AdaSwitch3D: Reliability-Aware Human-in-the-Loop Artificial Intelligence for Breast Tumor Segmentation in Dynamic Contrast-Enhanced MRI Across Unseen Centers

Background/Objectives: Deep learning models for breast tumor segmentation in dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) often lose accuracy at centers that did not contribute training data. We investigated whether case-wise routing between two low-resource adapted models, combined with an uncertainty-based review-prioritization signal, behaves consistently under severe cross-center domain shift. Methods: Using the four public collections of the MAMA-MIA dataset (DUKE, ISPY1, ISPY2, NACT; 1231 cases), we performed four-fold leave-one-center-out experiments in which each held-out center contributed only five labeled cases and a pool of unlabeled scans. We compared source-only transfer, five-shot fine-tuning, and teacher–student self-training with AdaSwitch3D, which routes each case between the two adapted branches using 17 label-free features and produces a heuristic review-prioritization signal. The complete pipeline—data splits, five-shot selection, training, adaptation, router construction, and referral calibration—was repeated with five random seeds, and each seed was evaluated once on 80 untouched test cases (400 seed-specific test observations). Results: Across five independent repetitions, mean Dice was 0.325 ± 0.051 for source-only transfer, 0.353 ± 0.052 for five-shot fine-tuning, 0.372 ± 0.047 for teacher–student adaptation, and 0.377 ± 0.057 for AdaSwitch3D; a validation-selected per-center strategy reached 0.376 ± 0.044. AdaSwitch3D exceeded five-shot fine-tuning in all five seeds and teacher–student adaptation in four, but the difference between AdaSwitch3D and teacher–student adaptation was small (mean +0.005) and varied by seed, while inference required both branches (approximately twice the single-model time). Observed test referral rates ranged from 23.8% to 56.3% against a 20% validation target, and low-Dice (<0.20) failure capture ranged from 21.7% to 71.8%. In the prespecified case-level analysis of seed 42, AdaSwitch3D improved on source-only transfer and five-shot fine-tuning (Holm-adjusted p=0.021 for each), whereas the comparison with teacher–student adaptation did not provide evidence of a difference (Holm-adjusted p=0.508). Conclusions: Absolute segmentation accuracy remained low, the incremental benefit of case-wise routing over teacher–student adaptation was small and seed-dependent, and referral behavior was unstable across repetitions. The principal finding is therefore that cross-center adaptation, case-wise model selection, and heuristic referral remain sensitive under limited target supervision. AdaSwitch3D should be interpreted as an exploratory case study of these challenges rather than as a clinically validated segmentation or referral system.