Gaussian Primitive Optimization in Retinal Registration

• Gaussian Primitive Optimization (GPO) is a deformable retinal registration method that utilizes Gaussian-modeled keypoints to propagate local deformation in sparse feature areas.
• It employs a four-stage pipeline combining coarse alignment, descriptor-based control nodes, Gaussian parameterization, and KNN interpolation to construct a globally coherent displacement field.
• The method significantly reduces target registration error from 6.2 px to approximately 2.4 px and improves AUC@25 from 0.770 to 0.938 on the FIRE dataset.

Gaussian Primitive Optimization (GPO) denotes the iterative framework introduced in "Gaussian Primitive Optimized Deformable Retinal Image Registration" (Tian et al., 23 Aug 2025). It is a deformable retinal image registration method designed for a regime in which large homogeneous / textureless areas provide almost no usable gradients, while sparse vascular structures are highly informative but occupy only a small fraction of the image. GPO addresses this by anchoring optimization on salient vascular keypoints, modeling them as Gaussian primitives with trainable position, displacement, and radius, and propagating their displacement information outward using Gaussian primitives and KNN-based message passing. After an initial coarse alignment, the method constructs a globally coherent displacement field from a minimal set of descriptor-based control nodes, and optimizes the whole framework end-to-end with both keypoint-level and image-level supervision. On the FIRE dataset, it reduces the target registration error from 6.2 px to approximately 2.4 px and increases the AUC at 25 px from 0.770 to 0.938 (Tian et al., 23 Aug 2025).

1. Problem setting and motivation

Deformable retinal registration is described as notoriously difficult because retinal images contain large homogeneous / textureless areas with almost no usable gradients, sparse vascular structures that are highly informative but occupy only a small fraction of the image, and therefore standard dense similarity losses are often dominated by easy-to-match background regions while crucial vessel alignment signals are diluted (Tian et al., 23 Aug 2025). The paper further states that vessels occupy less than 15% of the image area, so the gradients associated with the vasculature can be overwhelmed by the rest of the image.

Within this problem formulation, GPO is motivated by a specific optimization failure mode: classical similarity-driven optimization and many deep learning approaches struggle because homogeneous regions provide little or no gradient, vessels are sparse but critical, and dense deformation estimation is inefficient. The framework therefore proposes a sparse, structured optimization strategy: use key anatomical landmarks as control points and let them broadcast deformation information to nearby pixels. The emphasis is not only on alignment accuracy, but on finding enough reliable gradient signal to drive learning.

A common misconception is to regard GPO as a conventional dense-field predictor. The method is instead explicitly presented as a sparse, anatomically grounded registration formulation that places optimization effort on informative anatomical structures and then propagates local deformation information to textureless areas. This shift in optimization locus is central to the method’s design.

2. Registration pipeline and control-node construction

GPO is organized as a four-stage pipeline (Tian et al., 23 Aug 2025). The first stage performs initial coarse alignment, in which a descriptor-based model estimates a global transform to roughly align the moving image $I_m$ to the fixed image $I_f$:

$I_m^{(\text{coarse})}(x) = I_m(Ax + b).$

This coarse step reduces large pose or global misalignment so that subsequent optimization can focus on local nonrigid deformation.

The second stage extracts keypoints at salient anatomical structures from the coarse-aligned images. The descriptor network produces matched keypoints

$\{(g_i^f, g_i^m)\}_{i=1}^N,$

placed on meaningful structures such as major vessel junctions, vessel edges, and other distinctive anatomical sites. These matched points become descriptor-based control nodes (DCNs). When descriptor matches are unavailable or too sparse, the framework also defines grid-based control nodes (GCNs), sampled on a regular lattice.

The distinction between DCNs and GCNs is important because it clarifies the role of descriptors in GPO. Descriptors are used to initialize meaningful control nodes; they are not the full registration model. This means GPO differs from descriptor-only pipelines that often produce a single global transformation or limited refinement.

Control-node type	Initialization source	Intended role
DCN	Matched keypoints on salient anatomical structures	Minimal set of descriptor-based control nodes
GCN	Regular lattice	Fallback when descriptor matches are unavailable or too sparse

This design makes the control-node set anatomically meaningful when possible, while preserving a regularized alternative when descriptor support is limited.

