Predicting the Biological Behavior of Cervical Squamous Cell Carcinoma: A Machine Learning Approach Using Apparent Transverse Relaxation Rate (R2* maps) Radiomics Nomogram

Abstract Background Accurate prediction of the biological behavior of cervical squamous cell carcinoma (CSCC) is essential for optimizing therapeutic strategies and enhancing patient outcomes. This study aims to develop and validate a radiomics nomogram based on the apparent transverse relaxation rate (R2* maps) to predict deep stromal invasion (DSI), lymph node metastasis (LNM), and lymph-vascular space invasion (LVSI) in CSCC. Material and methods A total of 136 patients with CSCC were included. Patients were divided into two groups for each clinical characteristic: DSI (n = 61) vs. non-DSI (n = 75), LNM (n = 24) vs. non-LNM (n = 112), and LVSI (n = 61) vs. non-LVSI (n = 75). Radiomic features were extracted from axial MRI scans using the ESWAN sequence and post-processed to generate R2* maps, yielding 1,476 features. Clinical factors, including age, tumor diameter, SCC antigen levels, NLR, PLR, WBC count, Hgb level, tumor differentiation grade, history of irregular vaginal bleeding, and menopausal status, were also analyzed. Feature selection was performed using PCC, univariate logistic regression, ANOVA, RFE, and Relief. Clinical risk factors were identified via logistic regression. Machine learning classifiers were used to construct radiomics models, and combined models were developed by integrating the radiomics score with clinical factors. Model performance was assessed using ROC analysis, with AUC, accuracy, specificity, and sensitivity calculated. Radiomics nomograms were constructed for each condition. Results In the training set, radiomics models outperformed clinical models for DSI (AUC 0.824 vs. 0.724), LNM (AUC 0.827 vs. 0.783), and LVSI (AUC 0.897 vs. 0.712). Combined models further improved performance, with AUC values of 0.870 for DSI, 0.837 for LNM, and 0.876 for LVSI. In the testing set, combined models outperformed clinical models for LNM (AUC 0.735 vs. 0.633) and LVSI (AUC 0.775 vs. 0.664), but not for DSI (AUC 0.619 vs. 0.677). Conclusions The R2*-based radiomics nomogram with machine learning outperforms clinical models in predicting CSCC behavior in the training cohort. Combined models boost performance further. Although validation results are mixed, this method shows potential for enhancing personalized treatment and patient management.