Neural networks for thin curved structures such as retinal vessels, coronary arteries and nerve fibres work best when the image is first "lifted" into a space that stores, at every pixel, a response for every orientation. In that space two vessels that cross in the picture occupy different orientation layers, so a filter can smooth each one without smearing it into the other. The most parameter-efficient networks of this kind replace ordinary convolutions with learned partial differential equations (PDEs) that push information along the local orientation. The published versions all push it along straight lines, or more precisely along the group's straight "horizontal" directions. A straight-line smoother flattens tight bends, and it treats the end of a sharp curve the same way it treats noise.
We propose a trainable, rotation- and translation-equivariant diffusion layer whose direction of flow bends. At each point of the lifted space the layer estimates how fast the local orientation is turning, which is the curvature, and it diffuses along the matching circular arc rather than along the tangent line. A learned confidence gate returns the layer to the existing straight-line behaviour wherever the curvature estimate is unreliable. The claim is narrow and testable. Straight-gauge PDE layers carry a measurable straightening bias that concentrates on high-curvature segments. Curvature steering removes that bias at almost no parameter cost. The gain will be invisible in whole-image Dice scores and visible once errors are grouped by curvature.
What exists and what is missing
Three bodies of work supply the pieces. Franken and Duits, in Crossing-Preserving Coherence-Enhancing Diffusion on Invertible Orientation Scores, estimated local curvature $\kappa$ from the Hessian of the orientation score, measured in the moving frame. They used that estimate to steer handcrafted diffusion along circles, and we reuse their curvature estimator. Smets et al., in Total Variation and Mean Curvature PDEs on the Homogeneous Space of Positions and Orientations, built locally adaptive frames from a singular value decomposition of that Hessian. They reported 26.1 dB PSNR against 23.9 dB for BM3D on "Mona Lisa". Their flows are not learnable, need hand-chosen stopping times, and are slow. Their result motivates the proposal, because nonlinear geometry-adaptive flows help strongly exactly where Smets et al. in PDE-based Group Equivariant Convolutional Neural Networks found that linear diffusion layers added little.
The closest work is Diffusion-Shock PDEs for Deep Learning on Position-Orientation Space by Sherry et al. It puts data-driven frames inside a trainable network and shows that the right second-order operator for such frames is a generalized Laplacian built from a Lie-Cartan connection, not the standard Laplace-Beltrami operator. It explicitly does not steer on curvature. Its frames align with the local direction but still generate straight exponential curves. We take their connection-based Laplacian as the discretisation, and we change only which curve the frame follows. Analysis of (sub-)Riemannian PDE-G-CNNs by Bellaard et al. constrains the design. At high spatial-to-angular anisotropy, Riemannian approximations fail, so the curved frame must be built so that the sub-Riemannian limit, with no sideways motion, stays well-defined.
Core mechanism
Write $\mathcal{A}_1, \mathcal{A}_2, \mathcal{A}_3$ for the left-invariant frame on the lifted space. These are the derivative along the orientation $\theta$, the derivative perpendicular to it, and the derivative in $\theta$. Let $W$ be a feature channel. The layer first computes curvature from left-invariant derivatives of a Gaussian-smoothed copy of $W$, following Franken and Duits. The Gaussian scale $s$ is learned per layer. The layer then defines a bent tangent and runs a few explicit diffusion steps:
$$\mathcal{B}_1 = \frac{\mathcal{A}_1 + \beta\,c\,\kappa\,\mathcal{A}_3}{\sqrt{1+\beta^2 c^2\kappa^2}}, \qquad \partial_t W = D_{11}\,\mathcal{B}_1^2 W + D_{22}\,\mathcal{A}_2^2 W + D_{33}\,\mathcal{A}_3^2 W .$$
Here $\beta$ is the spatial-angular scale that makes $\kappa$ dimensionless. The confidence $c \in [0,1]$ is produced from the Hessian's orientation confidence by a small learned gate, and the $D_{ii}$ are learned diffusion weights. The curves generated by $\mathcal{B}_1$ are circular arcs in the image plane with curvature $c\kappa$. Setting $c=0$ recovers a straight-gauge layer exactly. The proposal is therefore a strict generalisation of the straight-gauge layer, and the ablation is clean.
Every quantity is built from left-invariant derivatives, so the layer commutes with roto-translations. Second-order terms are discretised with the Lie-Cartan construction of Sherry et al. The layer is placed between the convection and morphological (dilation/erosion) blocks of an existing PDE-G-CNN.
Two engineering questions determine whether this works. The first is noise in second derivatives. The learned scale $s$ and the gate $c$ are the answer we propose, and we will check whether training drives $c$ toward zero on noisy data. The second is stiffness. We will use at most four explicit steps under a stability bound on the step size, and we will report the wall-clock overhead against a straight-gauge layer.
Research program
The work has three stages. First, build the layer in an existing PDE-G-CNN codebase on $\mathbb{M}_2$ with 8 and 16 orientations. Second, construct a curvature-controlled synthetic benchmark. It extends the $64\times64$ "Lines" completion task used by Bellaard et al. and Sherry et al. with arcs and spirals of prescribed curvature, crossings at controlled angles, gaps, and correlated noise, so that ground-truth curvature is known everywhere. Third, evaluate on real vessel data, where curvature of the ground-truth centreline is computed from skeletons and used to stratify errors.
