| Sumario: | Simple Summary: Cervical cancer continues to affect many women worldwide, with a considerable number experiencing the return of the disease after treatment. Standard imaging methods, while valuable for planning therapy, are limited in their ability to predict which patients are at higher risk of recurrence. In this study, we analyzed PET/CT scans using a computer-based approach called radiomics, which extracts detailed information about tumor shape, intensity, and texture that is not visible to the human eye. By combining these imaging features with clinical information and applying modern computer algorithms, we created models that predict the likelihood of remaining disease-free after treatment more accurately than current approaches. Our findings highlight the potential of advanced image analysis to improve treatment planning and follow-up care, moving closer to personalized strategies that may benefit women with cervical cancer. Background: Cervical cancer remains a major global health concern, with high recurrence rates in advanced stages. [18F]FDG PET/CT provides prognostic biomarkers such as SUV, MTV, and TLG, though these are not routinely integrated into clinical protocols. Radiomics offers quantitative analysis of tumor heterogeneity, supporting risk stratification. Purpose: To evaluate the prognostic value of clinical and radiomic features for disease-free survival (DFS) in locoregionally advanced cervical cancer using machine learning (ML). Methods: Sixty-three patients (mean age 47.9 ± 14.5 years) were diagnosed between 2015 and 2020. Radiomic features were extracted from pre-treatment PET/CT (IBSI-compliant PyRadiomics). Clinical variables included age, T-stage, Dmax, lymph node involvement, SUVmax, and TMTV. Forty-two models were built by combining six feature-selection techniques (UCI, MD, MI, VH, VH.VIMP, IBMA) with seven ML algorithms (CoxPH, CB, GLMN, GLMB, RSF, ST, EV) using nested 3-fold cross-validation with bootstrap resampling. External validation was performed on 95 patients (mean age 50.6 years, FIGO IIB–IIIB) from an independent cohort with different preprocessing protocols. Results: Recurrence occurred in 31.7% (n = 20). SUVmax of lymph nodes, lymph node involvement, and TMTV were the most predictive individual features (C-index ≤ 0.77). The highest performance was achieved by UCI + EV/GLMB on combined clinical + radiomic features (C-index = 0.80, p < 0.05). For single feature sets, IBMA + RSF performed best for clinical (C-index = 0.72), and VH.VIMP + GLMN for radiomics (C-index = 0.71). External validation confirmed moderate generalizability (best C-index = 0.64). Conclusions: UCI-based feature selection with GLMB or EV yielded the best predictive accuracy, while VH.VIMP + GLMN offered superior external generalizability for radiomics-only models. These findings support the feasibility of integrating radiomics and ML for individualized DFS risk stratification in cervical cancer.
|
|---|