Radiomics Study For Differentiating Borrmann Type ? Gastric Cancer From Primary Gastric Lymphoma

Background:As the two most common gastric malignancies,gastric cancer(GC)and primary gastric lymphoma(PGL),endoscopist and radiologist encounter difficulty differentiating these two malignancies,especially,for differentiating Borrmann type IV GC from PGL.It is necessary to explore a more valuable imaging biomarker.The term radiomics has become a hot spot of research,which play an important role in disease diagnosis,response evaluation and prognostic prediction.Whether CT-based volumetric radiomics could accurately differentiate Borrmann type? GC from PGL has not been reported.Purpose:To evaluate the feasibility and the value of the radiomics signature for differentiating Borrmann type ? GC from PGL in clinical practice.Methods:This retrospective study was supported by XX and approved by the institutional review board of our hospital,informed consent was waived.The patient cohort included 40 Borrmann type ? GC patients and 30 PGL patients.All of them with complete clinicopathologic data and pretreatment multiphasic dynamic CT imaging of the upper abdomen in our hospital.The subjective CT findings for each patient were independently evaluated and recorded by two associate senior doctors with rich experience in the interpretation of abdominal CT,which including gastric wall peristalsis,perigastric fat infiltration,lymphadenopathy below the renal hila and enhancement pattern.Inter-reader variability was evaluated by the kappa statistics.A subjective findings model was constructed by multivariate analysis.In addition,manual segmentation was conducted by the two readers and radiomics features were extracted and selected by the least absolute shrinkage and selection operator(LASSO)logistic regression model to bulid the radiomics signature.The interclass correlation coefficient(ICC)was used to estimate the reproducibility.A independent t-test or Mann-Whitney U test was used to assess the differences in age and radiomics signature between the two groups.The differences in gender and the subjective CT findings between the two groups were analyzed using a Chi-square test or the Fisher's exact test.The combined diagnosis model,was built with multivariable logistic regression analysis using the significant predictors from the univariate analysis as inputs.Diagnostic performance of the subjective findings model,the radiomics signature and the combined model for differentiating these two malignancies was accessed by the area under the curve(AUC),the sensitivity,the specificity and the accuracy.The performance of these models were evaluated by using a 7-fold cross-validation.The differences in the AUC between the models were assessed using the Delong test.Results:Gender,gastric wall peristalsis,enhancement pattern and radiomics signature were the significant predictors in the univariate analysis.For the CT subjective findings model,the radiomics signature and the combined diagnosis model,their AUCs(95%CI,confidence interval)were 0.806(0.696-0.917),0.886(0.809-0.963),0.903(0.831-0.975),respectively;their sensitivity values were 63.33%,86.67%,70%,respectively;their specificity values were 95%,82.5%,100%,respectively;their accuracy values were 81.43%,84.29%,87.14%,respectively.The CT subjective findings model showed relatively lower AUC,sensitivity,and accuracy compared to that of the radiomics signature and the combined diagnosis model.Radiomics signature demonstrated both high sensitivity and specificity.Additionally,the combined diagnosis model achieved a highest AUC,a highest accuracy,and an excellent specificity.However,Delong test showed there were no significant differences in AUC among these three models(the CT subjective findings model vs the radiomics signature,P=0.188;the CT subjective findings model vs the combined diagnosis model,P=0.051;the radiomics signature vs the combined diagnosis model,P=0.422).In addition,delong test showed differences in cross-validation AUC through pairwise comparison were not statistically significant(P = 0.065-0.279).Conclusion:Radiomics analysis could be used to accurately differentiate Borrmann type IV GC from PGL before treatment.As a alternative to the visible CT findings,it would help less experienced radiologists and is apt to integrate into daily work,therefore,it has the potential for broader clinical use.