| Objective(s):According to the criteria of pathological examination or more than two kinds of enhanced imaging examination or the confirmation of benign lesions after more than6 months of clinical follow-up.This study aimed to investigate the prediction value of three different models for predicting the benign or malignant of Focal liver lesions(FLLs).The models include the lesion model,the lesion-liver background model,and the lesion-liver background-clinical risk factors model.The study utilized deep learning networks to learn lesion ultrasound images and liver background images and clinical risk factors of FLLs patients.Methods:1.Case collection: A total of 581 cases of non-simple cyst liver focal lesions detected by liver ultrasonography from May 2019 to May 2021 at The Second Affiliated Hospital of Kunming Medical University were collected according to the inclusion criteria.After strict screening based on the exclusion criteria,542 patients were included in this study.2.Ultrasonic images collection: The ultrasonic image of the patients were stored during routine ultrasound examinations according to the predesigned storage standard.The ultrasonic images include maximum section of lesion,measurement of maximum section of lesion,dynamic image of lesion and hepatorenal section,longitudinal section of left liver lobe,transverse section of left liver lobe,right intercostal section with displaying gallbladder,right intercostal section with displaying the right branch of portal vein.3.Collection of clinical datas of patient: Patients who met the inclusion criteria after being evaluated by a sonographer filled out a follow-up form to collect their clinical data.Clinical data include patient’s age,history of hepatitis(chronic viral),history of extra-hepatic tumours,family history of malignant tumor,weight loss of5 KG within 3 months,CT and/or MRI examination results.4.Supplement of clinical datas of patient: The sonographer supplemented the AFP,history of hepatitis,CT and\or MRI of patient and test results.Two sonographers cross-checked the ultrasonic characteristics of the lesions.The ultrasonic characteristics include tumour size,tumour shape,tumour margin,tumour echogenicity,distribution of tumour echogenicity,vascular invasion,Ascites.5.Selection of risk factors of patients: After conducting univariate Logistic regression analysis and multivariate Logistic regression analysis on the two kind features,those with statistical significance(P<0.05)were selected as clinical risk factors for FLLs.6.Construction of the lesion model: The lesion ultrasonic images were input into an improved Efficient Net-Attention deep learning network to construct a lesion model through continuous training,and testing.7.To establish the lesion-liver background model: the Res Net50 network was utilized to extract features from three types of liver background ultrasound images.The liver background features were then merged and normalized with lesion features and inputted into the Softmax classifier to construct the lesion-liver background model.8.To establish the lesion-liver background-clinical risk factors model: The R language was employed to integrate the prediction probability values of the lesion-liver background model with the clinical risk factors through logistic regression,resulting in the construction of the lesion-liver background-clinical risk factors model.9.To compare the evaluation indicators of each model.The Receiver Operating Characteristic(ROC)curves of all models were plotted,and the Area under the ROC Curve(AUC),accuracy,sensitivity,and specificity were obtained for each model to compare the evaluation metrics.10.The lesion-liver background-clinical risk factors model was visualized.predictive performance and clinical applicability of this model are evaluated.The lesion-liver background-clinical risk factors model was visualized using a nomogram,and the calibration curve and Decision Curve Analysis(DCA)curve were drawn to evaluate the goodness of fit and clinical applicability of the model.11.Statistical analysis: We built the models on Python(Version 2022.3),Python3.9,and R(Version 4.2.2).The SPSS(Version 25)was performed the screening of clinical features and ultrasound features.Logistic regression model was used for univariate and multivariate analysis,and statistically significant differences were considered.The Delong test is used to evaluate whether there are Statistical differences between the AUC values of the three models.The AUC value>0.5indicates that the predictive ability of the model has clinical significance.The distribution of AUC at 0.5-0.7,0.7-0.9,and>0.9 indicates that the prediction ability is lower,moderate,and high.The Hosmer-Lemeshow test was performed to examine the goodness of fit of the lesion-liver background-clinical risk factors model.Results:1.Cases collection: A total of 542 patients were included,among which 228 were diagnosed as benign and 314 as malignant.156 cases were diagnosed with normal liver background,37 with liver cirrhosis,and 12 with fatty liver.We collected2168 lesion ultrasound images,10840 liver background ultrasound images,and clinical data of all 542 patients.2.Clinical risk features screening: Age,hepatitis,history of primary tumor,and AFP and lesion echogenicity,lesion echogenicity distribution,and major vessel invasion were identified as clinical risk factors for FLLs(P<0.05).3.Comparative predictive performance of deep learning models.The AUC was0.709(95% confidence interval(CI): 0.609-0.808),the specificity as 0.578,and the sensitivity was 0.762 of the lesion model.The AUC achieved 0.937(95% CI: 0.863-1),the specificity achieve 0.933,and the sensitivity achieve 0.885 of the lesion-liver background model.The AUC reached 0.994(95% CI: 0.998-1),the specificity reached 0.972,and the sensitivity reached 0.985 of the lesion-liver background-clinical risk factors model.Notably,the AUC,specificity,and sensitivity of the lesion-liver background-clinical risk factors model were higher than those of the other two models(Both Delong test P<0.05).4.The ROC curves of the three models demonstrated that the lesion model had a lower predictive performance,the lesion-liver background model had a better predictive performance,and the lesion-liver background-clinical risk factors model had the best predictive performance.5.Evaluation of the predictive performance and clinical applicability of the lesion-liver background-clinical risk factors model revealed a well goodness of fit(Hosmer-Lemeshow test P < 0.05)and a high degree of clinical applicability based on the DCA curve and calibration curve.Conclusion(s):1.In this study,the predictive performance of the lesion model constructed is lower.the lesion-liver background model indicate better predictive performance,while the lesion-liver background-clinical risk factors model have the best predictive performance.Despite the use of limited data,the models achieved basic or even high predictive performance.2.The addition of liver background features and clinical independent risk factors of this study improved the predictive performance of the models.The liver background features and clinical risk factors affect the predictive performance of the models.3.Analysis of the calibration curve and DCA curve of the lesion-liver background-clinical risk factors model demonstrated its goodness of fit and clinical applicability.4.The use of a nomogram to visualize the lesion-liver background-clinical risk factors model is more interpretable than traditional logistic regression analysis. |
PDF Full Text Request