We will also analyse what the gate learns, meaning where $c$ is large, as a check that the network uses curvature where geometry predicts it should. As an exploratory downstream use, we will feed the network's output as the cost for the asymmetric Finsler tracker of Geodesic Tracking via New Data-Driven Connections of Cartan Type for Vascular Tree Tracking. Both that paper and its 2025 follow-up list learned costs as future work.
Baselines
The naive reference point is a standard 2D CNN trained with rotation augmentation. On DRIVE, PDE-based Group Equivariant Convolutional Neural Networks reports a 12-layer CNN at Dice 0.8189 with 129,432 parameters. We rerun it on our splits with the same epochs and augmentation budget.
The strongest existing learned method on the segmentation question is the 12-layer PDE-G-CNN from the same paper, reported at Dice 0.8220 with 3,678 parameters on DRIVE. It is the strongest competitor because it already has the equivariance and parameter efficiency our layer inherits. The only thing it lacks is adaptive geometry. We match its depth, orientation count and training schedule, and allow a parameter budget within 10% of its own.
The most direct competitor is the gauge-frame diffusion-shock PDE-G-CNN of Sherry et al. For it, no reported segmentation number was found on Lacuna, and its denoising PSNR on DRIVE and "Spiral" is described as beating BM3D and NLM without tabulated values in the file. We will therefore retrain it ourselves under identical data, tuning trials and GPU hours.
On DCA1, Bellaard et al. report only parity with a CNN at 35 times fewer parameters. No reported Dice value was found on Lacuna for DCA1, so we rerun their $\rho_{com}$ network.
For denoising, the reference numbers are Gauge TVF at 26.1 dB and BM3D at 23.9 dB on "Mona Lisa" (Smets et al. 2021). We also include the handcrafted orientation-score Gaussian-derivative filter of Robust and Fast Vessel Segmentation via Gaussian Derivatives in Orientation Scores, which reports DRIVE accuracy 0.9446 and STARE accuracy 0.9511.
The key ablation replaces $c\kappa$ with zero, which gives the straight-gauge twin with identical parameters otherwise. We also test a fixed, non-learned smoothing scale and an oracle-curvature variant on synthetic data, to separate "curvature helps" from "our curvature estimate is good enough."
Beating the baselines means the following. On the top curvature tercile, error must drop in a way that holds across five seeds (paired test across images). Whole-image Dice must be no worse than the straight-gauge twin and the PDE-G-CNN. Runtime overhead must stay under a factor of two.
Why not just…?
Why not just stack more straight-gauge PDE layers, or add more orientations? Composing a convection step with a diffusion step can approximate a bent kernel. This is a real threat. The experiment that settles it is a depth- and orientation-matched sweep of the straight-gauge twin, extended up to our compute budget. If a deeper straight network closes the high-curvature gap, curvature steering is not worth its cost, and we will report that.
Why not just use the diffusion-shock gauge-frame network as is? Sherry et al. already adapt the frame direction to the data, and the shock term sharpens edges, which could hide much of the straightening bias. Our layer differs only in the bent tangent, and the $c=0$ ablation isolates exactly that difference. We expect the gap to appear on tight arcs and spirals and not on straight vessels. If it does not appear there, the idea fails.
Why not just preprocess with Franken–Duits curvature-steered diffusion and feed an existing network? Handcrafted steering does straighten less, but its parameters and stopping time are fixed by hand, and its errors propagate into the network uncorrected. We run this pipeline as a baseline under the same training budget. If it matches the learned layer, the contribution reduces to engineering, and we will say so.
Why not just a big 2D U-Net with heavy augmentation? The file's evidence is about parameter efficiency, not raw accuracy: a PDE-G-CNN beats a 35-fold larger CNN on DRIVE. A large enough CNN may match us on aggregate Dice. Our claim is about curvature-stratified errors and crossing handling at a small parameter budget, so we report accuracy against parameter count.
Decisive checks
The first test is the synthetic curvature sweep. Plot completion and denoising error against ground-truth curvature for the straight-gauge twin and for our layer. The straight twin's error should rise with $\kappa$, and ours should stay flat up to the curvature the orientation sampling can resolve, roughly one orientation step per spatial scale. A flat curve for both would falsify the premise.
The second test uses correlated noise. At high noise, the learned gate should turn curvature off, and performance should fall back to the straight twin rather than below it. Falling below it would mean second-derivative noise wins.
The third test is on DRIVE and DCA1. The curvature-stratified Dice gain should concentrate in high-curvature vessel segments while aggregate Dice stays at or above 0.8220.
The final, exploratory test uses STAR. Plugging the learned output into the asymmetric Finsler tracker should reduce crossing mistakes below the 3 reported with optical enhancement plus TV-flow in Geodesic Tracking of Retinal Vascular Trees with Optical and TV-Flow Enhancement in SE(2). Crossings immediately followed by a bifurcation will be reported separately as the hardest case.
References
- Crossing-Preserving Coherence-Enhancing Diffusion on Invertible Orientation Scores
- Total Variation and Mean Curvature PDEs on the Homogeneous Space of Positions and Orientations
- PDE-based Group Equivariant Convolutional Neural Networks
- Diffusion-Shock PDEs for Deep Learning on Position-Orientation Space
- Analysis of (sub-)Riemannian PDE-G-CNNs
- Geodesic Tracking via New Data-Driven Connections of Cartan Type for Vascular Tree Tracking
- Robust and Fast Vessel Segmentation via Gaussian Derivatives in Orientation Scores
- Geodesic Tracking of Retinal Vascular Trees with Optical and TV-Flow Enhancement in SE(2)