Classification Algorithm Of Liver Focal Lesions Based On Multi-mode Ultrasound Imaging

The high mortality rate of liver cancer is closely related to the time of diagnosis.The development of artificial intelligence technology has brought new opportunities for the early diagnosis of liver disease and provided more space for patients to be cured.In recent years,more and more studies have verified the value of multi-modal ultrasound under the framework of deep learning in assisting doctors in the diagnosis process,but most of the studies focus on breast and thyroid diseases,and there are few studies on liver diseases.In addition,compared with the single-mode ultrasonic diagnosis model,the feature fusion and optimization methods of multi-mode models also affect the classification performance.At the same time,most studies usually treat all modes equally,ignoring the fact that different modes make different contributions to disease diagnosis.To solve the above problems,this paper uses multi-mode ultrasound images,clinical data and other information to develop an automatic classification model for focal liver lesions.Related work is as follows:Firstly,a multi-modal hepatic focal classification model was constructed,which was mainly composed of four kinds of ultrasound,namely focal ultrasound,Doppler ultrasound,liver background ultrasound and text clinical features.Firstly,valid cases were screened according to certain criteria,and data sets were made according to the characteristics of each mode.For ultrasonic images,batch operations such as labeling,cropping and enhancement of ROI regions were carried out;After feature screening,table summary,assignment,missing value and normalization of text clinical indicators,data sets matching four modes were obtained.Secondly,the optimal feature extraction network was explored for each mode.Four sub-models were designed according to the idea of adding one mode at a time by using the B-ultrasound model as the baseline,and the feasibility of multi-modal diagnosis of liver disease was verified by comparing evaluation indexes.At the same time,it was found that the sensitivity of different diseases to the modes was different when the significance test was carried out between models.Secondly,parameter gradient adjustment strategy and modal weight index are designed for multi-modal models.By analyzing the optimization process of parameter updating,the unbalance problem was found when multiple modes were trained at the same time.In order to alleviate this imbalance,the gradient adjustment strategy was proposed to adapt to control the optimization of each mode,so that all modes could be fully trained.At the same time,a modal score index is proposed to verify the difference of the contribution of each mode to the learning goal,and the modal score of the sample is quantitatively represented.The importance law of the weight of benign and malignant modes is obtained,which is further applied to the model of benign and malignant diagnosis,so that the model treats different modes differently in the prediction and emphasizes the more sensitive modal information.The above strategies effectively improve the model performance,make the diagnosis process more scientific,and provide more ideas for the full utilization of multiple modes.