VRUM: Voxel Refinement Upsampling Module

• VRUM is a voxel upsampling module that refines low-resolution 3D feature maps through dual-branch interpolation and deconvolution strategies.
• It combines stable trilinear interpolation with learnable multi-scale transposed convolutions to correct local geometric and semantic errors.
• VRUM enhances segmentation boundary accuracy in medical imaging, as evidenced by improved Dice scores in brain tumor studies.

Voxel Refinement UpSampling Module (VRUM) denotes a class of voxel-space upsampling and refinement operators that increase volumetric resolution while correcting local errors in geometry or semantics. The term is introduced explicitly in the brain tumor segmentation model GMLN-BTS, where VRUM is a decoder-side block that “synergistically combines linear interpolation and multi-scale transposed convolutions to suppress artifacts while preserving high-frequency details, improving segmentation boundary accuracy” (Huo et al., 14 Jul 2025). Related arXiv work uses closely aligned mechanisms under different names: hierarchical conditional diffusion for coarse-to-fine voxel detailisation, 3D U-Net-based refinement of semantic voxel predictions, density-guided voxel resampling for point generation, adversarial conditional refinement of coarse voxel shapes, and analytic subdivision of voxel volumes before surface extraction (Rasoulzadeh et al., 2024, Zhang et al., 20 Dec 2025, Du et al., 2024, Chen et al., 2020, Stock et al., 2023).

1. Origin and conceptual scope

In its explicit formulation, VRUM is part of the decoder in GMLN-BTS within the Edge Iterative MRI Lesion Localization System (EdgeIMLocSys). The processing chain is: multi-modal MRI inputs are encoded by the Modality-Aware Adaptive Encoder (M2AE), fused by the Graph-based Multi-Modal Collaborative Interaction Module (G2MCIM) and a lightweight Transformer, and then passed to VRUM for upsampling and voxel-level refinement before tumor or region prediction. Within EdgeIMLocSys, VRUM is updated together with the rest of the segmentation model during Continuous Learning from Human Feedback, so the module adapts to scanner-specific artifacts and boundary characteristics (Huo et al., 14 Jul 2025).

The immediate motivation is the complementarity between two standard 3D upsampling primitives. Trilinear interpolation is stable, parameter-free, and free of checkerboard artifacts, but behaves as a low-pass filter that blurs boundaries and suppresses fine textures. ConvTranspose3D can reconstruct high-frequency details such as tumor edges and small structures, but is prone to checkerboard artifacts caused by zero insertion and uneven overlap. VRUM is therefore designed to combine a smooth, artifact-free base with learnable detail recovery.

The broader literature shows that “voxel refinement” is not a single canonical architecture. In some papers it refines occupancy voxels, in others semantic voxel labels, signed distance values, latent voxel features, or point densities defined over voxel cells. This suggests that VRUM is best understood as a functional category: a module operating on volumetric discretizations to improve fidelity after coarse prediction, coarse generation, or low-resolution reconstruction.

2. Internal architecture in GMLN-BTS

The GMLN-BTS VRUM takes a low-resolution 3D feature map

$$\mathbf{X} \in \mathbb{R}^{B \times C \times \frac{D}{4} \times \frac{H}{4} \times \frac{W}{4}}$$

and processes it through two parallel branches followed by channel fusion (Huo et al., 14 Jul 2025).

The first branch is the interpolation plus Spatial Pixel Refinement branch. It begins with trilinear upsampling by a factor of $2$:

$$\mathbf{X}_{\mathrm{interp}} = \mathrm{TrilinearUpsample}(\mathbf{X}; s=2).$$

A $3 \times 3 \times 3$ Conv3D with stride $1$ and padding $1$ refines the interpolated features:

$$\mathbf{X}_{\mathrm{conv}} = \mathrm{Conv3D}(\mathbf{X}_{\mathrm{interp}};\ \mathbf{W}_{\mathrm{conv}},\ \mathrm{kernel}=3,\ \mathrm{stride}=1,\ \mathrm{pad}=1).$$

After a LeakyReLU with $\alpha=0.01$, the paper applies a ConvTranspose3D with kernel $4 \times 4 \times 4$, stride $2$, and padding $2$0:

$2$1

This branch is described as using interpolation as a stable base and transposed convolution as a region-specific correction mechanism.

