A multi-organ segmentation model trained on one set of MRI scanners can lose more than twenty points of Dice (overlap between predicted and true masks) on a scanner it has not seen. This was reported on the AMOS abdominal benchmark (AMOS). Hospitals cannot fix this by annotating new scans from scratch, because one volumetric scan takes a radiologist hours. Yet AMOS itself shows the way annotation is actually done: a model proposes a mask and radiologists correct it. This proposal makes the radiologist's correction the unit that an adaptation method spends its budget on. The query is "fix the left adrenal gland in scan 17." It is not a scattered set of voxels, and it is not "label all 15 organs in scan 17."
The idea in two sentences: when adapting to a new scanner, pick which (scan, organ) pairs a radiologist should correct by weighing how much each correction is expected to help against how long it is expected to take, with the cost predicted from the model's own proposal. Then fine-tune on these partially corrected scans with a loss that treats uncorrected organs as unknown rather than as background. The hypothesis is that this recovers more of the scanner gap per radiologist-minute than labeling a few whole scans. It should also beat uncertainty sampling, which piles queries onto large organs. The gain should be concentrated on the small organs (adrenal glands, duodenum) that AMOS reports as the persistent failures.
Core Claim
Two existing lines of work ask which labels to buy under domain shift: HALO and CBDA. HALO ranks pixels by a rarity signal times entropy. CBDA caps each class's share of queried pixels so that dominant classes do not absorb the budget. Both motivate this proposal, since they show that naive uncertainty sampling starves rare classes. Both also constrain it. They query isolated pixels in 2D driving scenes, they assume every pixel costs the same, and the file notes that scattered pixel queries are impractical for clinicians working on contiguous anatomy.
Two other papers supply the "model proposes, expert edits" pattern. SALA does this for clinical text adaptation, with a language model editing a small model's extractions and beating human active learning after 384 annotations. AMOS does it for its 3D labeling pipeline. Neither treats the edit itself as an object to be selected and costed.
This proposal claims something none of them test. Under scanner shift in 3D, the right query is an organ-level correction. Its cost differs by orders of magnitude across organs and scans and can be predicted before asking. Selecting by predicted gain per predicted cost, with a floor for small organs, is what makes a small expert budget go far.
Organ-level queries leave every queried scan partially labeled. That reintroduces the problem identified by LTUDA: if uncorrected organs are trained as background, the model learns that real anatomy is background. So the second, inseparable part of the claim is that the target fine-tuning loss must merge each unlabeled organ's probability with the background. A voxel labeled "not a corrected organ" is then penalized only if the model assigns it to an organ that was actually corrected in that scan.
Research Program
The source model is a 3D-UNet trained on AMOS CT plus MRI from a subset of scanners. The 3D-UNet is chosen deliberately, since AMOS reports it beats transformer backbones in this data regime (89.04% vs. 86.32% mean Dice on CT).
On each unlabeled target-scanner scan, the model produces a 15-organ proposal and a small ensemble of disagreement maps. The ensemble comes from test-time augmentation plus a few training seeds.
For every (scan, organ) pair, three quantities are estimated:
- Correction cost $c_{jk}$. This is the predicted number of slices on which the proposed boundary of organ $k$ in scan $j$ would need editing. It is computed from the slices where ensemble disagreement exceeds a threshold within the organ's predicted extent.
- Expected gain $g_{jk}$. This is the estimated Dice deficit of organ $k$ on the target scanner, taken from ensemble agreement across target scans. It is scaled by how atypical scan $j$'s proposal is for that organ.
- Selection. Pairs are chosen greedily by $g_{jk}/c_{jk}$ until a budget in edited slices is spent. A minimum share of the budget is reserved for the four smallest organs, borrowing CBDA's class-balancing logic at the organ level.
The radiologist, simulated by ground truth in the main study, replaces the proposal for each chosen organ with the true mask. The model is then fine-tuned with the merged-unknown loss described above. The fine-tuning regime is a controlled factor: normalization layers and decoder only versus all weights. Target data here amounts to a few scans' worth of corrections, and both AMOS and FEDD report that high-capacity updates overfit small medical cohorts.
There are two rounds of acquisition, so the cost predictor can be checked against realized costs after the first round.
Cost is measured in two ways. The primary measure is a simulated correction cost: the number of slices where the proposal and ground truth differ by more than a small boundary tolerance, summed over corrected organs. A whole-scan label is costed by the same rule applied to all organs, so the two query types are compared on one scale. Because slice counts are only a proxy, a small timing study is a secondary check. Two annotators correct a fixed set of proposals in a standard editing tool, which tests whether the proxy ranks real editing time correctly. If the study cannot be run, the cost claims stay explicitly about the proxy.
Baselines
The naive reference is the unadapted source model evaluated on the held-out scanner. AMOS reports a drop of more than 20% mean Dice on unseen MRI scanners. That figure sets the gap to be recovered, and every method's result is reported as the fraction of the gap closed.
The baseline a skeptic will find most convincing is whole-scan labeling. Spend the same simulated correction budget on fully correcting randomly chosen target scans, then fine-tune with the same schedule and the same choice of updated layers. A variant chooses scans by mean ensemble disagreement instead of at random. To beat it, the proposal must close more of the gap at equal budget, on mean Dice and on small-organ Dice, across at least three of four budget levels and across held-out scanner splits.
