Research Of Tumor Tissue Classification Based On Medical Hyperspectral Imaging Analysis

With the rapid development of hyperspectral imaging technology and theories of precise medicine in recent years,the application of hyperspectral technology in close-range medical diagnosis has become a new research trend.Based on micro-hyperspectral technology,combined with spectral feature analysis and deep learning method,it can provide more auxiliary information for medical pathological diagnosis and more precise spectral features for biological histology.Therefore,in this paper,the micro-hyperspectral imaging system,standardized data acquisition process,micro-hyperspectral data pretreatment and analysis methods are described in detail.Taking the pathological section of gastric cancer as an example,the classification performance based on convolution neural network and migration learning model is deeply studied,and a set of integrated space-spectral joint classification and diagnosis method is proposed.At the same time,micro-hyperspectral technology is applied to hyperspectral virtual staining,bacterial microbiological detection and melanoma diagnosis,which shows a broad application prospect and research value.The main research achievements and innovations in this paper are as follows:(1)Standardized data acquisition process and preprocessing method based on micro-hyperspectral imaging technology.Including slice sample preparation and marking,data acquisition,experimental sample extraction,spectral data pretreatment and spectral analysis.According to the characteristics of the instrument used in this experiment and the unique texture structure of the collected hyperspectral microscopic images,the best data preprocessing and visualization scheme is proposed,which can be used for efficient model discrimination and pathological analysis in the later period.At the same time,the micro-hyperspectral imaging system and data acquisition system are described in detail.(2)A micro-hyperspectral database of 30 patients with gastric cancer was established.And CNN models were established based on this database.For the subtle spectral difference between gastric cancer and normal tissues,the 1D-CNN model was used to extract spectral features.The optimal model structure and parameters of 1D-CNN for hyperspectral application in gastric cancer are studied,including the effects of convolution layer,pool type,and the number of neurons in the full junction layer on the performance of the model.The final classification accuracy reaches 94.39%.At the same time,the individual differences of spectral characteristics of gastric cancer tissue are studied and compared with the medical characteristics.(3)Tumor tissue classification and diagnosis based on multidimensional convolution neural network and transfer learning model.Aiming at the complex texture structure and rich details of pathological tissue,the deep spectral features extracted by 2D-3D CNN model are combined with the cell space features,and the classification accuracy is improved to 97.57%.In view of the relatively limited training samples of hyperspectral medicine,a pre-training model based on VGG-16 is established by using the open data set of gastric cancer digital pathological section,and the prior knowledge of microscopic pathological tissue is fully learned.On this basis,An efficient space-spectral joint classification model is proposed by integrating data preprocessing,spectral analysis and one-dimensional and multi-dimensional CNN classification models.(4)The application of micro-hyperspectral technology based on space-spectral joint classification diagnosis model.Micro-hyperspectral technology is applied to more biomedical fields.Firstly,in view of the time-consuming and laborious staining process in traditional pathological diagnosis,hyperspectral virtual staining of unstained tissues is realized by combining unsupervised spectral markers with morphological information of single spectral image.Aiming at the detection of pathogenic bacteria in urinary tract infection,a classification and recognition method based on microscopic hyperspectral characteristics and CNN model was proposed.The classification accuracy of four types of bacterial sample reached 98.62%.In addition,the micro-hyperspectral characteristics of bacteria,skin melanoma,liver cancer and giant panda hair are also studied,which provides a reference for other related researches of tissue,organ and biological sample.