DWI-derived intratumoral, peritumoral, and habitat features for preoperative prediction of lymph node metastasis in early-stage cervical cancer using machine learning method.

OBJECTIVES: This study aimed to develop a novel radiomic model by incorporating features from both habitat subregions and peritumoral regions to preoperatively predict lymph node metastasis (LNM) in early-stage cervical cancer using diffusion-weighted imaging (DWI). METHODS: 433 early-stage cervical cancer patients from four hospitals undergoing DWI were enrolled. Peritumoral regions were delineated by 1-4 mm expansion, and habitat analysis identified two intratumoral subregions, named Habitat 1 and Habitat 2 respectively. Intratumoral, peritumoral, and habitat features were extracted for model development. Prediction models included: Intra, Peri 1-4 mm, Habitat (1 and 2), and Fusion model. Performance was assessed via receiver operating characteristic curve, calibration, and decision curve analyses. RESULTS: Among the peritumoral models, the 3 mm peritumoral model demonstrated the best performance for LNM prediction, with AUCs of 0.867 (95% CI: 0.805-0.929), 0.747 (95% CI: 0.608-0.886), and 0.815 (95% CI: 0.743-0.887) in the training, validation, and test set, respectively. The Habitat 1 model also showed favorable performance, achieving AUCs of 0.838 (95% CI: 0.774-0.901), 0.712 (95% CI: 0.556-0.867), and 0.782 (95% CI: 0.694-0.869) in the training, validation, and test groups, respectively. Notably, Fusion model, combining Peri 3 mm and Habitat 1 features, achieved the best overall performance, with AUCs of 0.910 (95% CI: 0.868-0.953), 0.747 (95% CI: 0.600-0.894), and 0.837 (95% CI: 0.767-0.907) across the training, validation, and test sets, respectively and outperformed other models in calibration and decision curve analyses. CONCLUSION: The Fusion model enables superior and noninvasive prediction of LNM in early-stage cervical cancer patients.