From Intra- to Inter-habitat Analysis: Spatial Structure Features for Axillary Lymph Node Metastasis Prediction on Breast DCE-MRI.

Habitat radiomics characterizes texture within tumor subregions but discards the spatial relationships between them. We developed a landscape-ecology-inspired framework quantifying inter-habitat spatial structure on DCE-MRI, evaluated its complementary value alongside within-habitat radiomics, and compared early versus late fusion, using axillary lymph node metastasis (ALNM) prediction as the evaluation task. In this multicenter retrospective study, 284 patients from one center were randomly split into training (n = 198) and validation (n = 86) cohorts, with 120 patients from an independent institution as the external test set. Tumors were partitioned into six habitats by k-means clustering of 19 voxel-wise texture descriptors, with centroids derived exclusively from the training cohort. Three feature domains were constructed: 8 clinical parameters, 642 within-habitat radiomics features, and 94 inter-habitat spatial structure features spanning fragmentation (n = 30), morphology (n = 24), inter-habitat contact (n = 15), inter-habitat distance (n = 15), area ratio (n = 6), and diversity (n = 4). In the external test set, progressive fusion yielded incremental gains in AUC: clinical alone 0.754, plus within-habitat radiomics 0.843, plus inter-habitat spatial structure 0.863 (95% CI 0.792-0.934). The full early fusion model achieved higher AUC than the clinical-only (DeLong P = 0.011) and spatial-structure-only models (P = 0.044); early fusion yielded a numerically higher AUC than late fusion without reaching significance (0.863 vs 0.851, P = 0.636). The two single-domain imaging models were statistically indistinguishable (0.817 vs 0.784, P = 0.512). Inter-habitat spatial structure captures tumor organization overlooked by conventional habitat radiomics and provides complementary information for preoperative ALNM prediction.