Non-invasive endometriosis staging prediction using integrated radiomics and spatiotemporal transformer model based...

Precise staging of endometriosis remains a clinical challenge, as current diagnosis depends almost entirely on laparoscopic visualization-an invasive procedure marked by considerable inter-observer disagreement and diagnostic delays. Existing non-invasive approaches, whether based on conventional machine learning with handcrafted features or end-to-end deep learning architectures, have shown limited success in capturing the complex spatiotemporal characteristics of endometriotic lesions on dynamic imaging. Traditional radiomics methods struggle to model temporal enhancement patterns, while pure deep learning approaches often overlook domain-specific textural signatures that correlate with disease severity. To address these limitations, we developed a computational framework that combines the interpretability of engineered radiomics features with the representational power of spatiotemporal Transformers, specifically designed for dynamic contrast-enhanced magnetic resonance imaging sequences. Our dual-pathway architecture employs cross-modal attention mechanisms that allow bidirectional refinement between texture-based descriptors and learned spatiotemporal representations, alongside adaptive weighting that adjusts feature contributions based on individual case characteristics. When evaluated on a multi-institutional cohort of 486 surgically confirmed cases, the model achieved 87.8% accuracy with a macro-averaged F1-score of 0.854 on the test set, substantially outperforming both traditional machine learning baselines and single-modality deep networks. Independent validation on 127 retrospective patients from an external center demonstrated consistent performance with 85.0% accuracy, suggesting reasonable generalization across different clinical settings. Attention weight visualizations revealed that the model appropriately focuses on anatomical structures documented in surgical staging criteria, including ovarian endometriomas, deep infiltrating sites, and peritoneal implants. By providing interpretable staging predictions with quantified confidence levels, this framework may support preoperative risk stratification and assist clinicians in identifying patients who would benefit most from surgical intervention, though its ultimate impact on treatment decision-making requires further retrospective evaluation.