URDT-net: uncertainty-ranked dual-teacher learning with organ-topology consistency and budget-aware early exit for semi-supervised abdominal organ segmentation

Semi-supervised abdominal organ segmentation must use scarce annotations effectively while controlling anatomical errors and inference cost. Existing consistency methods generally use a single pseudo-label source and report efficiency only for a fixed full network. We developed URDT-Net, a 3D Res-UNet framework with two functionally complementary exponential-moving-average teacher views, uncertainty-ranked pseudo-label retention, organ-aware topology terms, and an intermediate segmentation head for confidence-based early exit. The revised formulation does not treat the two teachers as independently learned experts: both follow the same student, while complementary input views and region-dependent fusion produce morphology- and boundary-oriented targets. The topology objective combines a soft morphological-survival term for solid organs, soft-clDice for tubular structures, and a low-weight inter-class ambiguity term. At inference, the early head is accepted when a composite confidence score reaches a threshold selected on the held-out development split. Under a unified protocol in which every supervised method used the same 40 labeled volumes and every semi-supervised method used the same 40 labeled and 2,000 unlabeled volumes, the URDT-Net full path achieved mean Dice similarity coefficient (DSC) 0.8543, normalized surface distance (NSD) 0.8031, and 95th-percentile Hausdorff distance (HD95) 13.42 mm on the 50-case visible tuning cohort. The strongest comparator, bidirectional copy-paste, obtained DSC 0.8461, NSD 0.7934, and HD95 15.62 mm. We interpret this 0.82%-point DSC difference descriptively and do not claim statistical superiority. In the matched routing audit, the selected operating point (threshold 0.75) exited 38.7% of volumes and changed DSC from 0.8551 for the full path to 0.8502, while reducing measured mean latency from 6.14 to 4.60 s per volume and mean cost from 328.7 to 238.8 GFLOPs. Training was more expensive than Mean Teacher (73.4 versus 38.5 GPU-hours with eight Monte Carlo passes). URDT-Net provides a favorable descriptive accuracy–efficiency operating point on FLARE22, with the largest gains on small and low-contrast organs. The evidence supports an integrated benchmark-specific design rather than a universal claim of superiority. External validation and a fully observed accuracy sweep across routing thresholds remain future work; checkpoint-linked qualitative prediction overlays for six held-out development cases are included in Fig. 6 and the revision data package.