Yazhou Zhu

Perspective paper · arXiv:2609.10001 · v2

From Few-Shot Segmentation to Clinician-in-the-Loop Medical Image Analysis

Decidable Self-Assessment as the Precondition for Interaction and Adaptation

Yazhou Zhu rangerzhuyazhou@gmail.com
Download PDF View on arXiv

Central claim

Scarce expert attention should be spent only where a bounded intervention is expected to reach a clinically better outcome.

Conventional few-shot medical image segmentation asks how much can be learned from a small support set. Cross-domain FSMIS asks whether that evidence transfers to unfamiliar acquisition domains. Both improve the evidence already available to the model.

Rare, ambiguous, and shifted cases expose a different problem: the decisive evidence may be missing. Before asking a clinician or changing itself, the system must determine whether a permitted intervention can plausibly repair the case.

A strict dependency order

Self-assessment comes first.

Promptable interfaces make interaction possible. Adaptation mechanisms make rapid change possible. Neither establishes when asking or changing is warranted.

01

Decidable self-assessment

Separate low-risk outputs, correctable failures, and failures outside the reachable set under the remaining interaction budget.

Output: accept, query, or defer

02

Selective interaction

Treat clinician time as a distinct, limited resource. Select clicks, boundary edits, reference cases, or text only when response-conditioned value exceeds burden.

Virtue: parsimony

03

Bounded adaptation

Update approved components within a trust region, reassess with an independent safety monitor, and roll back when benefit or capability retention fails.

Virtue: reversibility

Conceptual trajectory

From preserving evidence to diagnosing what is missing.

Sparse support representation and cross-domain robustness improve what the support set retains. Decidable self-assessment governs what happens when the evidence needed for a reliable decision is absent.

Four-stage trajectory from sparse support representation and robustness under shift to decidable self-assessment, selective interaction, and bounded adaptation.
Figure 1. The first two stages preserve available evidence; the third diagnoses missing evidence and licenses the fourth.

Selective interactive adaptation

Ask only when help is reachable.

A foundation model can provide the representation and prompt interface, while a separate controller governs interaction, stopping, reassessment, and rollback.

Selective interactive adaptation loop with a reachability gate, clinician response, bounded adaptation, frozen safety monitor, and rollback.
Figure 2. Each update is independently reassessed against clinical risk and protected capabilities.

Two memory horizons

Learn from a case without turning deployment into uncontrolled online training.

Layer A · per-case episode

Repair the current case

Support examples, confirmed masks, and interaction history remain local to the episode. Updates are bounded, reversible, and discarded after the case unless separately admitted.

Layer B · cross-case experience

Retain correction priors

What transfers across rare cases is not a disease-specific mask prior, but a reproducible correction prior over failure modes, gated by cross-reader agreement and governance.

Falsifiable research program

Six hypotheses, one dependency chain.

The program is sequentially abandonable. If multi-level risk and repairability cannot be validated without clinicians, downstream interaction and adaptation claims should not proceed.

Hypothesis dependency graph from multi-level risk and repairability through query value, representation update, correction-prior transfer, and bounded adaptation.
Figure 3. H1 and H2 form the foundation; every downstream hypothesis presupposes them.
  1. Multi-level risksupports accept, query, and defer decisions beyond voxel entropy.
  2. Repairabilityis estimable, externally valid, and distinct from predicted error.
  3. Dynamic queriesimprove value per unit of real clinician effort.
  4. Feedback updates task representationand improves unedited regions without propagation harm.
  5. Correction priors transferacross rare presentations where case-similarity memory does not.
  6. Bounded adaptationretains benefit while reducing drift, forgetting, and unsafe personalization.

Abstract

A decision problem, not simply a stronger segmentation model.

Few-shot medical image segmentation seeks to delineate unseen structures from a small support set, but its standard formulation fixes task-defining evidence before inference. This assumption is fragile under acquisition shift, atypical pathology, ambiguous boundaries, and poor image quality. Adding clinician interaction and rapid adaptation is not sufficient: the binding constraint is deciding when asking or changing is warranted.

We reframe FSMIS as a three-layer sequential decision problem. Decidable self-assessment separates errors that a bounded intervention can repair from those no admissible intervention can reach. Selective interaction allocates a distinct expert-attention budget by response-conditioned net expected value of information. Bounded adaptation emphasizes reversibility and independent safety reassessment rather than speed.

A complementary cross-case memory stores reproducible correction priors over failure modes rather than disease-specific mask priors. The framework links sparse support representation, cross-domain robustness, multi-level risk estimation, clinician feedback, and governed experience transfer. Six hypotheses and a minimal pilot make the foundational self-assessment claim falsifiable before a clinician study.

Citation

Use this work.

The paper is available as arXiv:2609.10001 in Computer Vision and Pattern Recognition. 

@article{zhu2026clinicianloop,
  title={From Few-Shot Segmentation to Clinician-in-the-Loop Medical Image Analysis: Decidable Self-Assessment as the Precondition for Interaction and Adaptation},
  author={Zhu, Yazhou},
  journal={arXiv preprint arXiv:2609.10001},
  year={2026}
}
BibTeX copied