Potent chemotherapy drugs poison the body broadly before they kill the tumor: this systemic toxicity is what dose-limits treatment, and a whole pharmaceutical literature now cages the drug in injectable gels or heat-triggered nanoparticles so that it stays where it was put (Hyperthermia-targeted rectal delivery of thermosensitive liposomal doxorubicin, Comparison of Conventional Chemotherapy, Stealth Liposomes and Temperature-Sensitive Liposomes in a Mathematical Model). But caging only works if the clinician can see that the cage stayed closed, and today that verification is done by hand or with fragile classical image processing (In vivo imaging and pharmacokinetics of percutaneously injected ultrasound and X-ray imageable thermosensitive hydrogel loaded with doxorubicin versus free drug in swine, Evaluating the echogenicity of ethyl cellulose-ethanol gel for tracking biodistribution during liver ablation). The verification problem is, at its core, a domain shift problem: the depot's appearance drifts over hours as contrast agents diffuse, it changes from scanner to scanner, and the depot itself degrades the image — dense gel casts acoustic shadows that hide the very margin whose containment matters. Deep networks that could track depots live have never been built, and the evidence from multi-center benchmarks says naive ones would collapse exactly when they are needed (AMOS: A Large-Scale Abdominal Multi-Organ Benchmark for Versatile Medical Image Segmentation).
This proposal builds the first learned, domain-robust delineation system for imageable therapeutic depots, and — the part I care most about — converts the depot's own imaging artifacts into calibrated safety flags. The claim is not merely "a better segmenter." It is that a network trained to expect shadowing, contrast leaching, and scanner change can both delineate an amorphous depot under those corruptions and tell the interventionalist, per boundary region, whether the deep margin is actually visible or is being inferred behind an acoustic shadow. Occluded or leaking margins are precisely the situations where systemic toxicity returns, so the system's uncertainty output is not a nicety: it is the safety mechanism.
Core Claim
Existing out-of-domain segmentation defenses lean on shape priors and post-hoc fixes, and both fail here. The calibration network of Improved post-hoc probability calibration for out-of-domain MRI segmentation relies on a denoising autoencoder trained on ground-truth anatomical contours and, as its authors concede, works only for regular organs like the left ventricle — a fluid gel depot has no atlas, no consistent morphology, and no fixed topology. ErrorNet: Learning error representations from limited data to improve vascular segmentation shows the way around this: learn an internal error-correction mapping by synthetically degrading outputs (broken vessels, hollow cores) on source data alone, needing no target-domain labels — but it has never touched amorphous lesions or artifact-occluded regions. Deep Poincare Map For Robust Medical Image Segmentation shows that a policy acting on local oriented patches, with boundaries as attractors, transfers zero-shot across scanners (Dice 0.74 on an unseen challenge set vs. 0.66 and 0.51 for baselines) because it never commits to a global appearance or shape — but it assumes the boundary is visible.
My claim is that these three ingredients compose into something none provides alone: a depot tracker whose robustness comes from (i) local boundary dynamics instead of a shape prior, (ii) a corrector trained on a degradation curriculum that simulates the depot's own pathologies — acoustic shadowing of the deep margin, beam-hardening streaks from iodinated contrast, and temporal appearance drift as microbubbles diffuse — and (iii) a per-boundary visibility model that distinguishes "shadowed, hence unknown" from "visible and confined." The safety output is a running risk score. Let $V_t$ be the tracked depot volume at acquisition time $t$ and $\bar{\kappa}_t$ the fraction of the depot boundary that is occluded or low-confidence at that time; the flag is
$$R_t \;=\; \alpha \, \frac{\max(0,\, V_t - V_{t-1})}{V_{t-1}} \;+\; (1-\alpha)\,\bar{\kappa}_t,$$
with $\alpha$ tuned on labeled leakage events. The logic: a confined, dissolving depot should never grow, and growth that coincides with an occluded margin — the signature of gel tracking along a fascial plane behind its own shadow, which the swine hydrogel study observed with multi-pronged needles (In vivo imaging... thermosensitive hydrogel) — is the highest-risk pattern in percutaneous delivery.
