Answers are grounded in this paper and cite their sources.
Accurate segmentation of brain metastases remains difficult in the presence of small and multifocal lesions, inter-institutional MRI variability, and non-identical annotation schemes. We developed a three-dimensional target-image-free, target-ontology-aware framework that combines a source-aligned atomic tissue representation, deterministic categorical ontology projection, multi-ontology projection supervision, and training-only lesion-specific counterfactual supervision. Pretreat-MetsToBrain-Masks was used exclusively for model development, whereas 38 brain-metastasis cases from MOTUM were reserved for locked external evaluation without target-domain training, fine-tuning, normalization recalibration, threshold adjustment, probability calibration, or target-specific post-processing. For contrast-enhancing tumor, the primary cross-dataset endpoint, the proposed framework achieved Dice scores of 0.9037 on Pretreat (95% CI, 0.880–0.924) and 0.7676 on MOTUM (95% CI, 0.727–0.804). On MOTUM, the paired Dice difference relative to nnU-Net was + 0.0802 (95% CI, 0.034–0.126). The secondary harmonized hNCE and whole-abnormality endpoints yielded external Dice scores of 0.8426 and 0.8729, respectively; these projections represent operational cross-dataset harmonization rather than biological equivalence of the underlying tissue compartments. At the lesion level, external sensitivity reached 0.8648 with 1.0842 false-positive lesions per patient, while small-lesion sensitivity increased from 0.7061 with nnU-Net to 0.8063. Lesion-wise Dice remained 0.7246, indicating residual segmentation error despite improved lesion detection. These findings support the feasibility of combining ontology-aware projection with lesion-specific training under locked external evaluation, while broader generalization requires validation in additional independent and prospective cohorts.
No takes yet. Share an insight, caveat, or question.
Moradi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: