Slice Spacing as Imaginary Time: Resolution-Consistent 3D Segmentation from 2D Models

Most strong image models are 2D. When they are applied to CT or MRI volumes one slice at a time, they lose what lies above and below each slice: a blood vessel becomes a sequence of unrelated circles, and an organ mask flickers from slice to slice. The usual fixes are a full 3D network, which is expensive and must be trained for each acquisition protocol, or a smoothing layer that links neighbouring slices with a hand-tuned strength. This proposal borrows an exact correspondence from physics. A 2D quantum system that evolves in imaginary time is mathematically the same as a 3D classical system, where the extra axis is time cut into small steps. We read the scan's depth axis as that imaginary time. The coupling between neighbouring slices is then not a free knob. It is a fixed function of the physical slice spacing and of one "persistence length" per anatomical class, measured in millimetres.

The concrete claim is that this function, which comes from the Suzuki-Trotter construction, gives a slice-coupling model that keeps the same meaning when slice spacing changes. A model calibrated once on thin-slice CT should transfer without retuning to thick-slice CT and to anisotropic cardiac MRI. Constant couplings fail at that. So does the natural engineering guess of scaling the coupling inversely with spacing. The physics mapping supplies the parameterisation and a cheap falsification test. We do not use quantum hardware, and we do not claim a quantum speedup.

Core Claim

Each slice $z$ carries a label field. An off-the-shelf 2D segmenter provides per-pixel class evidence, and the slices are coupled in-plane by an ordinary Potts smoothness term. Treat the depth axis as continuous imaginary time. Let the labels at each in-plane location jump between the $q$ classes at rate $\Gamma_c$ for class $c$. This is the q-state analogue of a transverse field: it lets a pixel's label change as you move through depth. Trotterising this 2D quantum Potts model with step $\varepsilon=\Delta z$, the physical slice spacing in millimetres, yields a 3D classical Potts model. Its inter-slice coupling between equal labels is

$$K_c(\Delta z) = \ln\!\left(1 + \frac{q}{e^{\Gamma_c \Delta z}-1}\right).$$

Here $q$ is the number of classes and $1/\Gamma_c$ is the expected depth extent, in millimetres, over which class $c$ persists. The formula follows from exponentiating the uniform-jump generator over one step: the ratio of the stay probability to any single jump probability is $1+q/(e^{\Gamma_c\Delta z}-1)$.

The formula behaves very differently at the two ends of the spacing range. For thin slices, $K_c$ grows only logarithmically, as $\ln\big(q/(\Gamma_c\Delta z)\big)$. For thick slices, it decays exponentially, as $q\,e^{-\Gamma_c\Delta z}$. The common heuristic $K\propto 1/\Delta z$ matches neither regime. The parameters $\Gamma_c$ are physical, so they are fit once and reused at any spacing. The correlation length along depth is set by the gap of the 2D quantum model, not by the grid.

Relation to Prior Work

The mapping itself is standard physics. Phase transitions in models for coupled charge-density waves maps a zero-temperature quantum system to a classical model in one extra dimension. Disordered flat phase and phase diagram for restricted solid-on-solid models of fcc (110) surfaces runs the reverse direction: it reads one spatial axis of a 2D classical surface as imaginary time for a 1D quantum chain. That reverse direction is the one we reuse, with a CT depth axis in place of a crystal-surface axis. Both papers supply the method; neither touches images.

Three works sit closest to our setting, and each leaves the depth axis untouched. Q-Seg casts 2D segmentation as a binary optimisation problem on a D-Wave annealer, tiled into 32×32 patches, with no extra Trotter dimension. Quantum Annealing for Clustering couples Trotter replicas so that they jointly escape poor optima. Its replicas are abstract copies indexed by an algorithmic index, never physical slices, and its data are non-spatial vectors. A New Active Contour Method Based on Elastic Interaction starts from 3D elasticity but collapses it onto 2D slices and leaves volumetric extension open.

Our difference from all three is that the Trotter step is identified with a measured physical quantity, $\Delta z$. That identification is what produces a testable prediction about resolution transfer.

Two physics results constrain the design. Simple Glass Models and Their Quantum Annealing shows the spectral gap closes as $2^{-N/2}$ at first-order transitions, which makes annealing exponentially slow. We therefore use the mapping only to define the model. Inference is deterministic mean-field with exact forward–backward passes along depth, in the spirit of the quasi-static treatment in Quantum effects in rotationally invariant spin glass models. Self-organized error correction in random unitary circuits with measurement shows that mean-field estimates of critical exponents can be off ($\nu=1$ versus 1.1–1.3 numerically). This warns us not to trust mean-field near the regime where label fields would sit at a critical point. In practice, strong evidence terms from the 2D segmenter keep the model far from that regime.

On the vision side, the motivation comes from three findings. A Novel Domain Adaptation Framework for Medical Image Segmentation argues that slice-wise networks lack the volumetric context needed to resolve ambiguity. Your other Left! finds slice-level models falling back on memorised priors instead of spatial evidence. Tri-Plane Mamba recovers pseudo-3D context through heuristic orthogonal-plane scans with no account of slice spacing.

Research Program

The work has three stages, ordered so that the cheapest possible falsification comes first.

The first stage tests the physics premise directly on ground-truth labels, with no model involved. We take BTCV and MSD Spleen label volumes, subsample or average-pool them along depth by factors 1, 2, 4 and 8, and count label persistence and transitions between consecutive slices per class. From these counts we estimate an empirical coupling $\hat K_c(\Delta z)$. We then ask which family fits this curve with a single $\Gamma_c$ per class: the Trotter form, a $1/\Delta z$ law, or a constant. If the Trotter form does not fit, the project stops or reframes, because the core claim would then be false.

