Research On Classification And Recognition Of Medical Images Based On Microscopic Hyperspectrum

Traditional medical images are generally color images,although the images at the lesion can be obtained,the information that can be responded is too single.In addition,when using traditional medical images for pathological analysis and classification diagnosis,it largely depends on the personal judgment and experience of the physician,and has a strong subjectivity.Therefore,a method to assist physicians in classification and identification is urgently needed for medical image diagnosis.Hyperspectral images not only reflect the spatial information of the image itself,but also provide spectral information of the entire image.Compared to traditional medical images,they can provide more abundant information for medical work and help with pathological analysis.Therefore,this paper combines microhyperspectral imaging technology with medical image classification and recognition technology to investigate the classification and recognition method of blood cells and the recognition and segmentation method of cancer regions of thyroid tumor and bile duct cancer samples.1)We built a micro hyperspectral image acquisition device,collected micro hyperspectral image data of thyroid tumor samples,preprocessed the data,and combined ENVI software to complete the annotation of sample data.In order to accurately identify the different tissue regions in the two types of samples,this paper carries out the research of medical image classification and identification method based on spectral information.In this paper,we selected thyroid tumor and blood as the research objects,and used three different methods such as Spectral Angle Mapping（SAM）,K-nearest neighbor algorithm（KNN）and Support Vector Machines（SVM）.The classification methods were implemented to classify and identify thyroid tumors and blood samples.The overall classification accuracy was above 85%for thyroid tumors,and above 98%for blood samples.2)The traditional hyperspectral image classification methods only consider the spectral features of hyperspectral images and overlook the spatial features,which loses some information,leading to low classification accuracy.In this paper,a joint method of spatial and spectral information of hyperspectral images is proposed.Firstly,the Gabor features,LBP features and some statistical features of the image are extracted and noted as spatial feature vector F₁,which is combined with spectral feature vector F₂to obtain a one-dimensional spatial-spectral joint feature vector F.Secondly,three temporal encoding methods are used to encode the spatial-spectral joint vector F into a two-dimensional featured map.Then,a classification model combining Convolutional Neural Network（CNN）and a Support Vector Machine（SVM）is designed to first extract the spatial-spectral information features using Convolutional Neural Network,and then apply SVM to achieve the classification.Finally,the performance of the algorithm is verified by using bile duct cancer samples as the research object in this paper.The experimental results show that the accuracy of the three joint spatial-spectral classification algorithms proposed in this paper have an accuracy of over 85%for the classification and recognition of bile duct cancer.