Most CT and MRI scans are stored as thick slices. Detail inside each slice is fine, but the gap between slices can be five times larger. Networks that fill in the missing slices learn a black-box mapping from neighbouring slices to the gap, and they are usually trained for one fixed gap size. This note proposes a different view. Read the scan axis $z$ as if it were time, and discover the small differential equation that carries one slice into the next. Most of that equation should be plain motion of tissue within the plane: vessels drift, organ boundaries slide. A sparse, penalised remainder should account for structures that appear, vanish or branch. Once the update is written as an equation in physical millimetres, filling a gap of any size means integrating that equation over the gap. Its learned terms also say how much of the through-plane change is motion and how much is genuinely new content.
The opportunity lies where two literatures meet without touching. Equation-discovery methods learn sparse differential operators from snapshots, but only for evolution in time on small synthetic grids. Medical slice synthesis treats through-plane continuity empirically. The claim is that a sparse transport law along $z$, discovered with spacing-aware differential filters, can match black-box slice synthesis at the gap size it was trained on. It should clearly beat that synthesis at gap sizes and scanner spacings it never saw, and it should yield an interpretable split of through-plane change into motion and new content.
Core Claim and Positioning
The method borrows directly from PDE-Net. PDE-Net learns convolution filters that are constrained to act as finite-difference derivatives. It stacks Euler-like update steps and shows that learnable constrained filters stay stable where frozen finite-difference stencils diverge. PDE-Net 2.0 adds a sparse symbolic module that recovers terms such as $u\,u_x$, and pseudo-upwind schemes that stabilise advection. Both assume time evolution on dimensionless $32\times32$ or $50\times50$ grids. We change three things: the evolution variable becomes $z$ in millimetres; the filters are scaled by the true in-plane voxel spacing; and the coefficients become spatial fields predicted per scan rather than global constants.
The PDE-Net choice is deliberate given the opposing position of Mechanistic PDE Networks. That paper backpropagates through a global differentiable multigrid solver for noise robustness, but it is limited to small 2D grids by memory. Our gaps are short, and both of their end slices are observed. Cheap explicit stepping within the plane is therefore enough, and it scales to $512\times512$ slices.
The closest neighbours in the volumetric work are the methods we compete against, not the source of the method. SAINT feeds physical slice spacing to a network that generates filters, and it handles arbitrary upsampling factors by fusing sagittal and coronal views. I³Net works in the axial plane and gets better accuracy and speed, but is tied to a fixed ratio. ConResNet uses unconstrained differences between adjacent slice features for segmentation. Inter-slice consistent noise interpolation for diffusion models (ISCS) uses a fixed geodesic between two anchor slices and names data-adaptive correlation as open. None of these methods writes the slice-to-slice change as an equation with separable, interpretable terms. We do not claim a universal law of anatomy. The coefficient fields are predicted per scan. What is shared across patients and scanners is the form of the equation and the discretisation, and that is what should transfer.
The update we discover has this form:
$$\frac{\partial I}{\partial z} = -\mathbf{v}(x,y,z)\cdot\nabla_{xy} I + \sum_{k} \xi_k(x,y,z)\,\theta_k(I,\nabla_{xy}I,\nabla^2_{xy}I) + s(x,y,z).$$
Here $I$ is image intensity, $z$ is the scan axis in millimetres, and $\nabla_{xy}$ and $\nabla^2_{xy}$ are in-plane gradient and Laplacian operators. They are implemented as moment-constrained filters divided by the physical spacing $\Delta x, \Delta y$. The vector field $\mathbf{v}$ is the in-plane drift of tissue per millimetre of $z$. The $\theta_k$ are a small library of candidate terms, for example in-plane diffusion and intensity-dependent contrast change, each with a coefficient field $\xi_k$. The term $s$ is an unconstrained source that absorbs everything else, such as a vessel bifurcating or an organ entering the field of view.