The second branch is a multi-scale transposed convolution branch that upsamples directly from $2$2 using two kernel sizes. The small-kernel path uses

$2$3

while the large-kernel path uses

$2$4

Both outputs have intermediate channel dimension

$2$5

They are concatenated, fused by a $2$6 Conv3D, normalized by BatchNorm3d, and activated by ReLU.

Final fusion concatenates the outputs of the two branches and compresses them with a $2$7 Conv3D:

$2$8

Operationally, the design is parallel in its feature generation and sequential in its fusion.

3. Voxel-level refinement semantics and supervision

In the VRUM formulation of GMLN-BTS, “voxel refinement” means learned correction at individual voxel positions on top of a coarse upsampled representation rather than resolution increase alone. The interpolation branch provides a smooth per-voxel scaffold, while the Spatial Pixel Refinement submodule and the multi-scale transposed convolutions supply additional local and contextual corrections. Although the paper does not define refinement as explicit residual addition, it describes the module as superimposing deconvolution-based corrections through fusion (Huo et al., 14 Jul 2025).

Boundary accuracy is addressed through four interacting mechanisms. First, interpolation prevents jagged boundaries and avoids the checkerboard behavior associated with naive deconvolution. Second, the small-kernel $2$9 transposed convolution is detail-oriented and targets fine local variations, edges, and textures. Third, the large-kernel $$\mathbf{X}_{\mathrm{interp}} = \mathrm{TrilinearUpsample}(\mathbf{X}; s=2).$$0 transposed convolution is smoothness-oriented and captures broader context that suppresses artifacts and improves coherence. Fourth, the $$\mathbf{X}_{\mathrm{interp}} = \mathrm{TrilinearUpsample}(\mathbf{X}; s=2).$$1 fusion layer learns the voxel-wise mixture of stable and detail-rich components.

VRUM is not trained with a module-specific auxiliary loss. Instead, it is optimized indirectly through the segmentation objective applied to the final decoder output:

$$\mathbf{X}_{\mathrm{interp}} = \mathrm{TrilinearUpsample}(\mathbf{X}; s=2).$$2

with

$$\mathbf{X}_{\mathrm{interp}} = \mathrm{TrilinearUpsample}(\mathbf{X}; s=2).$$3

where $$\mathbf{X}_{\mathrm{interp}} = \mathrm{TrilinearUpsample}(\mathbf{X}; s=2).$$4 and $$\mathbf{X}_{\mathrm{interp}} = \mathrm{TrilinearUpsample}(\mathbf{X}; s=2).$$5 denote predicted and ground-truth voxel sets. The paper states that no extra per-VRUM auxiliary supervision is used. This places VRUM within an end-to-end decoder optimization regime rather than a separately supervised refinement stage.

A recurrent misconception is that voxel upsampling in medical decoders is equivalent to interpolation followed by a convolution. VRUM is more specific: it is a dual-branch volumetric upsampling block that treats artifact suppression and high-frequency recovery as competing design objectives and resolves them by explicit branch specialization and learned fusion.

4. Empirical profile in brain tumor segmentation

Within the full GMLN-BTS model, the reported result on BraTS2017 is a Dice score of $$\mathbf{X}_{\mathrm{interp}} = \mathrm{TrilinearUpsample}(\mathbf{X}; s=2).$$6 with $$\mathbf{X}_{\mathrm{interp}} = \mathrm{TrilinearUpsample}(\mathbf{X}; s=2).$$7 million parameters, described as a $$\mathbf{X}_{\mathrm{interp}} = \mathrm{TrilinearUpsample}(\mathbf{X}; s=2).$$8 reduction compared to mainstream 3D Transformer models. The paper attributes the overall efficiency to the complete architecture, but VRUM is part of the decoder-side mechanism that transforms low-resolution fused features into higher-resolution maps under that parameter budget (Huo et al., 14 Jul 2025).

