| The efficacy of microwave ablation in the treatment of benign and malignant lesions in the liver has been widely recognized by physicians and patients.In particular,it has become an important part of the comprehensive treatment of intrahepatic malignant lesions due to its advantages of minimally invasiveness,preservation of more liver tissue,and activation of immune response.However,during treatment process,the experience and skills of physicians have a great influence on the technique efficiency of treatment and the prognosis of patients.In this study,we pay attention to the characteristic imaging performance of the lesion debris within the damage range and the prediction of the ablation area.We hope to provide evidence that what the image performance of the debris should be through image registration;training deep learning models to improve existing model of planning system.propose a prediction plan that is closer to the real ablation area,and provide ideas for the refined and intelligentized development of microwave ablation treatment of intrahepatic lesions.Objective:(1)Explore the relationship between the low signal area on the enhanced magnetic resonance image and the lesion debris after microwave ablation of the liver lesion.(2)Based on the enhanced magnetic resonance images before and after treatment,the deep learning model is trained to predict the ablation area during the microwave ablation treatment of intrahepatic lesions.(3)The trained deep learning model is compared with the current clinically used SAR model and ellipsoid model,and the actual ablation damage area is used as the standard to compare the three simulated ablation areas.Materials and Methods:Prospectively collect cases of intrahepatic lesions treated by ultrasound-guided microwave ablation in our department from January 2018 to April 2019,and screen the clinical data and imaging data of the cases before,during,and after the treatment,and retain the contrast enhanced magnetic resonance imaging of liver data within 1 month before and after the ablation and relevant clinical information in our hospital.(1)Randomly select cases to mark the ablation damaged area,and the enhanced MRI image shows a low signal area.Using the method of image registration before and after treatment,compare the exact position of the low signal area and the lesion in the three-dimensional space.Determine whether the low signal in the ablation zone is a characteristic manifestation on the magnetic resonance image of the intrahepatic lesion after ablation treatment.(2)Select the cases of a specific ablation treatment process and randomly divide them into a training set and a test set.The image of the ablation area after treatment in the training set is used as the gold standard,and the image before treatment and the parameters in the treatment process are used as input data.Based on U-Net,a deep learning model is trained.(3)Use the trained deep learning model to predict the ablation damaged area of the test set,and also use the pre-treatment magnetic resonance image and ablation parameters as the input segment to compare the similarity between the deep learning prediction result and the actual damaged area after ablation.It compares the results with two commonly used preoperative planning models,and compares whether the preoperative planning results of the deep learning model are better than the SAR model and the ellipsoid model.Results:(1)Randomly select 42 cases of magnetic resonance images after ablation treatment for preand post-treatment image registration,using a combination of rigid registration and non-rigid registration,the DSC of this group of data is 95.23±1.77%,MSD is 1.35±0.82 mm.The average distance between the centroid of the suspicious debris and the lesion before ablation in the three-dimensional space was 5.70±3.19 mm,and the average ratio of the overlapping area to the suspicious debris was 0.66±0.29.(2)125 cases of data that meet the conditions are screened and randomly divided into two groups,of which 104 cases are in the training set,which are used to train the deep learning model.The trained model verifies the prediction effect of the ablation damaged area on the test set of 21 cases,and compares the ablation results simulated by the deep learning model,the SAR model and the ellipsoid model with the gold standard,showing that the deep learning model is accurate(0.95341 vs 0.852676)vs0.896302),recall rate(0.837983 vs 0.397118 vs0.322114),DSC value(0.885948 vs 0.524972 vs 0.455591),Jaccard value(0.809671 vs0.362105 vs 0.302795)are better than the other two models(p<0.05).Conclusion:(1)The low signal area in the ablation area shown on the magnetic resonance image after ablation is closely related to the space of the lesion before ablation,which is likely to be the characteristic image of the debris of the intrahepatic lesion after microwave ablation.(2)Compared with the current ideal thermal field model used for intrahepatic lesion microwave ablation preoperative planning,the trained deep learning model can obtain more accurate damage area prediction results before treatment.The future application potential of the model needs to be further explored.It provides a new solution for the intelligentized and individualized development of microwave ablation for intrahepatic lesions. |
PDF Full Text Request