Training minimises the error at held-out intermediate slices plus sparsity and smoothness penalties:
$$\mathcal{L} = \sum_{z^\ast}\|\hat I(z^\ast)-I(z^\ast)\|_1 + \lambda_\xi\sum_k\|\xi_k\|_1 + \lambda_s\|s\|_1 + \lambda_v\|\nabla \mathbf{v}\|_2^2.$$
Here $z^\ast$ ranges over the thin slices removed when a thick-slice input is simulated, and $\hat I(z^\ast)$ is the prediction at that slice. The coefficients $\lambda_\xi$, $\lambda_s$ and $\lambda_v$ weight the sparsity of the library coefficients, the cost of the source term and the smoothness of the motion field. The penalty on $s$ is the mechanism that forces explanation into the structured terms. Its size after training is itself a scientific readout.
Research Program
A small 2D encoder takes the two observed slices that bound a gap, plus the physical spacings, and predicts the fields $\mathbf{v}$, $\xi_k$ and $s$ at a few control depths. The fields are linearly interpolated in $z$ between those depths. We integrate forward from the lower slice and backward from the upper slice in sub-steps of at most the in-plane spacing, using pseudo-upwind advection. Sub-steps of that size keep the step-size stability condition for advection satisfied whenever the drift is below one pixel per millimetre. The two directions are blended with weights that depend on distance.
Because each step is a 2D operation on one slice, memory grows with the number of sub-steps rather than with a 3D receptive field. Gradient checkpointing keeps a gap of 10 mm at about 1 mm sub-steps tractable at full $512\times512$ resolution. This sidesteps the dimensionality bottleneck that MechNN-PDE reports.
Training data come from thin-slice volumes in MSD, the Medical Segmentation Decathlon CT collection (liver, colon and hepatic vessel tasks). We follow the SAINT and I³Net protocol of dropping slices to simulate ×2 and ×4 thick-slice inputs, randomising both the ratio and the absolute spacing. The analytical output is a per-anatomy decomposition: the fraction of the integrated $\partial I/\partial z$ explained by motion, by each library term and by the source. We will test whether source mass concentrates where expected, namely at annotated hepatic vessel branch points and at the top and bottom edges of organs in the MSD masks. A source that is diffuse everywhere would mean the structured form is not what explains the data.
Baselines
The naive reference point is linear and cubic interpolation along $z$. SAINT reports bicubic interpolation at 28.36 dB PSNR on ×4 MSD Liver (SAINT). Any learned method must clear this by a wide margin, and we report it on every split.
The strongest arbitrary-ratio competitor is SAINT. It reports 34.91 dB PSNR on ×4 MSD Liver, against 34.12 dB for a 3D residual dense network. The strongest fixed-ratio competitor is I³Net, which reports 43.90 dB PSNR on MSD at ×2, 1.14 dB above SAINT (I³Net). Both will be retrained from their released descriptions on our exact splits and with matched GPU-hour budgets. Each gets the same hyperparameter search budget as our method, and we will not compare against our numbers next to their reported ones from different splits. For I³Net at untrained ratios we use the fairest available extension: train it with ratios randomised the same way ours are, and also cascade its ×2 model.
The classical structured competitor is deformable-registration interpolation. It estimates a dense displacement between the two bounding slices and warps them partway along it. This is effectively our equation with only the motion term and a field constant in $z$. We use a standard registration toolkit and tune its regularisation on validation data. No reported result was found on Lacuna for registration-based slice interpolation on MSD.
Three ablations isolate the key ingredients. The first is a black-box neural ODE in $z$: the same integrator and training, but with the right-hand side replaced by an unconstrained CNN of matched parameter count. This isolates whether the sparse, interpretable form matters at all. The second removes the source term, setting $s=0$, and the third removes the library while keeping motion and source. Together these test whether the decomposition carries weight. The fourth uses frozen finite-difference stencils instead of moment-constrained learned ones, which tests whether PDE-Net's stability finding carries over to $z$-evolution.
