Stop When the Interval Is Tight: Certified Slice Inspection for 3D Lesion Measurement

Software that reads CT or MRI scans for clinicians increasingly works one 2D slice at a time. An agent built on a vision-language model looks at selected slices, writes down what it sees, and assembles an answer to questions like "how large is the liver tumor?" or "how many lesions are in the left kidney?". It works this way because 2D models are cheap, available, and promptable in plain language. The cost is volumetric context loss: whatever lies on slices the agent never opened is invisible to it. These agents also have no principled way to decide when they have looked at enough slices. They stop when the language model feels done.

We propose to fix the stopping problem with an idea from heuristic search. Search algorithms track a lower and an upper bound on the answer and halt when the gap is small. For a quantitative volumetric question, the inspected slices give a lower bound on, say, lesion volume. A cheap 3D "envelope", meaning a region that should contain the lesion, gives an upper bound on what the unopened slices could add. The agent opens the slices that shrink this gap fastest. It stops when the gap falls below a tolerance the user sets, and it reports an interval ("3.1–3.4 mL, 14 of 82 slices inspected") instead of a single unqualified number. Search research also predicts the main way this fails: everything depends on how good the initial bound is. We therefore treat envelope quality as the central experimental variable, not a detail.

Core Claim

For volumetric measurement questions about focal targets (lesion volume, lesion count, how far a lesion extends along the scan's long axis), an inspection loop driven by the bound gap will do two things. First, it will reach a fixed error tolerance with fewer expensive per-slice inspections than fixed-interval sampling or uncertainty-based slice recommendation. Second, its reported intervals will contain the true value at a measurable, high rate, provided the envelope does not clip the lesion. The second property is the new part. No agent in the file attaches a checkable certificate to its answer. Where the envelope is wrong, the method should detect its own failure and widen the envelope, rather than silently report a confident wrong number.

Positioning

The closest system is 3DMedAgent. It segments organs with 3D tools to fill a memory bank (organ size, extent, intensity) and finds candidate regions with text-image heatmaps. It then runs a "Think-with-1-Slice" loop in which a 2D multimodal model inspects slices and records textual findings. It reaches 66% on DeepTumorVQA, against 37% for RadFM and 39% for M3D. That paper motivates our work: it shows slice-wise agents beat native 3D models at clinical question answering. It also leaves the loop without any stopping criterion or error bound, and once its initial region of interest is set it never revisits it. We keep its architecture and replace the loop's control logic.

Slice and Conquer propagates a user-annotated 2D slice into 3D and recommends the next slice by lowest maximum softmax probability. It reports 0.849 Dice on MSD-Lung against 0.815 for a prior interactive method. It supplies our propagation component and one strong baseline. Its recommender, however, uses an uncertainty score with no link to a bound on the final error.

How to select slices for annotation constrains the design. It found that clever slice selection with masked-autoencoder features often fails to beat random or fixed-interval sampling on LiTS17 and breast MRI. It also found that fixed sampling fails completely on small stroke lesions in ATLAS below 5% slice density, and that mask interpolation between sampled slices can hurt. We read this as a warning that slice ranking alone may not help. We therefore make the stopping certificate the main contribution and test selection separately.

The bounding machinery comes from search. From Using Alternative Suboptimality Bounds in Heuristic Search we borrow additive tolerances ($C \le C^* + \gamma$) in place of multiplicative ones. A clinician's natural request is "within 0.5 mL", not "within 10%". The lesson from Saturated Post-hoc Optimization is that combining several valid bounds tightens estimates but cannot rescue uninformative ones. We combine several envelope sources by taking their intersection. The intersection is still a valid bound if each source is valid, and it is tighter than any single source. The anytime Marginal MAP work (Anytime Best+Depth-First Search for Bounding Marginal MAP) and Hansen and Abdoulahi document the cold-start cost of loose initial bounds. That is the failure mode we expect and measure.

Research Program

Take a question with a target quantity $Q$, such as lesion volume. The scan has slices $z = 1,\dots,Z$, and $V$ is the set of slices inspected so far. An expensive per-slice oracle returns the lesion mask on a slice. In the main setting the oracle is a simulated clinician click-through using ground truth; in the realistic setting it is a SAM-based 2D segmenter. Let $a(z)$ be the lesion area returned on slice $z$, $E(z)$ the envelope area on slice $z$, and $s$ the slice thickness. The bounds are

$$ L(V) = s\sum_{z\in V} a(z), \qquad U(V) = L(V) + s\sum_{z\notin V} E(z). $$

$U$ is valid whenever the envelope contains the lesion on every unopened slice. The loop stops at the first $V$ with $U(V) - L(V) \le \gamma$, where $\gamma$ is the user's tolerance. Lesion count and long-axis extent get analogous bounds: each unopened stretch of slices whose envelope is non-empty can contribute at most a bounded number of new lesions, or extend the measured extent by at most its own length.