The strongest existing acquisition method is CBDA. Its class-balanced dynamic acquisition is ported from pixels to 3D regions, here the (scan, organ) pair. The comparison is fair because CBDA uses identical proposals, budget, loss and fine-tuning. Only its uncertainty-with-class-caps score replaces gain over cost.
HALO's radius-times-entropy score is included where its hyperbolic projection head can be attached to the 3D-UNet decoder. It gets the same tuning budget as the proposal: a fixed number of hyperparameter trials on a validation scanner, never the test scanner.
No reported result was found on Lacuna for active domain adaptation of 3D multi-organ MRI segmentation on AMOS. The closest reported figures are from driving scenes and are not comparable. CBDA reports a minimum-class gain of +4.6 IoU at a 20% budget, and HALO reports 74.5% mIoU at 5% on GTAV to Cityscapes.
A label-free baseline uses source-free adaptation in the style of MIMPB: an EMA teacher-student with a prediction bank to balance foreground and background, which reached 87.80% Dice on REFUGE to RIM-ONE-r. It is extended to 15 classes and run on the target pool. It also serves as an initialization that the active methods can start from.
Two ablations isolate the key ingredients. The first keeps organ-level queries but trains uncorrected organs as background, which isolates the merged-unknown loss. The second keeps the loss but selects by gain alone with uniform cost, which isolates the cost model.
Why not just…?
Why not just label a few whole target scans? This is the real threat, and the proposal is designed so it can lose to it. The argument for organ-level edits is that on a new scanner most of a proposal is already nearly right: large organs like the liver are segmented well even under shift, according to AMOS. Whole-scan correction therefore spends much of its budget confirming easy organs. If the cost proxy shows that most of a whole-scan correction's slices sit on the few organs that are wrong anyway, organ-level selection can only reorder spending and will not help much. The budget-curve comparison settles it. If whole-scan labeling matches the proposal within seed variance, that is reported as the finding.
Why not just use source-free adaptation and annotate nothing? MIMPB-style adaptation was shown on 2D optic disc and cup, a two-structure problem. The file notes that its prediction bank and pseudo-labels are unproven on multi-class 3D volumes. Under the documented miscalibration of segmenters out of domain (post-hoc MRI calibration), its pseudo-labels for small organs are likely to be confidently wrong. The experiment measures this directly by reporting how much small-organ Dice source-free adaptation recovers. It also tests whether initializing from it reduces the expert budget the proposal needs.
Why not just apply CBDA or entropy sampling at the organ level? This differs from the proposal only in ignoring cost. That is exactly why the ablation with uniform cost exists. If the gain-over-cost ratio adds nothing beyond CBDA-style class caps, the contribution shrinks to "organ-level queries plus the right loss." That would still be useful, but it is a smaller claim, and the paper would say so.
Why not just train jointly on more source data, including CT? AMOS reports that joint CT+MRI training adds 2.14% Dice on MRI, an order of magnitude smaller than the unseen-scanner drop. The source model here already uses joint training, so this gain is in the starting point rather than a competitor.
Proposed Experiments and Decisive Checks
The main study uses AMOS MRI with leave-one-scanner-group-out splits. A secondary harder shift is CT-only source to MRI target. Budgets are set at four levels, expressed as the simulated cost of roughly one, two, four and eight whole-scan corrections. Results use five seeds. Metrics are mean Dice over 15 organs, Dice averaged over the four smallest organs, worst-organ Dice, and fraction of the scanner gap closed per unit cost.
Three checks decide the outcome. First, at equal budget the proposal must beat whole-scan labeling on small-organ Dice at the lower budgets, where the hypothesis predicts the largest advantage; at the largest budget the two should converge. Second, the background-trained ablation should show depressed Dice specifically on organs that are often left uncorrected, which confirms the LTUDA failure mode in this setting. Third, the predicted cost must rank realized cost with a clearly positive rank correlation after round one. If it does not, the cost model is not doing work, and the method reduces to a CBDA variant.
A partial-label transfer check is also run. LTUDA's single-organ CT union is the source and BTCV is the target, with the same organ-level queries, to test whether the conclusions hold when the source itself is partially labeled.
References
- AMOS: A Large-Scale Abdominal Multi-Organ Benchmark for Versatile Medical Image Segmentation
- Hyperbolic Active Learning for Semantic Segmentation under Domain Shift
- Class Balanced Dynamic Acquisition for Domain Adaptive Semantic Segmentation using Active Learning
- Exploring LLM Annotation for Adaptation of Clinical Information Extraction Models under Data-sharing Restrictions
- Labeled-to-unlabeled distribution alignment for partially-supervised multi-organ medical image segmentation
- FEDD - Fair, Efficient, and Diverse Diffusion-Based Lesion Segmentation and Malignancy Classification
- Source-Free Domain Adaptation for Medical Image Segmentation via Mutual Information Maximization and Prediction Bank
- Improved Post-hoc Probability Calibration for Out-of-Domain MRI Segmentation