What Exists and What Is Missing
Two experimental papers sit exactly at this intersection and both stop short of learning. The echogenic ethyl cellulose–ethanol study built a classical pipeline (contrast-limited histogram equalization plus a Fast Marching front-propagation segmenter) that reached Dice scores above 0.7 in excised liver, but the authors state plainly that at ethyl cellulose concentrations of 9% and above the depot's acoustic shadow blinds the algorithm, and they explicitly name instantaneous deep-learning segmentation as unaddressed future work (Evaluating the echogenicity of ethyl cellulose-ethanol gel). The POLDOX swine study tracked dual-modality imageable hydrogels with manual 3D Slicer contouring, showed that plasma exposure drops by more than 60% versus free drug (area under the curve 852.1 ± 409.1 vs. 2283.4 ± 377.2 ng/mL·h), and observed the ultrasound appearance turn heterogeneous over four hours as microbubbles diffused — with no algorithmic answer to that drift (In vivo imaging... thermosensitive hydrogel). Neither touches domain shift. On the imaging side, the domain-shift literature supplies methods but never depots: multi-center robustness via hybrid CNN–Transformer backbones (TransNetR: Transformer-based Residual Network for Polyp Segmentation with Multi-Center Out-of-Distribution Testing), partially-supervised organ alignment (Labeled-to-Unlabeled Distribution Alignment for Partially-Supervised Multi-Organ Medical Image Segmentation), OAR contouring under scarce labels (Separated Contrastive Learning for Organ-at-Risk and Gross-Tumor-Volume Segmentation with Limited Annotation), and input-space adaptation for frozen, regulator-certified models (Adapt the Data, Not the Model: Input-Space Adaptation for Frozen Time-Series Predictors). The OAR work motivates the clinical framing — contouring organs is how radiotherapy prevents collateral injury — but it is single-modality CT and never handles the severe artifacts an injected depot creates. The gap is a system, and a benchmark to train and judge it on; neither exists.
Research Program
Three things get built. First, a depot-shift benchmark: agar phantoms and excised bovine liver following the echogenicity study's design (ethyl cellulose from 6% to 12%, depots around 4 cm², which is exactly the concentration range that transitions from visible to shadow-casting), plus an in vivo swine arm following the POLDOX protocol (liver, kidney, and muscle beds; single end-hole, multi side-hole, and multi-pronged needles; hourly CT and ultrasound over four hours; plasma sampled for high-performance liquid chromatography so that "did it leak" has a biochemical answer, not just an imaging one). Domain shift is introduced deliberately: at least three ultrasound machines or radically different presets, at least two CT reconstruction kernels, and the natural temporal axis of microbubble diffusion and contrast leaching. Expert contours on a labeled subset provide Dice ground truth; Torchio-style simulated corruptions (bias fields, ghosting, speckle), as used in the cardiac calibration work (Improved post-hoc probability calibration), augment training.
Second, the model: a local-patch boundary-dynamics segmenter in the spirit of the Poincaré-map agent, paired with an ErrorNet-style corrector whose variational degrader is retrained on depot-specific failures — shadowed deep margins, streak artifacts, fragmentation from microbubble heterogeneity — and a visibility head that predicts, per boundary element, whether it lies in a shadowed or degraded region. Third, the safety layer: the risk score above, plus a depot-to-organ margin computed against an AMOS-style multi-organ context model (we start from the public MASILab/deform-uxnet codebase), so the report reads as "depot margin 6 mm from renal capsule, deep margin occluded, volume stable" rather than a bare mask.
