Predictive performance of MRI and CT radiomics in predicting the response to induction chemotherapy in nasopharyngeal carcinoma: A network meta-analysis

Yongjie Jian¹Jiaxuan Peng²Gan Yang³Xiaojuan He¹Jing Wang⁴Hui Shi¹Qiyu Lan¹Zuogang Yang⁵Zhenyu Shu^2*

• ¹The Third People's Hospital of Sichuan Province, Chengdu, China
• ²Department of Radiology, Zhejiang Provincial People’s Hospital, Hangzhou, Jiangsu Province, China
• ³Shehong People's Hospital, suining, China
• ⁴Sichuan Nursing Vocational College, Chengdu, Sichuan Province, China
• ⁵Rangtang Country People's Hospital, Aba Tibetan and Qiang Autonomous Prefecture, China

Objectives: To evaluate the accuracy of different radiomics methods in predicting the response of nasopharyngeal carcinoma (NPC) to induction chemotherapy (IC). Methods: A systematic search was conducted in PubMed, Embase, Web of Science, and Cochrane Library. Radiomics studies utilizing CT and MRI were included in this network meta-analysis. The quality of the studies was appraised via the PROBAST, RQS, and IBSI guidelines. The sensitivity, specificity, and accuracy of different radiomics models were analysed. Results: Ten eligible studies involving 1550 subjects were included. The pooled sensitivity and specificity of the radiomics models were 0.86 (95% CI: 0.78-0.91) and 0.69 (95% CI: 0.62-0.75), respectively. The AUC based on the SROC curve was 0.83 (95% CI: 0.70-0.91). The predictive performance of each model was rated using SUCRA values. The MRI-based support vector machine radiomics model had the highest specificity, and accuracy, at 80.7% and 73.2%, respectively. The MRI-based SVM radiomics combined with clinical features model had the highest sensitivity (82.0%). Among the CT methods, the deep learning (DL)-based convolutional neural network model had the highest sensitivity, and accuracy, at 51.0% and 44.9%, respectively. The PROBAST showed that 7 studies were at risk for bias. Conclusion: This study synthesized existing evidence to confirm that radiomics serves as a viable exploratory tool for predicting IC efficacy in NPC. MRI-based nonlinear models and clinical-radiomics fusion models exhibit considerable promise, whereas clinical translation necessitates three critical steps: (1) standardized protocols following IBSI/METRICS/RQS guidelines; (2) prospective multicenter validation; and (3) investigating tumor microenvironment mechanisms. These measures will facilitate the transition of radiomics from technical exploration to clinical utility.