3. Gaussian primitive parameterization and KNN interpolation

Each control node is modeled as a Gaussian primitive with three learnable attributes: position $\mathbf{g}_i \in \mathbb{R}^2$, displacement / translation $\mathbf{t}_i \in \mathbb{R}^2$, and radius $r_i \in \mathbb{R}^+$ (Tian et al., 23 Aug 2025). For DCNs, the position is initialized by the fixed-image keypoint, $\mathbf{g}_i \equiv g_i^f$, and the initial displacement is

$\mathbf{t}_i^{(0)} = g_i^m - g_i^f.$

The radius is parameterized through a learnable scalar $\beta_i$:

$I_f$0

where $I_f$1 is the sigmoid. The constant $I_f$2 is included so that the radius never collapses to zero, which helps prevent vanishing gradients.

The Gaussian radius gives each node an adaptive spatial influence: small radius implies localized deformation support, whereas larger radius implies broader smooth influence. This allows the deformation representation to adapt to local deformation scales rather than enforcing a uniform support width across the image.

The final displacement field is not predicted densely. Instead, for each pixel location $I_f$3, GPO synthesizes the displacement from the top $I_f$4 nearest control nodes using Gaussian weights:

$I_f$5

with normalized affinities

$I_f$6

The normalization ensures $I_f$7.

The use of the top $I_f$8 nearest control nodes is justified in the paper as a way to reduce computation, avoid unnecessary global coupling, preserve local structure, and still propagate enough information to keep the field coherent. The reported setting is $I_f$9 as a good balance of accuracy and runtime (Tian et al., 23 Aug 2025). The interpolation mechanism is explicitly interpreted as structured message passing: each control node stores a local displacement signal, each pixel queries its nearest control nodes, and the pixel’s deformation is formed by blending those node signals with Gaussian weights.

4. Optimization objective, warping, and gradient flow

At iteration $I_m^{(\text{coarse})}(x) = I_m(Ax + b).$0, the displacement field is used to warp the coarse-aligned moving image:

$I_m^{(\text{coarse})}(x) = I_m(Ax + b).$1

The model then iteratively updates the Gaussian primitive parameters $I_m^{(\text{coarse})}(x) = I_m(Ax + b).$2 (Tian et al., 23 Aug 2025).

The objective is described as a two-term loss comprising a global cross-correlation loss with matched control nodes and a normalized cross-correlation loss between the fixed image and warped image. The implementation details report $I_m^{(\text{coarse})}(x) = I_m(Ax + b).$3 and $I_m^{(\text{coarse})}(x) = I_m(Ax + b).$4. In the paper’s interpretation, this combines point-level consistency from the matched keypoints with image-level intensity alignment from the NCC term.

A central claim of GPO is that it improves gradient flow in sparse-feature regions. Because the displacement field is a weighted sum over nearby nodes, gradients flow back through the Gaussian weights. The paper writes this as

$I_m^{(\text{coarse})}(x) = I_m(Ax + b).$5

The stated consequence is that pixels near vessel edges still produce strong gradients, those gradients are routed back to nearby control nodes, and each node accumulates signal from a neighborhood of informative pixels. This is presented as a mechanism for mitigating vanishing gradient signal in textureless areas.

The resulting optimization behavior differs from dense similarity-driven training in a specific way: optimization is anchored where the anatomy is informative rather than being overwhelmed by large uninformative regions. This also explains why the method emphasizes salient vascular keypoints rather than uniform spatial coverage.

5. Experimental setting and reported performance on FIRE

The principal evaluation is conducted on the FIRE dataset, which contains 134 retinal image pairs, 10 expert-annotated landmarks per pair, and image resolution 2912 $I_m^{(\text{coarse})}(x) = I_m(Ax + b).$6 2912 (Tian et al., 23 Aug 2025). The reported categories are S: 71 pairs (minimal distortion), A: 4 pairs (anatomical changes), and P: 49 pairs (perspective distortion). As preprocessing, images are resized to 1024 $I_m^{(\text{coarse})}(x) = I_m(Ax + b).$7 1024 and Gaussian blur is applied for anti-aliasing.

Implementation details specify PyTorch, an NVIDIA A100 GPU, the Adam optimizer, learning rates $I_m^{(\text{coarse})}(x) = I_m(Ax + b).$8 for node position and $I_m^{(\text{coarse})}(x) = I_m(Ax + b).$9 for displacement and radius, and $\{(g_i^f, g_i^m)\}_{i=1}^N,$0. The GPO-DCN configuration uses $\{(g_i^f, g_i^m)\}_{i=1}^N,$1 and 100 iterations, while GPO-GCN uses a $\{(g_i^f, g_i^m)\}_{i=1}^N,$2 grid and 200 iterations. Evaluation uses TRE (Target Registration Error), where lower is better, and AUC for TRE thresholds at 15, 25, and 50 pixels, where higher is better.

