (ML) algorithm, and the new proposed extended procedure based on ML algorithm. Comparing with kNN classifier, they all improve the efficiency and classification performance greatly. Ultimately, comprehensive experimental results show that the ML algorithm combining with relevant component analysis (RCA) is best for the stained nerve sample, and for the unstained sample, the ML algorithm performs best.(4) Based on the comparative analysis of above algorithms, and considering the existing system software can only realize the micro-hyperspectral data acquisition, the thesis has builded the preprocessing algorithms and some excellent classification algorithms into the software to meet the requirements of non-professionals by combining strengths of C++ builder and IDL.