Beating the strongest baselines means two things. At the trained ratio, we must be within 0.5 dB of retrained I³Net. At untrained ratios (×3, ×5, and ×4 when trained only on ×2), we must exceed both retrained SAINT and ratio-randomised I³Net by at least 1 dB PSNR. That gap must also show up in a through-plane metric, not only per-slice PSNR.
Why not just…?
Why not just use deformable registration between adjacent slices and interpolate along the flow? This is the most serious threat, because the hypothesis itself predicts that most through-plane change is motion. Registration has two weaknesses here. It assumes a single displacement across the whole gap, and it has no mechanism for content that appears or disappears, so it smears new vessels and organ tips. The no-library, no-source ablation of our model is close to learned registration. The comparison against classical registration therefore measures exactly the value of the extra terms. If registration matches us everywhere except at source-heavy locations, the honest conclusion is a narrower one: motion plus source suffices. We would report that.
Why not just train I³Net or SAINT on randomised ratios and let the network learn scale? Ratio-randomised I³Net is run as a baseline for exactly this reason. Our expectation that it degrades out of range rests on one point. A network that maps slices to gaps has no built-in notion that a 6 mm gap is the composition of two 3 mm gaps, whereas integrating an equation enforces that composition by construction. The test that settles this trains every model on ratios from ×2 to ×3 and evaluates at ×4 and ×5. It also evaluates on KiTS19 kidney CT, whose spacing differs from training. No reported number was found on Lacuna for KiTS19 slice synthesis, so all KiTS19 results will be our own runs.
Why not just use a black-box neural ODE in $z$, which also integrates to any depth? It shares the composition property, so it may capture most of the generalisation gain. That is precisely why it is the central ablation. If it matches our accuracy, the contribution reduces to "integrate along $z$" plus the interpretable decomposition, and we would say so. We expect the sparse form to help mainly under spacing shift, because filters scaled by physical spacing transfer across scanners while unconstrained filters implicitly memorise pixel scale.
Why not just smooth along $z$ with total variation or a 3D Gaussian? Smoothing does help in some settings: Roth et al. report that 3D Gaussian smoothing raised pancreas Dice from 56% to 68% (Roth et al.). But ISCS reports that a total-variation penalty along $z$ erases small lesions. Smoothing is included only as a sanity point, evaluated on small-structure metrics.
Decisive Experiments
The first experiment tests generalisation to unseen gaps. All methods are trained on MSD Liver at ×2 only and then evaluated at ×3, ×4 and ×5. At ×4 they are scored against the full-protocol SAINT reference of 34.91 dB, using our retrained numbers. Metrics are PSNR and SSIM on held-out slices, plus gradient error measured in sagittal and coronal reformats. Those reformats are the views where staircase artefacts appear.
The second experiment tests generalisation to unseen spacing. Models trained on MSD are evaluated on KiTS19 without fine-tuning.
The third experiment tests small structures. The MSD hepatic vessel task is run through a fixed segmentation network on volumes interpolated by each method. Vessel Dice and centreline continuity are compared against the result on the original thin-slice volume.
The fourth experiment tests interpretability. We will check whether source-term mass is significantly enriched at vessel branch points and organ edges relative to random locations. We will also check whether the fraction explained by motion is stable across scanners.
The proposal fails if the black-box neural ODE or registration matches it on all four experiments. It also fails if the source term carries most of the signal everywhere. Either result would itself be a clean negative finding about whether the through-plane structure of anatomy is low-dimensional.
References
- PDE-Net: Learning PDEs from Data
- PDE-Net 2.0: Learning PDEs from Data with A Numeric-Symbolic Hybrid Deep Network
- Mechanistic PDE Networks for Discovery of Governing Equations
- SAINT: Spatially Aware Interpolation NeTwork for Medical Slice Synthesis
- I³Net: Inter-Intra-Slice Interpolation Network for Medical Slice Synthesis
- Inter-Slice Context Residual Learning for 3D Medical Image Segmentation
- Improving 2D Diffusion Models for 3D Medical Imaging with Inter‑Slice Consistent Stochasticity
- Deep convolutional networks for pancreas segmentation in CT imaging