Research On The Recognition Of Cholangiocarcinoma Microscopic Hyperspectral Image Based On Deep Learning

Cholangiocarcinoma is a relatively rare but highly malignant tumor.Since the early symptoms are not obvious,patients often miss the best opportunity for treatment when they find themselves ill.Pathological diagnosis is the"gold standard"for the diagnosis of cholangiocarcinoma,often operated by the experienced physicians who perform cumbersome and time-consuming microscopic examination of pathological sections.In this process,misdiagnosis or missed diagnosis may occur due to the inadequate experience in observing slices or different judging criteria of physicians.The proposal of precision medicine and the rapid development of artificial intelligence provide a new solution to the above problems,and also put forward higher requirements for the amount of information of medical data.The analysis based on traditional color medical images can provide some reference for computer-aided diagnosis,but the amount of information contained in traditional color images is limited.As an emerging technology,hyperspectral imaging technology can simultaneously acquire the spatial and spectral information of the sample which need to be collected.Based on this,the paper studies the region recognition of cholangiocarcinoma microscopic hyperspectral image based on deep learning.This paper proposes a 3D-Res-CNN?3 Dimensional Residual Convolutional Neural Network?model to recognize cholangiocarcinoma regions,hepatic fibrosis regions,and other tissue regions in cholangiocarcinoma microscopic hyperspectral images.Firstly,this paper analyzes the composition units and network architecture of convolutional neural networks.Then,the structure of the network is designed based on the"spatial-spectral combined"characteristics of hyperspectral images.The 3D convolutional layer with dilated structure in the spatial dimension is used to extract features from the 3D hyperspectral data,and the receptive field in the spatial dimension is expanded without increasing the amount of network parameters.The 3D ConvPool layer implemented by the strided convolution operation contains trainable parameters,which can reduce network complexity and has more flexibility than the pooling layer without trainable parameters.Considering the contradiction between model performance and calculation efficiency,a residual module with a bottleneck structure is added,and the problem of gradient disappearance can be avoided at the same time.Finally,in order to optimize the training process of the model,a dynamic adjustment strategy of learning rate is proposed to suppress the possible oscillations in the accuracy curve of the validation set,further guarantee the convergence of the network,and lay the foundation for the subsequent region recognition of the image.The experimental results show that,the 3D-Res-CNN model performs better compared with the 1D-CNN?1 Dimensional Convolutional Neural Network?and 2D-Res-CNN?2 Dimensional Residual Convolutional Neural Network?in region recognition of cholangiocarcinoma microscopic hyperspectral images,and the overall accuracy,Macro F₁.and Kappa coefficient are 91.6%?0.841 and 0.777,respectively.Compared with the recognition results of the pseudo-color composite image,the overall accuracy of region recognition on the hyperspectral image is improved by about 12.5%,indicating that the hyperspectral image is helpful to improve the region recognition performance of different tissues.The area ratio of cholangiocarcinoma region and hepatic fibrosis region is calculated based on the region recognition results of cholangiocarcinoma microscopic hyperspectral images,which can provide a certain reference value for the auxiliary diagnosis of different lesion regions in the pathological section of cholangiocarcinoma.