Multimodal Feature-based Ancillary Diagnosis Of Liver Tumours

In China,liver cancer is one of the most common malignant tumors,but the onset of this disease is extremely insidious,and symptoms usually appear only in the middle and late stages.Consequently,patients can receive timely treatment and improve their chances of surviving liver cancer if it is detected early when symptoms are not obvious.In recent years,deep learning has shown great potential in assisting ultrasound imaging diagnosis,and many great advances have been made in using computer-assisted doctors for diagnosis,and smart medical care has become one of the most important application fields of artificial intelligence.Presently,deep learning and ultrasound-assisted diagnosis are being used to detect thyroid nodules,detect breast lesions and stage liver fibrosis,but liver ultrasound still has relatively few applications.Therefore,in this paper,we used clinical information such as relevant tumor markers and non-invasive findings such as ultrasound as a basis for a study on ancillary diagnosis of liver tumors based on multimodal features,which includes the following three aspects of the study:First,study the feasible method to check the data processing of raw data for patients with liver disease.As the first step in the experimental process,the cases were screened according to certain inclusion and exclusion criteria,and then the ultrasound picture base dataset and clinical data base dataset of liver disease patients were constructed sequentially.When constructing the ultrasound image dataset,the main task was to crop and label the original ultrasound images.When constructing the clinical data dataset,we mainly finish the summary and data processing of the original clinical data.Second,ultrasound pictures and clinical data were used to construct benign and malignant classification models for liver tumors.When using the ultrasound images,we first trained Efficient Net,Res Net50,Mobile Net,and VGG16 respectively,and then selected the Efficient Net with the best diagnostic performance according to the experimental results,added the soft attention mechanism and improved the network structure.When using clinical data to build models,the variables significantly affecting the model were first established by binary logistic regression statistical analysis and clinical experience screening,and then the processed data were fitted using logistic regression,random forest,and XGBoost.Third,design an effective multimode feature fusion strategy to construct a liver tumor assisted diagnosis model based on multimode features.First,the trained deep learning model is used to automatically extract the ultrasonic image features,and then is fused with the clinical information to form the multi-mode features.Finally,the multi-mode feature data set is used to fit the random forest and XGBoost classifier.Based on the idea of end-to-end network,the above three experimental processes are encapsulated into a model to realize the automatic diagnosis of benign and malignant liver tumors based on multimode features.In short,this paper will deep learning technology applied to benign and malignant liver tumor auxiliary diagnosis,using convolutional neural network to extract ultrasound information,fusion clinical data to build multimodal features,through the experiment can fit a more effective model,can help clinicians,especially inexperienced doctors to make timely and accurate diagnosis of patients with liver disease,has important practical significance.