The second stage builds the inference layer. The evidence terms come from a 2D TransUNet trained on slices using the Beckschen/TransUNet code; frozen SAM features are an optional second source. The in-plane Potts weight and the per-class $\Gamma_c$ are fit on the training volumes at native spacing. Inference alternates two steps. In-plane mean-field updates within each slice are followed by an exact forward–backward pass along depth at each pixel, using the transition matrix $e^{\Delta z\, L}$. Memory stays at the cost of two slices plus the stored messages. The ingredients are therefore an existing 2D model, a few dozen scalar parameters, and the Trotter coupling.

The third stage stress-tests transfer. We fit at one spacing and evaluate at others without refitting. Evaluation covers BTCV with simulated thick slices, the natively thick and anisotropic cine-MRI of ACDC, and AMOS, whose merged prostate/uterus label probes whether a single $\Gamma_c$ can serve a class with mixed shapes.

Baselines

The naive reference is slice-wise argmax of the same 2D TransUNet, with no depth coupling. Every proposed gain is measured against it using identical evidence terms. For context, the file reports this 2D architecture at 77.48% average Dice on Synapse and 89.71% on ACDC, per TransUNet.

The key ablation keeps everything fixed except how the inter-slice coupling depends on spacing. One variant uses a 3D Potts model with a constant per-class coupling tuned at native spacing. A second uses a coupling scaled as $1/\Delta z$. A third is an oracle constant coupling re-tuned separately at every test spacing. All variants share the evidence terms, the inference code, and the number of tuned parameters, and each gets the same search budget on the same training volumes. The Trotter model beats this ablation if, at untuned spacings, it lands within the noise of the per-spacing oracle while the constant and $1/\Delta z$ variants do not. The noise level is set by the across-fold standard deviation.

The strongest existing method is 3D TransUNet, which the file reports at 88.11% Dice on BTCV, surpassing nnU-Net (Beckschen/TransUNet). UNETR is a second reference; it reports a 1.238% average-Dice margin on the BTCV leaderboard (UNETR). We will retrain 3D TransUNet on the same 24/6 split. Each model is tested both at its training spacing and at shifted spacings, under two input-handling protocols: the standard isotropic-resampling pipeline, and no resampling. We do not expect to beat it at matched native spacing with full data. Beating it means a smaller drop in Dice and 95th-percentile Hausdorff distance (HD95) at shifted spacings, or higher Dice in the 3-volume regime.

TP-Mamba is the strongest adapter-style competitor. Its paper reports Dice 12% higher than conventional 3D networks with 3 training volumes and peak performance in 200–400 epochs (Tri-Plane Mamba), but gives no absolute BTCV Dice in the file. We will rerun it under our split and spacings rather than quote a number.

Why not just…?

Why not just resample every volume to isotropic spacing and train a 3D network? This is the most serious threat. Resampling plus a 3D model is the standard practice behind the 88.11% BTCV figure. Interpolation, however, cannot recreate anatomy that was never sampled between thick slices. It also commits the network to whatever spacing distribution it was trained on. The thick-slice BTCV and ACDC experiments settle the comparison directly. If resampled 3D TransUNet degrades no more than our layer as spacing grows, the proposal's practical case collapses to the low-data regime, and we will report that.

Why not just a 3D CRF with a learned inter-slice weight, or a weight scaled by $1/\Delta z$? At a single fixed spacing, our model is a 3D CRF. The physics contributes only the dependence on spacing, so the ablation above is exactly this alternative. If $1/\Delta z$ scaling matches the Trotter form in stage one, the quantum mapping adds nothing beyond intuition and we will say so. The two forms differ sharply for thick slices (exponential versus inverse-linear decay), which is precisely the clinically common regime.

Why not just learn a recurrent or Mamba module along depth? A learned depth module can fit any spacing it has seen, but it has no reason to extrapolate to spacings it has not. It also needs far more parameters and data than a few per-class persistence lengths. We will train a depth-wise recurrent head on the same 2D features, using the same data and epochs, and compare transfer to unseen spacings and accuracy with 3 training volumes. This is the setting where a handful of physical parameters should matter most.

Proposed Experiments

The decisive check is a fit-once, test-everywhere protocol on BTCV. We fit $\Gamma_c$ at native spacing, then evaluate Dice, HD95, and a depth-coherence score at 2×, 4× and 8× slice thickness. The coherence score is the number of 3D connected components per organ relative to ground truth, which directly measures the "disjoint circles" failure. Three outcomes would support the claim. First, the stage-one label statistics follow the Trotter curve with one $\Gamma_c$ per class. Second, the Trotter-coupled layer stays within fold noise of the per-spacing oracle while the constant and $1/\Delta z$ couplings fall outside it. Third, on natively thick ACDC MRI, $\Gamma_c$ values fit on CT-derived thin slices for analogous structures give a non-trivial gain over slice-wise TransUNet.

Secondary checks probe the limits of the model. On MSD Spleen, the layer should help most on fragmented masks. On AMOS, we test whether the merged prostate/uterus class needs two persistence lengths, which would show where a single uniform-jump rate is too crude. In a sweep of evidence strength, we test whether mean-field accuracy degrades as the evidence weakens toward the critical regime flagged in Self-organized error correction in random unitary circuits with measurement. A negative result in stage one is itself publishable: it would show that the imaginary-time picture, however elegant, does not describe how anatomy persists along the scan axis.