There are three components. The first is the envelope. It is the intersection of the organ mask from the 3D segmenter, a dilated mask from a coarse 3D lesion model, and an intensity window in Hounsfield units. After each inspection we tighten it on neighboring slices using Slice-and-Conquer-style propagation, but only as a candidate. Whenever propagation would shrink an envelope, we re-check validity on held-out slices. The second is selection. We open next the slice with the largest current gap contribution, scored as envelope area minus expected lesion area under the propagated estimate. The third is self-repair. If the oracle finds lesion pixels touching the envelope boundary on an inspected slice, the envelope is evidently clipped. The loop then dilates it locally and recomputes $U$. This is the "feedback to refine initial bounds" that 3DMedAgent lacks.

The vision-language model is kept out of the measurement path. Your other Left! shows GPT-4o falls from 75% to 50% on relative-position questions when CT slices are flipped, which means it recites textbook anatomy instead of reading pixels. So the language model only parses the user's question into a target quantity and tolerance, then phrases the certified interval as an answer. The accessibility goal from CellAgent is preserved: a plain-language request yields an answer a non-programmer can audit. The difference is that the answer's reliability comes from the bound, not from a model judging its own plots.

Baselines

The naive reference is the coarse 3D model alone, with zero slices inspected. On small, varied targets it is weak: The Medical Segmentation Decathlon reports 0.21 Dice for pancreas tumors and 0.16 for colon cancer, against 0.94 for liver and spleen. It sets the error our method must reduce. For it to count as beaten at a given budget, our method must cut absolute volume error at that budget by a margin confirmed with a paired test across cases.

The strongest selection baseline is fixed-interval sampling with mask interpolation, the policy How to select slices found hard to beat. It runs on the same scans, with the same oracle and the same per-case slice budget as our method. No result is reported in the file on LiTS17, DBC, or ATLAS for this policy; the paper gives only qualitative findings. We will regenerate it ourselves.

The strongest existing interactive method is Slice and Conquer's lowest-confidence recommender, run with its released propagation model and the same oracle budget. Its reference figure is 0.849 Dice on MSD-Lung.

The strongest agentic baseline is 3DMedAgent's slice loop with its own language-model stopping rule, run on the quantitative subset of DeepTumorVQA. Its reference figure is 66% average accuracy on DeepTumorVQA overall. We report accuracy on our subset, together with the number of slices it opens.

Two ablations isolate the key ingredient. In the first, fixed-interval selection is paired with our bound-gap stopping rule. In the second, our selection is paired with a fixed budget and no stopping rule. If the first ablation matches the full method, the certificate is the contribution and the selection rule is not, and we will say so.

All methods share the coarse model, oracle, and compute. Every method gets the same budget for tuning hyperparameters on the training folds.

Why not just…?

Why not just run a full 3D segmenter and measure from its mask? That is our zero-slice baseline. On large organs it may well be enough, and we expect no gain there. On small tumors, the Decathlon's 0.16–0.21 Dice means volume and count estimates from the mask alone are unreliable. More importantly, a 3D mask comes with no statement of how wrong it might be. The decisive comparison is error at matched clinician effort and interval coverage, both reported per target size.

Why not just sample every k-th slice? This is a real threat. On LiTS17 and breast MRI, fixed sampling matched cleverer schemes. Our first ablation (fixed-interval selection with our stopping rule) is designed to show whether the stopping rule, rather than the ranking, carries the gain. The prediction that would distinguish them comes from the ATLAS failure: on focal lesions, fixed spacing misses whole lesions. Bound-gap selection should not, because it targets the unopened slices where the envelope still has area.

Why not just let the vision-language agent decide when to stop? That is 3DMedAgent's design. The flipped-image collapse in Your other Left! suggests its confidence is not tied to the pixels it has seen. We measure directly how often its stopping point leaves the true value outside any honest interval.

Proposed Experiments

Data come from MSD liver-tumor, pancreas-tumor, and colon tasks, from KiTS19, LiTS17, and ATLAS, and from the quantitative subset of DeepTumorVQA. We run a ground-truth oracle, to test the bound logic cleanly, and a SAM-based oracle, to test realistic use. Metrics are slices inspected to reach tolerance $\gamma$, absolute error of the interval midpoint, interval coverage (the fraction of cases where the truth lies inside), and wall-clock time. All are reported in bins of lesion size and slice thickness, because thick-slice scans inflate $s$ and loosen every bound.

The decisive check for the cold-start thesis is a controlled sweep of envelope quality. We erode or dilate the coarse mask and inject clipping at known rates. The expected outcome is three curves: slices needed rising as the envelope loosens, coverage collapsing as it clips, and self-repair restoring coverage at modest extra cost. If coverage stays high only under ideal envelopes and repair fails, the certificate is not usable in practice and the claim is falsified.

A second check measures coverage under the imperfect SAM oracle. Oracle errors break the lower bound's validity, so we need to know how much coverage we lose and whether a calibrated error margin on each slice's area restores it.

A third check is the ablation pair described under Baselines. It decides whether this work is a stopping rule, a selection rule, or both.