Application Of A Model Based On Deep Learning For Predicting Recurrence Of Hepatocellular Carcinoma After Transarterial Chemoembolization

Background In China,primary liver cancer(PHC)is a kind of malignant tumor with high fatality rate,in which hepatocellular carcinoma(HCC)accounts for the vast majority,and its incidence is on the rising year by year.At present,the main treatment methods include surgical resection,liver transplantation,radiofrequency ablation and interventional therapy.However,due to the high recurrence rate and poor prognosis condition,the occurrence of intrahepatic metastasis in HCC post-operative patients has a very important impact on their survival cycle,which is also an important guiding factor for the selection of pre-operative treatment methods and the establishment of post-operative follow-up cycle.At present,the rapid development of artificial intelligence(AI)field has been widely used in three different stages of HCC clinical problems:early diagnosis,the selection of treatment methods and the evaluation of therapeutic efficacy.Therefore,it is of great clinical significance for HCC patients to combine AI and magnetic resonance imaging to establish a model for predicting post-operative recurrence of HCC patients,to explore the survival indicators for efficient prediction of post-operative recurrence,to implement more convenient post-operative management measures,and to extend the survival cycle and improve the prognosis.Objective Through a combination of visual attention mechanism and deep learning(DL)technology,using magnetic resonance imaging,a preliminary AI model has been established and repeated verification by cross-validation,to evaluate the predictive value of this model for patients with HCC whether occurs intrahepatic metastasis after the treatment of Transcatheter Arterial Chemoembolization(TACE).Method In the study of the application of AI model in predicting the recurrence of HCC,from January 2015 to December 2019,a total of 161 patients with HCC diagnosed by puncture pathology or clinical diagnosis and underwent TACE treatment in Shandong Provincial Hospital were retrospectively enrolled as research subjects.Through investigation and follow-up to find out whether the patients had intrahepatic metastasis within 3 years after discharge,the patients were divided into non-metastasis group(n=79)and intrahepatic metastasis group(n=82).All the above patients used magnetic resonance scanners to collect images before receiving TACE treatment,including conventional and enhanced MRI sequences of the upper abdomen.The magnetic resonance images of FS-T2WI(Blade)sequence and TIWI-vibe arterial phase sequence are input into itk-snap v3.8 software in DICOM format,and the largest tumor lesions were selected for the region of interest(ROI)segmentation,the marked binary images of the above sequence are output to.NTTFI format.Using the Python software to write a program,combined with DL technology and Vision Transformer architecture to establish an prediction AI model,and by adopting the method of 10%discount cross verification to match the accuracy index,obtained the accuracy,sensitivity and specificity,jointly evaluate the prediction efficiency of the AI model.At the same time,13 clinical variables were collected:gender,age,status of liver cirrhosis,hepatitis,glutamic oxaloacetic transaminase(AST),glutamic pyruvic transaminase(ALT),total bilirubin(TBIL),albumin(ALB),blood urea nitrogen(BUN),serum creatinine(Scr),number of tumors,tumor size,post-operative survival time or tumor-free survival time after surgery,to analyze and judge the clinical factors affecting the occurrence of intrahepatic metastasis after TACE treatment from ditferent angles,and to predict the possibility of patients with post-operative intrahepatic metastases.The above collected data were analyzed by SPSS 26.0 software for statistical univariate analysis and multivariate Cox survival regression analysis,and P<0.05 was considered to be statistically significant.Result ①According to the characteristics of clinical variables,univariate analysis showed that age,hepatitis,number of tumors and tumor size were risk factors for post-operative intrahepatic metastasis(P<0.05),and gender,status of liver cirrhosis,glutamic oxaloacetic transaminase(AST),glutamic pyruvic transaminase(ALT),total bilirubin(TBIL),albumin(ALB),blood urea nitrogen(BUN),serum creatinine(Scr),post-operative survival time or tumor-free survival time in patients with HCC had no significant effect on intrahepatic metastasis after TACE treatment.② Multiariable Cox regression analysis showed that tumor size(≤5 or>5cm)was an independent risk factor for predicting intrahepatic metastasis(HR 1.747,95%confidence interval,1.039 to 2.938,P=0.035),and age(≤55 or>55 years old),hepatitis and the number of tumors(single or multiple)on the survival time of HCC patients after TACE treatment was not statistically significant.③ Under the 10-fold cross-validation calculation,the highest correct rate of the AI prediction model is 0.8125,the lowest correct rate is 0.5625,the average correct rate is 0.6897±0.0818(95%confidence interval,0.659-0.841),the sensitivity is 0.732,and the specificity is 0.771.Conclusion ①By establishing an AI model combined with visual attention mechanism and DL algorithm,we can extract the imaging features of tumor images of HCC patients and effectively predict the high-risk recurrence of HCC after TACE treatment.These patients should be closely followed up after treatment and discharge,and guide the clinic to adjust the treatment mode in time.②Combined with the statistical analysis of preoperative clinical variables of HCC patients,it can be concluded that the largest lesion diameter is an independent factor affecting the postoperative survival time of HCC patients.This data may be used to evaluate the survival of HCC patients after TACE treatment.