Method	TRE	AUC@25
GPO-DCN	2.352 px	0.938
GPO-GCN	2.649 px	0.914
GeoFormer	6.201 px	0.770
RetinaRegNet	2.766 px	0.910

For GPO-DCN, the paper reports AUC@15 = 0.906 and AUC@50 = 0.964; for GPO-GCN, AUC@15 = 0.888 and AUC@50 = 0.926 (Tian et al., 23 Aug 2025). Relative to the best descriptor-based baseline, GeoFormer, GPO-DCN reduces TRE by about 3.8 px and boosts AUC@25 from 0.770 to 0.938. Relative to the strongest learning-based baseline, RetinaRegNet, it improves TRE by about 0.4 px.

These results support the paper’s main empirical claim: descriptor-based initialization alone is insufficient for fine local deformation, while dense learning-based registration can still be dominated by background regions. GPO positions itself between these regimes by using descriptors only to initialize meaningful control nodes and then iteratively optimizing a deformable field with local Gaussian blending.

6. Terminological scope and related Gaussian-primitive optimization directions

Within the literature cited here, the explicit acronym GPO refers to the retinal registration framework introduced in "Gaussian Primitive Optimized Deformable Retinal Image Registration" (Tian et al., 23 Aug 2025). At the same time, several contemporaneous papers in 3D Gaussian Splatting optimize Gaussian primitives themselves under different names, which suggests a broader conceptual use of the phrase “Gaussian primitive optimization” even when the paper-specific acronym is different.

"Optimized Minimal 3D Gaussian Splatting" introduces Optimized Minimal Gaussian representation (OMG), not GPO, and explicitly states that there is no acronym “GPO” in the paper (Lee et al., 21 Mar 2025). Its method first reduces the primitive set to a minimal, high-value subset using local distinctiveness-aware importance scoring, then compresses the remaining primitives with a compact representation and Sub-Vector Quantization; the paper reports nearly 50% storage reduction compared to the previous state-of-the-art and 600+ FPS rendering while maintaining high rendering quality.

"ProtoGS: Efficient and High-Quality Rendering with 3D Gaussian Prototypes" learns Gaussian prototypes to represent Gaussian primitives, using SfM anchor points to group primitives and K-means to derive prototypes within each group (Gao et al., 21 Mar 2025). It renders directly with prototypes rather than treating prototype learning as a separate post-training compression stage. The reported NeRF-Synthetic comparison lists ProtoGS at 45K primitives and PSNR 33.45, alongside 225 FPS.

"Optimized 3D Gaussian Splatting using Coarse-to-Fine Image Frequency Modulation" reduces the number of Gaussian primitives during training itself rather than compressing the model after training (Farooq et al., 18 Mar 2025). On the baseline 3DGS, it reports an average reduction of 62% in Gaussians, a 40% reduction in the training GPU memory requirements, and a 20% reduction in optimization time without sacrificing the visual quality.

"DashGaussian: Optimizing 3D Gaussian Splatting in 200 Seconds" reframes optimization as a scheduling problem over rendering resolution and primitive number, describing optimization complexity as dominated by these two coupled factors (Chen et al., 24 Mar 2025). "Plug-and-Play PDE Optimization for 3D Gaussian Splatting: Toward High-Quality Rendering and Reconstruction" treats Gaussian updates as a PDE system, introduces a viscous term to ensure stable optimization, and uses the Material Point Method to obtain a stable numerical solution (Mo et al., 17 Sep 2025). "Gradient-Driven Natural Selection for Compact 3D Gaussian Splatting" formulates compact 3DGS as selecting a retained subset under a budget by letting rendering gradients compete with a global opacity-decay pressure; it reports over 0.6 dB PSNR gain under 15% budgets (Deng et al., 21 Nov 2025).

This broader landscape indicates an important distinction. In retinal registration, GPO denotes a specific iterative deformable registration framework based on salient anatomical control nodes, Gaussian primitives, and KNN Gaussian interpolation (Tian et al., 23 Aug 2025). In 3DGS, closely related work uses other names for methods that optimize Gaussian primitives through pruning, prototype learning, scheduling, PDE stabilization, or gradient-driven selection (Lee et al., 21 Mar 2025). A plausible implication is that “Gaussian primitive optimization” now functions both as a named method in retinal registration and as a broader conceptual category for methods that treat the Gaussian primitive set itself as the optimization target.