Abstract
BACKGROUND: This study investigated the predictive ability of a transformer model utilizing intratumoral, peritumoral, and habitat features derived from pretreatment (18)F-FDG PET imaging to assess overall survival (OS) in patients with cervical cancer. METHODS: A retrospective analysis was performed using pretreatment PET data from 107 patients with cervical cancer across two medical institutions. The k-means unsupervised clustering algorithm categorized the tumor and its 4 mm peritumoral region into four distinct habitat subregions. Radiomic features were extracted from the intratumoral, peritumoral, and each habitat subregion to construct intratumoral, peritumoral, habitat, and combined transformer models. Model performance was evaluated using the area under the receiver operating characteristic (ROC) curve, calibration curves, and decision curve analysis. RESULTS: The habitat subregion 1 model demonstrated the highest performance. Among individual models, the habitat transformer model achieved the strongest results, with an external validation set area under the curve (AUC) of 0.778 (95% CI: 0.612–0.944), surpassing the intratumoral transformer model (AUC 0.714, 95% CI: 0.521–0.907) and the peritumoral transformer model (AUC 0.707, 95% CI: 0.517–0.896). The combined model further enhanced predictive accuracy, attaining a validation set AUC of 0.823 (95% CI: 0.677–0.970), and exhibited superior calibration and clinical applicability. CONCLUSION: This study highlights the efficacy of the transformer model based on (18)F-FDG PET habitat features in predicting cervical cancer prognosis. The combined model, integrating intratumoral, peritumoral, and habitat features, significantly improves predictive performance and provides valuable insights for personalized treatment planning. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12885-025-14977-1.