The discrimination system of pancreatic endoscopic ultrasonography image based on M-band wavelet transform

Abstract

Objective To develop and evaluate the digital discrimination system for pancreatic ultrasound endoscopy images. Methods EUS images of 153 pancreatic cancer and 63 non-cancer cases were selected. According to the multi-fractal feature vectors based on the M-band wavelet transform, we acquired the fractal features with lower dimension with the feature screening algorithm. With the optimal feature com- bination, cases were classified into pancreatic cancer group and non-pancreatic cancer group automatically. Then the sensitivity, specificity and accuracy of this method were calculated, and compared with those of tra- ditional 9 dimension fractal feature vectors. Results Three kinds of muhi-fractal dimensions were intro- duced to the framework of M-band wavelet transform according to the EUS images to form fractal vectors of 18 dimension. With the selection by sequence forward search （SFS） algorithm, 7 dimension of feature vectors were chosen and were combined with bi-order muhi-fractal dimension to a better feature combination. The Bayes, support vector machine （SVM） and ModestAdaBoost classifiers were introduced to evaluate the clas- sification efficiency, resulting in a classification accuracy of 97.98% and short running time of 0. 49 s with lower feature dimension. Conclusion These data suggest the feasibility, accuracy, noninvasiveness and efficacy of classification of EUS images to differentiate pancreatic cancer from normal tissue based on the Mband wavelet transform algorithm. It is a new and valuable research area in diagnosis of pancreatic cancer. Key words: Pancreatic cancer; Endoscopic ultrasonography; M-band wavelet transform; Multifractal