Research And Implementation On Pathological Images Fusion Recognition Based On Neural Network

With the rapid development of medical and computer image processing technology, automatic identification of medical image diagnosis is a current research focus in the cross field of computer image technology and medical image. Using computer image processing technology to complete the identification and pathological diagnosis, its purpose is how to construct a fast, high right rate pathological image recognition device and assist pathology experts in medical diagnosis. In this paper, human liver cell slice images are studied as the object, based on analysis of pathological image features, image fusion and recognition algorithm theory, feature extraction methods, based on pathological images, have researched and realized. Especially, a medical image classification and recognition algorithm was primarily designed and realized, which bases on neural network and combining feature-level data fusion with decision-level data fusion.In this paper, the research work is as follows:(1) Through summarizing home and abroad literature and analysis of the existing lack of pathological image recognition algorithm, a medical image classification and recognition algorithm frame has designed which bases on combining neural network with data fusion.(2) In order to make pathological image features more obvious and filter out noise disturbance, under the pathology expert guidance and help, through analyzing and summing up images characteristics, the original images pre-processing and simulation have been completed, including the gray transform, threshold, median filter, histogram equalization and image sharpening based on space-domain.(3) In order to make extracted features fully express the image content, through a lot of experimental comparison, the final features were extracted based on histogram color features 6-dimensional, based on combining wavelet packet with fractal texture features 18-dimensional and based on moment invariants shape features 7-dimensional. Together the 31-dimensional features form medical images feature vector.(4) In order to decrease reduce features' redundancy and improve the diagnosis recognition real-time, feature-level data fusion experiments, based on the principal component analysis method, have been completed, not only decrease the features' redundancy and reduce the dimension of feature space, but also reserve the required identification information.(5) A kind of medical image classification algorithm have been mainly designed and realized, which bases on neural network and combining feature-level data fusion with decision-level data fusion. First five kinds of recognition technologies, including self-organizing neural network, BP neural network, LVQ neural network, Bayes and Euclidean distance method, have been respectively adopted to complete medical image classification and recognition, finally the majority voting algorithm has been adopted to achieve the decision-level fusion recognition; Theoretical analysis and experimental results show that the recognition algorithm is effective to solve the lack that any feature can not well express medical images content, makes full use of the complementarities between different classifiers, further improves the medical images recognition rate.