Research On Recognition Method Of Microscopic Hyperspectral Image Of Cholangiocarcinoma Based On Convolutional Neural Network

Cholangiocarcinoma is a malignant tumor of the bile duct epithelium and is the second most common hepatobiliary and pancreatic tumor in the world.The onset of cholangiocarcinoma is insidious and often has developed to an advanced stage when diagnosed.Therefore,early diagnosis is very important.Clinically,histopathological diagnosis is necessary to determine the actual degree of tumor invasion.However,the characteristics of the morphological changes of histopathological images have brought challenges to expert diagnosis.The manual diagnosis process is tedious and timeconsuming,prone to misdiagnosis and missed diagnosis.The development of deep learning technology provides a new method for the recognition of microscopic pathological images.In addition,in contrast to color images,hyperspectral images can provide a wealth of spatial information and spectral information which reflecting the chemical composition of pathological slices simultaneously.Therefore,this thesis established a hyperspectral data set for the pathological tissue of cholangiocarcinoma,and performed deep learning technology to study the area recognition method of cholangiocarcinoma.This thesis established a cholangiocarcinoma hyperspectral data set containing the annotations of cancerous regions,and used this data set to study the area recognition method of cholangiocarcinoma.The thesis established a 3D-Dilated-U-Net network to identify cancerous regions and normal tissue regions in the microscopic hyperspectral image of cholangiocarcinoma,expanding the dilated convolution from the spatial dimension to the spatial-spectral dimension,and analyzed the recognition results of 3D dilated convolution with different dilated coefficients quantitatively.In order to identify the large-area cancerous area in the pathological image of cholangiocarcinoma better,a 3DD-NL-U-Net model was established by combining the Non-local module and 3DDilated-U-Net.The recognition performance was compared and analyzed.In addition,in order to verify the efforts of the spectral information do to the pathological tissues on model recognition,a 2D convolutional neural network was performed to identify one single-band cholangiocarcinoma images and hyperspectral cholangiocarcinoma images respectively.The results showed that the hyperspectral data set of cholangiocarcinoma established in this thesis can help to the study of tumor histopathology.The results showed that the 3D-Dilated-U-Net model performs better than ordinary 3D convolution and the 3D convolutional neural network that only performs dilated convolution on the two-dimensional image in recognition of the hyperspectral image of cholangiocarcinoma microscopic pathology.The 3DD-NL-U-Net model performed best among all models,and its regional recognition accuracy,precision,recall,Io U and F1 values for cholangiocarcinoma images were 0.8723,0.8375,0.7651,0.6609 and0.788,respectively.The above results show that the 3D dilated convolution and Nonlocal module can help to improve the model’s ability to recognize the region of the hyperspectral image of cholangiocarcinoma.According to the proportion of the cancerous area in the model segmentation results,it can provide a certain reference for the diagnosis and analysis of the degree of cancer cell infiltration in the pathological slices of cholangiocarcinoma.