[Intratumoral and peritumoral CT radiomics for predicting PD-L1 expression and pathological complete response after...

Objective: To construct a CT radiomics model based on intratumoral and peritumoral regions of hypopharyngeal carcinoma for noninvasive preoperative prediction of programmed death-ligand 1 (PD-L1) expression, and to evaluate its value in predicting pathological complete response (pCR) after neoadjuvant chemoimmunotherapy. Methods: A total of 364 patients with hypopharyngeal carcinoma who underwent contrast-enhanced CT and neoadjuvant chemoimmunotherapy from three medical centers, including Qilu Hospital of Shandong University, Qilu Hospital (Qingdao) of Shandong University and Weihai Municipal No.2 Hospital Affiliated to Qingdao University, were retrospectively enrolled. The intratumoral and 1-5 mm peritumoral regions of primary lesions were delineated, and radiomic features were extracted. The optimal features were selected by analysis of variance and least absolute shrinkage and selection operator (LASSO) regression. Multiple machine learning models including random forest, eXtreme Gradient Boosting (XGBoost) and support vector machine (SVM) were constructed with PD-L1 expression status as the endpoint. The predictive performance was evaluated by receiver operating characteristic (ROC) curve. The concordance index (C-index), net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were used to assess the predictive efficacy for pCR. Results: The overall positive rate of PD-L1 was 56.87%, and the pCR rate was 35.99%. The combined intratumoral+2 mm peritumoral feature model achieved the best predictive performance, with area under the ROC curve (AUC) values of 0.896, 0.782 and 0.754 in the training, internal validation and external validation sets, respectively. The C-index of the radiomics-clinical combined model for predicting pCR after neoadjuvant therapy reached 0.801. Using the clinical model as the reference, the NRI and IDI of the radiomics-clinical combined model were 0.471 and 0.184 (P<0.001), and its predictive performance was superior to the clinical-only model and the radiomics-only model. SHAP analysis demonstrated that intratumoral texture and morphological heterogeneity, as well as first-order gray-level statistical features of the 2 mm peritumoral region, were the key features driving PD-L1 prediction. Conclusion: The CT radiomics model combining intratumoral and peritumoral regions has certain predictive value for PD-L1 expression status and potential capability to evaluate pCR after neoadjuvant chemoimmunotherapy in hypopharyngeal carcinoma.