Research And Application On The Technique Of Computer Pattern Classification And Clustering Based On Medical Image

This research work is a part of the Research of Data Mining on Medical Image supported by the National Natural Science Foundation of China. Some key techniques and main algorithms of data mining — classification and clustering for the mammograms database are proposed, and a computer-aided medical diagnosing system on breast cancer is developed based on the original system. The main work is as follows:(1) Image PreprocessingA simple and effective median smooth filter with boundary holding is applied to eliminate noise in the digital mammograms. At the same time, to implement image enhance and improve the quality, A novel homogeneity measurement is used to enhance the contrast. Then it is contrasted with the histogram equalization technique. It is proved by experiments that an image can be enhanced without losing more information by homogeneity measurement.(2) Features ExtractionTo implement classifying the medical images as normal and abnormal classes, four GCMs are constructed in four directions and texture features independent of directions are extracted, further, 4 statistical features are added, and it is proved that the accuracy is improved by doing so. To implement classifying the abnormal medical images as benign and malignant tumor classes, a region growing method to extract accurate boundary of the breast tumor region was investigated, and the corresponding compactness, moment , Fourier descriptors and chord-length statistics are extracted, and it is proved by experiments that these features can describe the shape of breast tumor perfectly and are more effective in distinguishing the benign from the malignant tumor.(3) Classifier DesignA nonlinear proximal support vector machine ( NPS VM) classifier is proposed based on the proximal support vector machine (PSVM ) classifier. PSVM is not only runs faster than standard support vector machine classifiers but also is easy toimplement with satisfactory result for lower hardware. And, faced with the nonlinear real-life questions, NPSVM overcome the localization which PSVM have. It is proved by experiments that the training correctness and testing correctness of NPSVM are all higher than those of PSVM. (4) An Improved Density Based Clustering AlgorithmA reformative algorithm is presented by studying the clustering algorithm deeply. It keeps the good features of density based clustering method such as the ability of discovering clusters with arbitrary shape and insensitiveness to noise data, and it can also do efficiently when it face with the database with uneven distribution. Furthermore, it is linear time complex, so it can be used in mining very large databases. In this paper, this method is applied to the computer-aided medical diagnosing system on breast cancer system to help the doctor dealing with the cases in "batch" mode.