The ablation study isolates the contribution of major components. A baseline without G2MCIM, M2AE, and VRUM yields $$\mathbf{X}_{\mathrm{interp}} = \mathrm{TrilinearUpsample}(\mathbf{X}; s=2).$$9 mean Dice. Adding G2MCIM gives $3 \times 3 \times 3$0, adding both G2MCIM and M2AE gives $3 \times 3 \times 3$1, and the full model including VRUM reaches $3 \times 3 \times 3$2. On that ablation path, the incremental gain associated with adding VRUM on top of G2MCIM and M2AE is $3 \times 3 \times 3$3 percentage points:

$3 \times 3 \times 3$4

Because the surrounding encoder and fusion blocks remain unchanged in that comparison, the result is consistent with the intended role of VRUM as a boundary-sensitive decoder refinement block.

In the broader model comparison, GMLN-BTS reports Avg Dice $3 \times 3 \times 3$5, WT $3 \times 3 \times 3$6, ET $3 \times 3 \times 3$7, and TC $3 \times 3 \times 3$8. SegFormer3D, described as having a similar parameter scale but using conventional upsampling, reports Avg Dice $3 \times 3 \times 3$9 and ET $1$0. The paper does not attribute that entire gap to VRUM alone, but the authors explicitly connect VRUM to preservation of high-frequency information and boundary accuracy, which is particularly relevant for enhancing-tumor regions because they are small, irregular, and boundary-sensitive.

The implementation remains fully 3D. VRUM uses Conv3D, ConvTranspose3D, BatchNorm3d, LeakyReLU, ReLU, and trilinear upsampling; it does not rely on separate post-processing such as CRFs. Its novelty is therefore architectural rather than procedural: the refinement is embedded inside the decoder and learned jointly with segmentation.

5. Related voxel refinement paradigms in arXiv literature

The term VRUM is not used uniformly across arXiv, but several papers describe functionally similar voxel refinement or upsampling modules. They differ mainly in target representation, conditioning signal, and supervision regime.

Paper	Representation	Refinement mechanism
"ArchComplete" (Rasoulzadeh et al., 2024)	Binary occupancy voxels	Hierarchy of local 3D conditional diffusion models
"Point Cloud Upsampling via Cascaded Refinement Network" (Du et al., 2022)	Point sets, with voxel-oriented extension discussed in the supplied details	Coarse-to-fine residual refinement and multi-stage supervision
"DECOR-GAN" (Chen et al., 2020)	Coarse and detailed voxel shapes	Conditional adversarial refinement with style code and masks
"Arbitrary-Scale Point Cloud Upsampling by Voxel-Based Network with Latent Geometric-Consistent Learning" (Du et al., 2024)	Voxel densities for point generation	Density-guided grid resampling plus continuous point regression
"Enhancing 3D Semantic Scene Completion with a Refinement Module" (Zhang et al., 20 Dec 2025)	Semantic voxel predictions	3D U-Net refinement with PNAM and VLGM
"Towards smoother surfaces by applying subdivision to voxel data" (Stock et al., 2023)	Scalar voxel volumes	Analytic subdivision before surface extraction

ArchComplete gives a particularly direct voxel-space analogue to a refinement upsampler. It uses a two-stage pipeline in which a $1$1 coarse occupancy grid is refined hierarchically to $1$2 and $1$3 by local 3D conditional DDPMs trained on aligned coarse-to-fine patch pairs $1$4 and $1$5. The reverse process is conditioned on the spatially aligned coarse chunk:

$1$6

with conditioning implemented by concatenating the noised fine chunk and the subdivided coarse chunk as two volumetric channels in a 3D U-Net with residual blocks. For $1$7, the paper reports $1$8 s and $1$9 GB, versus $1$0 s and $1$1 GB for DECOR-GAN; for $1$2, it reports $1$3 s and $1$4 GB versus $1$5 s and $1$6 GB (Rasoulzadeh et al., 2024).