Baselines
The naive reference point is the clinical status quo: manual and semi-automated threshold contouring in 3D Slicer as used in the swine POLDOX study (In vivo imaging... thermosensitive hydrogel). It sets the latency and throughput bar — anything learned must be near-instantaneous during a live procedure — but its segmentation quality is operator-dependent and unmeasured on our benchmark, so the quantitative naive anchor is the strongest classical automated pipeline: the CLAHE-plus-Fast-Marching segmenter of Evaluating the echogenicity of ethyl cellulose-ethanol gel, which reported Dice scores greater than 0.7 in liver tissue in its own setting. We reimplement it faithfully, tune its free parameters once on our validation split, and run it on every acquisition, including the shadowed and late-timepoint cases it was never designed for. Beating it means holding Dice above 0.7 across scanners and timepoints, including the ≥9% ethyl cellulose regime where the original authors report it goes blind.
The strongest existing learned competitor is an ErrorNet-style error-correction U-Net, because it is the one published method designed for exactly this regime: no target-domain labels, amorphous topology, learned self-repair. In its own domain it achieved an average Dice of 71.0 when trained on one retinal dataset and tested on unseen ones, roughly 2.5 points above the next best domain adaptation method (ErrorNet). We train it on our source-domain depot data with its original degradation sampler (topological corruption only — no shadow or drift curriculum). We also run a plain 3D U-Net with aggressive standard augmentation, since AMOS shows this is the default expectation and its cross-vendor collapse — more than 20% mDice lost on unseen MRI vendors, from 89.04% mDice in-domain for a 3D U-Net (AMOS) — is the failure mode our claim predicts and must beat. All methods get the same training images, the same expert labels, the same GPU budget and tuning trials on the same validation split; only our curriculum differs. Beating ErrorNet-adapted means a statistically significant Dice and Hausdorff-95 improvement on held-out scanners and timepoints beyond hour two, plus lower expected calibration error, since correctness of confidence is half the deliverable.
The key ablation removes the degradation curriculum — shadow simulation, streak simulation, and temporal drift augmentation — while keeping the architecture intact, trained only on clean source images with standard augmentation. If the curriculum is the real ingredient, this ablation should reproduce the AMOS-style collapse pattern (large Dice loss on unseen scanners, overconfident calibration error near the 14–17% range that Improved post-hoc probability calibration reports for poorly-calibrated out-of-domain models) while the full system does not. A second ablation removes the visibility head, testing whether per-boundary occlusion modeling actually drives the leakage flag's usefulness.
Why not just…?
Why not just use the classical pipeline that already works? Because "works" was measured in its comfort zone: the Fast Marching pipeline's Dice above 0.7 was obtained on phantoms and excised liver at moderate concentrations, and its own authors report that at ≥9% ethyl cellulose the acoustic shadow makes the deep margin impossible to delineate (Evaluating the echogenicity of ethyl cellulose-ethanol gel) — and the higher concentrations are the ones that localize drug best, so the clinically desirable formulation is the one that breaks the tool. Manual 3D Slicer contouring does not break, but it is offline and too slow for live guidance, and the same study's four-hour observation of evolving heterogeneous ultrasound appearance is exactly the drift a static threshold cannot follow (In vivo imaging... thermosensitive hydrogel, exact id In vivo imaging... thermosensitive hydrogel). Our experiments settle this directly: both are baselines, run on every corrupted and late-timepoint case.
Why not just train a big 3D U-Net with heavy augmentation? This is a real threat and I will not argue it away — it might carry much of the segmentation quality. The evidence against sufficiency is that AMOS already did roughly this across eight scanners and two centers and still lost more than 20% mDice on unseen MRI vendors (AMOS), and out-of-domain deep segmenters are not merely wrong but overconfident, with expected calibration errors of 14.48% under artifact corruption and 16.78% across a bSSFP-to-LGE protocol shift even after local temperature scaling (Improved post-hoc probability calibration). An overconfident tracker that silently extends the depot boundary into a shadowed region is worse than no tracker. The plain U-Net is a baseline; if it matches our Dice and our calibration and our leakage-flag AUC, the proposal's added machinery is unjustified and I would report that.
