Classification Of Hyperspectral Medical Microscopic Imagery Based On Convolutional Neural Network

Recent years,with the rapid development of hardware technology and software technology,hyperspectral-imaging technology has been widely used in many' fields.Hyperspectral-imaging technology has not only been successfully applied in military,agriculture,food security,atmospheric monitoring,but also gradually penetrated into the biomedical field.Hyperspectral imaging can extract spatial and spectral information of the observed object at the same time.Compared with traditional optical microscopic imaging,it can provide more information to assist medical staff in pathological analysis.Deep learning in recent years gradually become a hot research topic in the field of computer vision and the computer vision and medical image diagnosis of interdisciplinary research.Compared with traditional methods,the deep learning method can adaptively extract the high identification degree features according to the data,it can provide more accurate and objective identification results to assist medical staff to make further clinical diagnosis.The main work of this paper is as follow:1.In cooperation with the China-Japan friendship,the hyperspectral imaging technique was used to capture glomeruli hyperspectral image,and we invited the nephrologists to label the immune complex areas.Firstly,we analyses the collected datasets and do some pretreatment.Then,we applied Convolutional Neural Network(CNN)to the classification of hepatitis B virus associated membranous nephropathy(HBV-MN)and idiopathic membranous nephropathy(I-MN),the experimental result shows that the deep learning methods is supervisor compared with traditional methods.This work proved that hyperspectral-imaging technique combined with deep learning is expected to be a new method for the diagnosis of membranous nephropathy.2.In view of the shortcomings of CNN with single scale feature and weak direction description ability,we used the Gabor filter incorporate into the CNN inner structure,which combines a modulated Gabor wavelet and deep CNN kernels,name as MGCNN.This method can lead CNN filter to obtain multi-scale and multi-direction description features in the convolutional layers.Experiments on the classification of five types of leukocytes hyperspectral image data show that the proposed MGCNN has significantly improved the classification accuracy compared with the traditional classification algorithm and other state-of-art Gabor and CNN based methods,especially in the case of small samples.