Intratumoral and peritumoral radiomics based on super-resolution T2-weighted imaging for prediction of normal-sized lymph node metastasis in cervical cancer

Background Preoperative identification of normal-sized lymph node metastases (LNM) remains clinically significant yet challenging in cervical cancer. Purpose To investigate the value of super-resolution T2WI-derived intratumoral and peritumoral radiomics for normal-sized LNM prediction in cervical cancer. Material and Methods A total of 257 patients from three sites of our hospital were divided into a development cohort (site 1, n = 97), a validation cohort (site 1, n = 42), and two internal test cohorts (site 2, n = 62; site 3, n = 56). Super-resolution reconstruction based on generative adversarial network was applied to all images. The volume of interest delineation encompassed primary tumor boundaries with outward expansions (1–5 mm increments) in super-resolution T2-weighted (T2W) imaging. Radiomics features were independently extracted from intratumoral and five peritumoral regions. The clinical, radiomics and combined models were built using multilayer perceptron. Model performance was evaluated through receiver operating characteristic (ROC) analysis and decision curve analysis (DCA). Results The IntraPeri3 mm radiomics model achieved superior discriminative performance compared to other radiomics models. The combined model integrated clinical variables (tumor size and squamous cell carcinoma antigen), intratumoral and peritumoral 3 mm radiomics features yielded optimal performance (AUC = 0.838 in the development cohort, 0.808 in the validation cohort, and 0.769 and 0.766 in the internal test cohorts). DCA confirmed the combined model's enhanced clinical utility across probability thresholds. Conclusion Super-resolution T2W-based radiomics aids in predicting normal-sized LNM in cervical cancer, especially the combined model incorporating clinical information, intratumoral and peritumoral 3 mm radiomics features demonstrates optimal diagnostic performance.