Why not just recalibrate a frozen segmenter post hoc? Because the best published post-hoc method still leaves 12.78% expected calibration error on corrupted scans and 15.37% across protocols (Improved post-hoc probability calibration) — better, not good — and its core mechanism, a denoising autoencoder shape prior over ground-truth contours, is explicitly restricted to regular anatomies and fails on amorphous structures, which is precisely what a fluid gel depot is. Calibration cannot restore information the shadow destroyed; it can only make ignorance honest. Our visibility head is built to make ignorance localized, which is what an interventionalist needs. That said, we will run local temperature scaling on our frozen base segmenter as an additional baseline, because if global calibration plus our curriculum suffices, the visibility head's extra complexity should be cut.
Why not just use unsupervised domain adaptation on the new scanner's images? Because the deployment moment is the intervention: there is no archive of target-scanner depot images to adapt to — depots are rare, patient-specific, and the scanner may be encountered for the first time mid-procedure. ErrorNet's cross-dataset result (71.0 Dice, roughly 2.5 points above the next best adaptation method with zero target data, ErrorNet) is the evidence that source-only error correction is the right family for this constraint, and our design inherits it rather than fighting it.
Proposed Experiments and Decisive Checks
The primary evaluation is a three-axis generalization matrix over the benchmark: train on one scanner set, phantom and ex vivo only; test on held-out scanners, held-out organs, and held-out timepoints. Metrics are Dice and 95th-percentile Hausdorff distance against expert contours (the deep-margin distance is the toxicity-relevant one), expected calibration error following the protocol of Improved post-hoc probability calibration, inference speed against the 54.60 FPS reference reported for interventional-speed segmentation in TransNetR, and — the decisive new quantity — leakage-flag AUC: does the risk score separate injections with HPLC-detectable plasma exposure from confined ones? The swine studies give the anchor numbers to interpret against: confined POLDOX runs showed plasma area-under-curve of 852.1 ± 409.1 ng/mL·h versus 2283.4 ± 377.2 for free drug (In vivo imaging... thermosensitive hydrogel), and depot sphericity near 0.9 for single end-hole needles — a quantitative signature of confinement our tracker can be checked against.
Three checks decide the claim. First, the shadow regime: at ≥9% ethyl cellulose, the classical baseline should fail on the deep margin (as its authors report) and the curriculum-trained system must retain usable Dice and, critically, must flag the occluded margin rather than hallucinate it — a hallucinated confident margin is a failure even with good Dice. Second, the time axis: Dice and calibration must stay flat across the four-hour appearance drift that the POLDOX study documented, where static appearance models should decay. Third, the safety correlation: across needle designs (the multi-pronged needle is the known leak-inducer), the risk score must rank leak events above confined events with AUC meaningfully above chance, tuned on one organ and tested on the others. If the plain U-Net matches all three, the interesting residue of this proposal is the visibility model alone — and that would still be worth reporting honestly.
References
- Hyperthermia-targeted rectal delivery of thermosensitive liposomal doxorubicin via intra-arterial and intravenous administration
- Comparison of Conventional Chemotherapy, Stealth Liposomes and Temperature-Sensitive Liposomes in a Mathematical Model
- In vivo imaging and pharmacokinetics of percutaneously injected ultrasound and X-ray imageable thermosensitive hydrogel loaded with doxorubicin versus free drug in swine
- Evaluating the echogenicity of ethyl cellulose-ethanol gel for tracking biodistribution during liver ablation
- AMOS: A Large-Scale Abdominal Multi-Organ Benchmark for Versatile Medical Image Segmentation
- Improved Post-hoc Probability Calibration for Out-of-Domain MRI Segmentation
- ErrorNet: Learning Error Representations from Limited Data to Improve Vascular Segmentation
- The Deep Poincaré Map: A Novel Approach for Left Ventricle Segmentation
- TransNetR: Transformer-based Residual Network for Polyp Segmentation with Multi-Center Out-of-Distribution Testing
- Labeled-to-unlabeled distribution alignment for partially-supervised multi-organ medical image segmentation
- Separated Contrastive Learning for Organ-at-Risk and Gross-Tumor-Volume Segmentation with Limited Annotation
- Adapt the Data, Not the Model: Input-Space Adaptation for Frozen Time-Series Predictors