The Research On Diagnosis System For Early Colorectal Cancer Based On Neural Network

Laser induced fluorescence(LIF) Spectroscopy is a new cancer-diagnosis method developed in last decade. Compared to other traditional diagnosis methods, it has many advantages, such as high-sensitivity, easy-accession, easy-to-operate, non-invasion, non-toxicity and non side effects, etc. All of these make LIF a novel method in the field of diagnosis and treatment for cancers in the early stages. This subject used Laser-induced autofluorescence diagnosis instrument for early-colorectal cancer as the research tool which is a key program of national "Tenth Five Project". All the research autofluorescence-spectrum samples of colorectal cancer were provided by National Hepatobiliary and Enteric Surgery Research Center, Ministry of Health. With the establishment of a rich experiment database, different fluorescence spectra can be analyzed and compared systematically. By adopting strictly Mathematical Statistics Principles in the analysis of Characteristics of fluorescence spectra and combining with BP neural network for spectrum-distinguished system, a self-adaptive fluorescence spectra diagnosis system was developed, which is very sensitive and special in clinical diagnosis. It offers a good basis for the invention of Intelligent Laser-induced Autofluorescence Diagnosis Instrument for Early-colorectal Cancer. To be more specific, all the work related to this dissertation is given as follows,(1) Mathematical Statistics Principles was first used in this dissertation. The amplitudes of 22 autofluorescence spectra samples of colorectal tissue were analyzed by t-test with SPSS. The amplitudes in all ranges of all these samples had no significant differences statistically. However, comparisons between autofluorescence spectra samples of cancerous tissues and normal tissues were made during 431-437nm,631-638nm,717-724nm.The result turned out be that most of samples had significant differences statistically (P<0.1), only a few didn't have. The result shows that distinguishment between normal tissues and cancerous tissues can not be made in statistical ways.(2) Based on the Preliminary result, BP neural network was adopted to distinguish 35 cancerous tissue samples and 6 normal tissue samples without pre-data cleaning and processing after the neural network training of 20 cancerous tissue samples and 6 normal tissue samples.The result was that the accuracy of Laser-induced autofluorescence diagnosis instrument for early-colorectal cancer based on BP neural network was 46.67%. This proved that only with BP neural network, the diagnosis instrument can not reach the clinical requirement.(3) In order to reduce the effect of noise data, pre-data cleaning and processing was conducted. A method for identification and calculation of amplitudes was put forwarded with a whole set of methods for data transforming and processing. Then with this basis, a mathematic discriminatory equation based on statistics was developed.The result was that with the method for identification and calculation of amplitudes mentioned in the dissertation, the ranges of the peak-amplitude characteristics in all spectra can be separated rapidly. The mathematic discriminatory equation for peak-amplitudes was composed of kurtosis, overshoot and absolute error and relative error of smoothing, etc. That is to say, F= W1×Q1 + W2×Q2 + W3×Q3 +×Q4 + W5×Q5. If F< 32, then the sample was diagnosed as a normal colorectal tissue; if else, the sample was a cancerous colorectal tissue. 35 samples was diagnosed at random, the accuracy was 93.4%.(4) Use BP neural network with pre-data cleaning and processing to diagnose between cancerous tissues and normal tissues.The conclusion was that with the modified BP neural network in the 35sample-distinguished cases, the results of 34 cases were right except one. Theaccuracy reached 97.1%.The research of this subject accelerates the process of applying Laser-inducedautofluorescence diagnosis for early-colorectal cancer in clinical use.