Deep Learning Radiomics Nomogram Based On Enhanced CT To Predict The Response Of Neoadjuvant Chemotherapy On Metastatic Lymph Nodes In Locally Advanced Gastric Cancer

ObjectiveNeoadjuvant chemotherapy(NACT)and radical gastrectomy have become the mainstay therapies for locally advanced gastric cancer(LAGC).However,not all patients with LAGC respond well to NACT.Timely and accurate determination of therapeutic response in patients with LAGC treated by NACT is important for individualized treatment.We aimed to construct and validate a deep learning radiomics nomogram based on baseline and restage enhanced computed tomography(CT)images and clinical characteristics to predict the response of NACT on metastatic lymph nodes in patients with LAGC.MethodsWe prospectively collected 112 patients with LAGC who received NACT from January,2021 to August,2022.We established and compared three radiomics signature based on arterial phase,venous phase and delayed phase CT images before and after NACT which are radiomics-baseline,radiomics-delta,radiomics-restage.We compared the three radiomics signatures and chose the signature with the best prediction performance.Then,the clinical model and the deep learning(DL)model were constructed separately,and the clinical features,the optimal radiomic signature and deep learning features were combined by multivariate logistic regression to create a nomogram to predict the response of LAGC with metastatic lymph nodes after NACT.Finally,the area under the receiver operating characteristic curve(ROC)curve was used to compare the predictive value of the clinical model,the DL model and the nomogram,and the clinical utility of the three models was assessed by decision curve analysis(DCA).ResultsAfter inclusion and exclusion criteria,a total of 98 patients were randomized 7:3 to the training cohort(n=68)and validation cohort(n=30).Three radiomics signatures were developed using arterial,venous,and delayed phase CT images before and after NACT,namely radiomics-baseline,radiomics-restage,and radiomics-delta signatures.In the training cohort and validation cohort,the AUC values were 0.86(95% confidence interval(95% CI),0.72-0.88)and 0.81(95% CI,0.71-0.85)respectively for radiomics-baseline,and 0.79(95%CI,0.74-0.86)and 0.78(95% CI 0.72-0.81)respectively for radiomics-restage,and 0.94(95%CI,0.85-0.97)and 0.86(95% CI,0.73-0.89)respectively for radiomics-delta.The radiomicsdelta was the best radiomics signature for predicting metastatic lymph node response.After multivariate analysis,the total number of cycles of NACT,the delta longest diameter of lymph node,and post-treatment CA-199 were used as independent risk factors to predict GR and developed the clinical model.A DL model was developed based on DL features.In both the training and validation cohorts,the AUC values were 0.90(95% CI,0.73-0.94)and 0.83(95%CI,0.76-0.85)respectively for the clinical model and 0.94(95% CI,0.85-0.97)and 0.86(95%CI,0.73-0.89)respectively for the DL model.Therefore,a DL delta radiomics nomogram(DLDRN)was developed by multivariate logistic regression of the total number of NACT cycles,delta lymph node longest diameter,post-treatment CA-199,DL features and radiomics-delta signature.The AUC values for DLDRN in the training and validation cohorts were 0.97(95% CI,0.78-0.98)and 0.94(95% CI,0.82-0.96),respectively,demonstrating adequate good response to NACT treatment.In addition,DLDRN significantly outperformed both the clinical and DL models in terms of predictive performance(P < 0.001).The superior net clinical benefit of DLDRN over the other models was confirmed by decision curve analysis.ConclusionIn LAGC patients receiving NACT,the DLDRN has proven to be effective in predicting therapeutic response on metastatic lymph nodes,which could provide valuable information for individualized treatment.