Deformation-Constrained Active Scanning for Frozen Ultrasound Models

A diagnostic network that matches experts on curated scans often fails on a cheaper handheld ultrasound probe. Regulators also usually forbid updating its weights after approval. The current answer is to repair the images after acquisition, for example by translating handheld frames into the style of a high-end cart machine before the frozen model sees them. When a robot holds the probe, a different option exists: move the probe so it records a better image in the first place. The obvious version of this, steering the robot to wherever the frozen model is most confident, has a hidden failure. Soft tissue deforms under the probe. Pressing harder improves acoustic contact and probably raises the model's confidence, but it compresses vessels and other anatomy. The robot can then "succeed" by producing confident measurements of distorted anatomy. Prior phantom work found probe-induced vessel diameter errors above 80% before correction (Deformation-Aware Robotic 3D Ultrasound).

We propose a robotic scanning policy that chooses probe pose and contact force to reduce a frozen model's uncertainty about a clinically used measurement. The search is constrained by a learned, patient-specific prediction of how much the tissue is being deformed. The application is carotid intima-media thickness (IMT, the thickness of the inner artery wall layers, a cardiovascular risk marker whose value depends on compression), measured with a handheld-grade transducer. Spine curvature (Cobb angle) on bone serves as a control case where deformation should matter little. The central claim is falsifiable: unconstrained confidence-seeking acquisition will show a measurable force-dependent bias in IMT and lumen diameter that the deformation constraint removes. The constrained policy should match or exceed post-hoc image translation without touching the model's weights.

Core Claim

The frozen model's confidence is an exploitable target. Ultrasound physics links contact force to both image quality and anatomical distortion, so any acquisition loop that maximizes diagnostic confidence alone will drift toward high force. A useful active-acquisition system must therefore keep two quantities separate. One is the epistemic uncertainty of the frozen model, which is what the robot should reduce. The other is the deformation the robot is causing, which it must bound and, where it cannot avoid it, correct.

Concretely, at each step the robot picks an action $a$ (a probe pose and a normal force) to solve

$$ a^\star = \arg\max_a \; \big[ U(x_t) - \mathbb{E}\,U(x_{t+1} \mid a) \big] \quad \text{s.t.} \quad \hat{D}_\theta(a) \le \delta , $$

where $U(x)$ is the frozen model's predictive uncertainty on the IMT estimate from image $x$, and $\hat{D}_\theta(a)$ is the predicted peak tissue strain along the vessel wall under action $a$. That prediction comes from a deformation model with patient-specific stiffness parameters $\theta$. The constant $\delta$ is a strain tolerance set from phantom calibration. The tolerance is a clinical choice, not a learned one.

How This Differs From Existing Work

Prior systems split into three groups.

The first group controls force only to keep acoustic contact. FARUS holds force at 2–4 N and orients the probe by image entropy before producing thyroid risk (TI-RADS) scores. UltraDP learns a force-aware diffusion policy, a controller that generates trajectories by iterative denoising, and steers it toward centering the carotid. Neither looks at the diagnostic model's uncertainty, and UltraDP produces no diagnosis at all.

The second group repairs images without acting on the patient. TISA projects handheld frames into cart style with a frozen diffusion model. INPUTADAPTER does the analogous input-space adaptation for health-record time series. Both supply our key constraint, that the model stays frozen. Neither can recover signal that was never recorded.

The third group supplies methods: active sensing that reveals hidden tissue parameters by physical probing. JIGGLE infers tissue attachment boundaries by deliberate jiggling under an energy safety bound. EDO-Net encodes a short exploratory pull into a latent stiffness code that conditions a graph dynamics model. Jiang et al. fit a second-order stiffness curve from robot "pokes" and invert in-plane deformation.

We borrow the exploratory-probe idea to estimate $\theta$ in a few seconds per patient. We use a conditional graph network in the style of the cGNN for soft tissue as $\hat{D}_\theta$, because it predicts displacement fields in about 3 ms, which is fast enough for the control loop. The new element is aiming active sensing at a frozen diagnostic model's uncertainty, with deformation as an explicit constraint. None of the cited works does this.

Research Program

Three things are built.

The first is a stiffness-estimation routine. The robot makes a short palpation sequence at the start of each scan. That sequence is encoded, as in EDO-Net, into a latent that conditions the deformation network. The network is pretrained on synthetic mass-spring simulations and fine-tuned on phantom trajectories, following the cGNN recipe, which reported a 34% error reduction from simulation pretraining.

The second is an uncertainty estimate for a frozen IMT segmentation model that requires no weight changes. We use test-time input perturbations and Monte Carlo dropout if the certified architecture already includes it. TISA's dual-uncertainty filter is the reference implementation.

The third is the controller. It wraps UltraDP-style centering as a base behavior. A sampling-based optimizer, the cross-entropy method used in D-Cubed, proposes pose-and-force perturbations, scores them with the objective above, and rejects those violating the strain bound. After acquisition, any residual predicted deformation is used to correct the reported IMT and diameter, in the manner of Jiang et al.

