Preoperative Prediction of Perineural Invasion Status of Rectal Cancer Based on Radiomics Nomogram of Multiparametric Magnetic Resonance Imaging

OBJECTIVES: To compare the predictive performance of different radiomics signatures from multiparametric magnetic resonance imaging (mpMRI), including four sequences when used individually or combined, and to establish and validate an optimal nomogram for predicting perineural invasion (PNI) in rectal cancer (RC) patients. METHODS: Our retrospective study included 279 RC patients without preoperative antitumor therapy (194 in the training dataset and 85 in the test dataset) who underwent preoperative mpMRI scan between January 2017 and January 2021. Among them, 72 cases were PNI-positive. Then, clinical and radiological variables were collected, including carcinoembryonic antigen (CEA), radiological tumour stage (T(1-4)), lymph node stage (N(0-2)) and so on. Quantitative radiomics features were extracted and selected from oblique axial T(2)-weighted imaging (T(2)WI), T(1)-weighted imaging (T(1)WI), apparent diffusion coefficient (ADC), and enhanced T(1)WI (T(1)CE) sequences. The clinical model was constructed by integrating the final selected clinical and radiological variables. The radiomics signatures included four single-sequence signatures and one fusion signature were built using the respective remaining optimized features. And the nomogram was constructed based on the independent predictors by using multivariable logistic regression. The area under curve (AUC), DeLong test, calibration curve, and decision curve analysis (DCA) were used to evaluate the performance. RESULTS: Ultimately, 20 radiomics features were retained from the four sequences—T(1)WI (n = 4), T(2)WI (n = 5), ADC (n = 5), and T(1)CE (n = 6)—to construct four single-sequence radiomics signatures and one fusion radiomics signature. The fusion radiomics signature performed better than four single-sequence radiomics signatures and clinical model (AUCs of 0.835 and 0.773 vs. 0.680-0.737 and 0.666-0.709 in the training and test datasets, respectively). The nomogram constructed by incorporating CEA, tumour stage and rad-score performed best, with AUCs of 0.869 and 0.864 in the training and test datasets, respectively. Delong test showed that the nomogram was significantly different from the clinical model and four single-sequence radiomics signatures (P < 0.05). Moreover, calibration curves demonstrated good agreement, and DCA highlighted benefits of the nomogram. CONCLUSIONS: The comprehensive nomogram can preoperatively and noninvasively predict PNI status, provide a convenient and practical tool for treatment strategy, and help optimize individualized clinical decision-making in RC patients.