Study On Lung Cancer Histopathological Analysis Based On Microscopic Hyperspectral Imaging And Deep Neural Network

Lung cancer is one of the most common malignancies,and its morbidity and mortality are usually the highest in the world.The therapeutic measures and prognoses for different subtypes of lung cancer vary,so it is crucial to identify the lung cancer subtypes.At present,histopathological diagnosis is still the ‘gold standard’ for the classification of lung cancer subtypes.Traditional histopathological diagnosis relies on the pathologist observing the histopathological section under the microscope and making a judgment.This process is complex,time-consuming,and closely related to the doctor’s experience and some subjective factors.Therefore,it came into being with analyzing and processing histopathological images by means of image processing technology.Compared with the traditional image,a microscopic hyperspectral image contains spectral information as well as spatial information,which provides a new approach for the classification of lung cancer subtypes.On this basis,the classification of pathological subtypes of lung cancer based on microscopic hyperspectral imaging is studied in this thesis.The research content of this thesis is mainly separated into the following three parts.Firstly,the spectral correction method based on Lambert-Beer Law and the spectral bands optimization method based on principal component analysis are used to preprocess the hypercube.In this way,the interference of system noise is reduced and the signal to noise ratio of the hypercube is improved.Then,in this thesis,a convolutional neural network based on the three-dimensional convolution and the convolution combination unit is proposed for the microscopic hyperspectral images of lung cancer sections.It is called 3D-CCU-CNN and is used to assist doctors in clinical to classify and identity the lung cancer subtypes.This thesis adopts the threedimensional convolution in the 3D-CCU-CNN to obtain the characteristics of the spectral dimension,which solves the problem that two-dimensional convolutional neural network is not applicable to the three-dimensional image.Besides,the convolution combination unit is designed to obtain the fusion of the features obtained by different convolution scales.Finally,aiming at the pathological feature of the lung cancer subtypes,this thesis studies the method of analyzing the pathological feature of microscopic hyperspectral images based on the 3D-UNet network and cell morphological feature.The image segmentation and the calculation of the pathological feature’s parameters can help explain the diagnosis and prognosis of the lung cancer subtypes to some extent.The experimental results show that the 3D-CCU-CNN model proposed in this thesis can classify the lung cancer subtypes more effectively compared with the commonly used two-dimensional models,with the overall classification accuracy up to 0.962,and the accuracy rate,recall rate and Kappa coefficient are not less than 0.920.The 3D-UNet that is used to segment cancer cells performs well,with an overall accuracy of 0.942.The study of lung cancer histopathological analysis based on microscopic hyperspectral imaging and deep neural network in this thesis can realize the automatic classification of lung cancer subtypes,give a certain explanation from the cell morphological feature,provide a new idea for the pathological study of lung cancer,and offer help and support to doctors to a certain extent.