The analysis component is a force-sweep study run before any control is built. On phantoms of known geometry, and on volunteers, we record the frozen model's confidence and its IMT and diameter estimates as a function of applied force. This tests the premise directly. If confidence does not rise with force, or the estimates do not drift, the Goodhart argument fails and the constraint is unnecessary. We will report that outcome and reframe the work around active search alone.

Baselines

The naive reference point is a source-only frozen model on handheld frames acquired at a fixed 3 N with UltraDP-style centering, and no adaptation. In TISA's carotid setting this is the configuration their method improves on. For spine, TISA reports a 35.76% F1 improvement over source-only on vertebral detection.

The strongest existing method is TISA applied to the same fixed-force robotic acquisitions. On handheld Clarius L7 carotid images, TISA reports 82.49% Dice for intima-media segmentation. On spine it reports 75.75% F1 for vertebral detection and a Cobb angle error of 2.6° ± 2.3° against expert measurement. It is the strongest competitor because it already respects the frozen-model constraint and outperformed TENT and COTTA by over 14% F1. It runs with the same frozen task model, the same robot-acquired frames, and its published threshold settings. We give it equal wall-clock budget per patient. Beating it means lower absolute IMT error against reference and lower force-dependent bias, at equal or better Dice. TISA is also composable with our method, so we report TISA on top of our acquisitions as well.

The acquisition-side competitor is FARUS-style entropy-driven probe orientation at constant force. It tests whether a generic image-quality signal does as well as diagnostic uncertainty. No reported carotid number exists on Lacuna for this exact configuration, so we reimplement it.

Two ablations isolate the key ingredients. Confidence-only active acquisition is our controller with the strain constraint removed. It isolates the deformation constraint, and we predict it achieves the highest confidence and Dice but the largest IMT and diameter bias. Passive dense sweep with best-frame selection has the robot traverse a fixed grid of poses and forces in the same time budget, after which we keep the frames on which the frozen model is most confident subject to the same strain bound. It isolates the value of closed-loop action. All variants share hardware, transducer, frozen model, time per patient, and controller tuning budget.

Why not just…?

Why not just translate images after acquisition, as TISA does? Translation cannot add acoustic signal lost to poor coupling or shadowing. UltraDP's ablation without force feedback fell to 0/20 successful scans, which shows acquisition itself fails without physical control. Translation also does nothing about anatomy that was compressed during acquisition. The experiment that settles this applies TISA to both fixed-force and constrained-active acquisitions. If TISA alone matches our IMT error, active acquisition adds nothing for this task.

Why not just sweep many poses and forces and pick the best frame? This is the real threat, and it may work as well. If the passive sweep with the strain filter matches the closed-loop controller at equal scan time, the honest conclusion is that the contribution lies in the deformation model and constraint, not in active control. We would report it that way. Closed-loop search should matter most when the good acoustic window is narrow, for example in patients with thick necks or with the lower-quality handheld transducer. We stratify results by those conditions.

Why not just steer by a generic image-quality score like entropy? FARUS shows entropy is a workable proxy for orientation. However, image quality and diagnostic uncertainty can disagree on exactly the frames that matter, and entropy contains no notion of deformation. The FARUS-style baseline measures this directly.

Why not just adapt the model's weights at test time? Approved devices generally cannot be updated online, and TISA found TENT and COTTA unstable on ultrasound streams, trailing it by over 14% F1. We include TENT only as a secondary reference.

Proposed Experiments

Phantoms come first. We use stiff and soft vascular phantoms matching Jiang et al. (about 3237 and 1489 N/m), where true lumen diameter and wall thickness are known. On those phantoms, Jiang's stiffness-aware correction reached Dice 0.89 on the unseen soft phantom versus 0.74 for a force-based model. That establishes cross-stiffness transfer as the test we must pass. The decisive check is whether the confidence-only ablation shows a positive association between applied force and the error of the reported diameter and IMT, and whether the constrained policy brings that error within the strain tolerance across both phantoms.

Next is a volunteer study with a handheld-grade linear probe on a robot arm, following UltraDP's protocol. The source model is trained on CUBS cart-system images. The reference measurement is a sonographer's light-contact cart-system IMT on the same day. Metrics are IMT absolute error, Bland–Altman bias against force, Dice, scan success rate, and agreement with guideline risk categories. That last metric addresses the call in Bridging the Clinical Translation Gap for outputs that change care rather than raise AUC.

The spine control uses the paired robotic and handheld lumbar dataset with CT ground truth, with Cobb angle compared to TISA's 2.6°. The prediction there is that the constraint is nearly inactive and confidence-only and constrained policies perform alike. If the constraint instead hurts on bone, it is over-conservative.

The proposal succeeds if three things hold together on carotid: unconstrained confidence-seeking shows clear force bias, the constraint removes it, and the constrained policy beats TISA on IMT error.