A Study On Automatic Grading Of Hepatocellular Carcinoma In MR Images Based On Deep Learning Methods

With the continuous development of computer hardware and software,many new deep learning computing platforms are emerging,making deep learning widely used in all aspects of life,especially in the field of image recognition and classification.The multi-layer structure of the deep learning network can effectively express complex functions,so that it can learn image features with strong representation ability and improve the accuracy of image recognition.In the medical field,the research on the combination of computer-aided diagnosis system and artificial intelligence has attracted more and more researchers' attention.This paper applies convolutional neural networks to the automatic pathological grading of hepatocellular carcinoma based on enhanced nuclear magnetic resonance imaging.The pathological grading of hepatocellular carcinoma is of great significance for the clinical diagnosis,treatment and prognosis of liver cancer.However,the evaluation of hepatocellular carcinoma grade based on medical imaging is a difficult challenge for radiologists.Therefore,achieving objective automatic grading of hepatocellular carcinoma has important clinical significance.This paper describes the development process of deep learning and mature deep learning techniques,and deeply studies the advantages of convolutional neural networks on image classification.Secondly,this paper applies convolutional neural networks to evaluation of hepatocellular carcinoma grade based on enhanced MR images.In the experiment,the experimental data set was obtained from two different hospital clinical centers.The data preprocessing and enhancement are carried out before the training and testing of the convolutional neural networks during the experiment.Finally,the paper proposes a new deep learning framework SE-DenseNet based on DenseNet(densely connected convolutional network)and SENet(squeeze-and-excitation networks).In the framework of SE-DenseNet,the SENet is added as an additional layer between the dense blocks of the DenseNet.The quantitative experimental results prove that the classification performance of SE-DenseNet outperforms that of the DenseNet.Owing to the advantage of useful automatic feature learning by the SE layer,the SE-DenseNet can simultaneously handle useful feature enhancement and superfluous feature suppression.To a certain extent,the SE-DenseNet can mitigate the impact of feature redundancy caused by the DenseNet,whose classification performance is improved.