Research On Liver Cancer Magnetic Resonance Imaging Classification Method Based On Residual Networ

Intrahepatic Mass-forming Cholangiocarcinoma(IMCC)is the second most common primary intrahepatic malignant tumor.Magnetic Resonance Imaging(MRI)is very important for the early diagnosis of IMCC.However,the morphological characteristics and microstructure of IMCC are easy to be confused with Hepatocellular Carcinoma(HCC),which is easy to cause misdiagnosis.The existing image classification and diagnosis methods of IMCC/HCC still have limitations.In addition,according to the characteristics of the medical image itself,the residual network can greatly improve the depth of the effective training network,which is more advantageous for training.Therefore,we intends to use two types of single-parameter MRI and multi-parameter MRI to carry out residual network classification diagnosis of IMCC/HCC.The details are as follows:(1)In the classification of IMCC/HCC single-parameter MRI,the existing research lacks attention to the edge features of the lesion and the poor fusion of multi-level features.At first,we propose a strided feature fusion residual network SFFNet,which retains more features of each level of the lesion by cross-layer shortcuts,optimizes the residual structure to improve the information loss of the network,and combines the attention mechanism to increase the attention to the key features.We also proposes new image preprocessing methods to highlight the edge features of IMCC/HCC lesions.The experimental results show that the classification accuracy of the proposed method is 92.64%,the Area Under The Receiver Operating Curve(AUC)is 0.9799,and it has high recall(91.05%)and precision(93.51%)for IMCC.Compared with the current similar methods,the proposed method shows better results.(2)In the classification of IMCC/HCC multi-parameter MRI,aiming at the problem that the model cannot enhance the identifiability of lesion features,we propose a camouflage classification residual network CCRNet.The model increases the feature extraction of the fusion multi-parameter MRI,enhances the identifiability and robustness of the lesion features by simulating the human receptive field module,and finally generates more discriminative fusion features from the feature information of each level.The classification accuracy of the proposed method is 96.97%,and the AUC is 0.9943,which is the best accuracy in this field by literature review,which proves the superiority of the model performance.The two new residual networks proposed in this paper have better performance in various evaluation indicators,and the feasibility and effectiveness of the models are further proved in various experiments.