DECOR-GAN represents another refinement lineage: a single-stage 3D CNN generator conditioned on an $1$7-dimensional style code upsamples coarse occupancy voxels by a factor of $1$8 per dimension and is supervised by a 3D PatchGAN with global and style-specific branches. Its generator is constrained by a loose generator mask, and discriminator masks enforce occupancy in the projected high-resolution regions. The design emphasizes content preservation, plausible local detail, and style-conditioned geometric variation rather than artifact suppression in a medical decoder (Chen et al., 2020).

PU-VoxelNet shows a different voxel-centric route: a 3D CNN predicts per-voxel occupancy $1$9 and density $$\mathbf{X}_{\mathrm{conv}} = \mathrm{Conv3D}(\mathbf{X}_{\mathrm{interp}};\ \mathbf{W}_{\mathrm{conv}},\ \mathrm{kernel}=3,\ \mathrm{stride}=1,\ \mathrm{pad}=1).$$0, defines a filled density $$\mathbf{X}_{\mathrm{conv}} = \mathrm{Conv3D}(\mathbf{X}_{\mathrm{interp}};\ \mathbf{W}_{\mathrm{conv}},\ \mathrm{kernel}=3,\ \mathrm{stride}=1,\ \mathrm{pad}=1).$$1, samples candidate cells by multinomial drawing, and then uses density-guided farthest point sampling with

$$\mathbf{X}_{\mathrm{conv}} = \mathrm{Conv3D}(\mathbf{X}_{\mathrm{interp}};\ \mathbf{W}_{\mathrm{conv}},\ \mathrm{kernel}=3,\ \mathrm{stride}=1,\ \mathrm{pad}=1).$$2

to suppress outliers before regressing continuous point coordinates from voxel centers. Its refinement target is not a dense voxel output but a high-fidelity point cloud reconstructed from a voxel density field (Du et al., 2024).

ESSC-RM demonstrates that voxel refinement can also be applied after a baseline network has already produced a full semantic voxel grid. The refinement module is a 3D U-Net operating on discrete predicted labels, with PNAM replacing intermediate decoder aggregation blocks and VLGM injecting text-derived priors. On SemanticKITTI, the paper reports mIoU improvement from $$\mathbf{X}_{\mathrm{conv}} = \mathrm{Conv3D}(\mathbf{X}_{\mathrm{interp}};\ \mathbf{W}_{\mathrm{conv}},\ \mathrm{kernel}=3,\ \mathrm{stride}=1,\ \mathrm{pad}=1).$$3 to $$\mathbf{X}_{\mathrm{conv}} = \mathrm{Conv3D}(\mathbf{X}_{\mathrm{interp}};\ \mathbf{W}_{\mathrm{conv}},\ \mathrm{kernel}=3,\ \mathrm{stride}=1,\ \mathrm{pad}=1).$$4 for CGFormer and from $$\mathbf{X}_{\mathrm{conv}} = \mathrm{Conv3D}(\mathbf{X}_{\mathrm{interp}};\ \mathbf{W}_{\mathrm{conv}},\ \mathrm{kernel}=3,\ \mathrm{stride}=1,\ \mathrm{pad}=1).$$5 to $$\mathbf{X}_{\mathrm{conv}} = \mathrm{Conv3D}(\mathbf{X}_{\mathrm{interp}};\ \mathbf{W}_{\mathrm{conv}},\ \mathrm{kernel}=3,\ \mathrm{stride}=1,\ \mathrm{pad}=1).$$6 for MonoScene, under multiscale supervision (Zhang et al., 20 Dec 2025).

