Preoperative prediction model of lymphovascular space invasion in endometrial cancer using ultrasound indicators and foundation model-based features: a multicenter study

Abstract

Preoperative assessment of lymphovascular space invasion (LVSI) in endometrial cancer remains unreliable, as biopsy samples are often superficial and fragmented. However, knowledge of this parameter in the preoperative setting could be clinically relevant for surgical planning and risk stratification. Most imaging-based predictive models rely on magnetic resonance imaging, whereas ultrasound, despite being widely available and low cost, has been scarcely explored for this purpose. In this study, we retrospectively analyzed 211 patients with histologically confirmed endometrial cancer from two Italian centers and developed predictive models based on preoperative ultrasound images processed through a transfer learning approach using a Vision Transformer foundation model. Deep radiomic features and ultrasound-derived indicators were evaluated alone and in combination for both binary and multiclass LVSI prediction. Model performance was assessed on an internal hold-out test set derived from the pooled two-center cohort. The combined approach improved performance, achieving a test Area Under the Receiver Operating Characteristic Curve (AUC) of 85.47% (CI: 73.16% − 95.71%) in the binary setting; for multiclass classification, a cascade of binary models outperformed native multiclass models, reaching an overall accuracy of 76.76% (CI: 62.79% − 88.37%) and a macro F1-score of 66.74% (CI: 49.44% − 82.21%). These findings suggest that ultrasound-based transfer learning provides promising non-invasive biomarkers for preoperative LVSI prediction in endometrial cancer, with the integration of ultrasound-derived indicators further improving performance and offering a pragmatic alternative to MRI.