The analytic subdivision work occupies yet another point in the design space. It does not learn parameters at all; instead it applies a three-dimensional subdivision scheme directly to scalar voxel volumes before isosurface extraction. One subdivision iteration doubles resolution in each axis and increases storage by $$\mathbf{X}_{\mathrm{conv}} = \mathrm{Conv3D}(\mathbf{X}_{\mathrm{interp}};\ \mathbf{W}_{\mathrm{conv}},\ \mathrm{kernel}=3,\ \mathrm{stride}=1,\ \mathrm{pad}=1).$$7. The nonlinear version avoids overshoots and produces smoothly interpolated data while also preserving edges (Stock et al., 2023).

The cascaded refinement network for point cloud upsampling is not itself a voxel module, but the supplied technical summary explicitly maps its three-stage residual logic to voxel representations: geometric displacement of voxel centers, residual correction of occupancy or signed distance values, or residual refinement of voxel features. This suggests a transferable design principle for VRUM-like systems: coarse-to-fine stages with offset or residual prediction can mitigate regression difficulty and reduce conflicts between stages (Du et al., 2022).

6. Limitations, misconceptions, and open directions

VRUM-style modules inherit the trade-offs of their underlying upsampling operators. In GMLN-BTS, the multi-scale deconvolution branch improves detail recovery but increases computation relative to interpolation-only schemes, and the module is designed within a U-Net-like decoder context rather than as a universally architecture-agnostic primitive. The paper does not list explicit VRUM-specific failure cases, but it notes the remaining complexity of ConvTranspose3D-based designs and the possibility that very resource-limited devices may still find multi-scale 3D deconvolution non-trivial (Huo et al., 14 Jul 2025).

The hierarchical voxel refinement literature identifies additional structural limits. ArchComplete emphasizes error accumulation across the hierarchy: higher-resolution grids cannot easily fix artifacts from prior levels, and purely local chunk-wise detailisation can cause slight topological inconsistencies because the model lacks awareness of global structures. The same paper proposes appending refinement networks to each level, adding a global shape encoding branch, exploring alternative diffusion parameterizations, and using sparse 3D convolutions, octree or sparse voxel grids, or mixed CNN-transformer architectures to scale beyond $$\mathbf{X}_{\mathrm{conv}} = \mathrm{Conv3D}(\mathbf{X}_{\mathrm{interp}};\ \mathbf{W}_{\mathrm{conv}},\ \mathrm{kernel}=3,\ \mathrm{stride}=1,\ \mathrm{pad}=1).$$8 (Rasoulzadeh et al., 2024).

Analytic voxel refinement has a different failure mode. Subdivision volumes improve smoothness and can preserve edges, but smoothing may change metrological measurements and blur very sharp corners. The reported $$\mathbf{X}_{\mathrm{conv}} = \mathrm{Conv3D}(\mathbf{X}_{\mathrm{interp}};\ \mathbf{W}_{\mathrm{conv}},\ \mathrm{kernel}=3,\ \mathrm{stride}=1,\ \mathrm{pad}=1).$$9 storage increase per iteration also makes repeated refinement impractical on large industrial CT volumes. This highlights a general point: “refinement” can mean higher visual smoothness, better semantic consistency, or more plausible local detail, but these objectives do not always coincide (Stock et al., 2023).

A common misconception is that voxel refinement is synonymous with interpolation. The arXiv record shows a broader design space: dual-branch interpolation/deconvolution hybrids, hierarchical diffusion conditioned on coarse chunks, adversarial conditional detailisation, density-guided resampling from voxel fields, post-hoc attention-based refinement of predicted semantic grids, and deterministic subdivision. A plausible implication is that future VRUM research will be organized less by the name of the module and more by four orthogonal design choices: the refined variable, the conditioning source, the scale schedule, and the supervision strategy.