OmniSlicer: Omnidirectional Volume Slicing

Updated 14 July 2026

OmniSlicer is an omnidirectional volume slicing method that extracts 2D cross-sectional images from 3D tomographic data using both canonical and optimized non-canonical orientations.
It employs spherical sampling with repulsion-based optimization to uniformly distribute slicing directions and maximize lesion visibility.
The framework integrates directly with 2D vision foundation models through a spherical graph structure, resulting in improved performance in medical image classification.

Searching arXiv for papers relevant to “OmniSlicer,” especially the TomoGraphView paper and any competing or overloaded uses of the term. OmniSlicer is the omnidirectional volume slicing component of TomoGraphView, together with its released Python library (OmniSlicer on PyPI). It is a mechanism that turns a 3D tomographic volume into a set of 2D cross-sectional images taken from many directions, including both canonical and non-canonical orientations, and is designed to feed these slices into powerful 2D vision foundation models while still preserving 3D spatial structure via a spherical graph (Kiechle et al., 12 Nov 2025). In this formulation, OmniSlicer addresses a specific limitation of conventional slice-based decomposition for 3D medical image classification: canonical axial, coronal, and sagittal planes may inadequately capture the spatial extent of target structures when these are misaligned with standardized viewing planes.

1. Definition and conceptual scope

OmniSlicer is introduced through a “novel omnidirectional volumetric slicing strategy” that “extends beyond axial, coronal, and sagittal orientations, thereby enabling a richer characterization of volumetric structures that are not naturally aligned with canonical axes” (Kiechle et al., 12 Nov 2025). Conceptually, the method encloses a 3D volume in a sphere centered at the volume center, places $N$ points on the sphere surface, and uses those points to define slicing directions. Three points are fixed to represent canonical planes, and the remaining $N-3$ points are optimized to be uniformly distributed on the sphere via a repulsion-based optimization. For each direction, the method extracts the “most representative” slice, operationally defined as the plane and position that maximizes visible lesion area.

The resulting representation consists of $N$ 2D slices: canonical in-plane slices plus many off-axis, non-canonical cross-sections. This construction is motivated by two considerations. First, tumors or anatomical structures may be elongated or oriented obliquely, so a single axial, coronal, or sagittal “largest lesion slice” can be a poor summary. Second, the approach is explicitly designed to reuse 2D vision transformers such as DINOv2 on 3D data, thereby avoiding training 3D models from scratch in low-data regimes.

Within the paper’s comparison protocol, OmniSlicer is defined against four canonical baselines. “2D-axial” denotes a single axial slice with maximum lesion area. “2D-axial+” denotes the axial “largest lesion slice” plus symmetric neighbors, for a total of $N$ axial slices. “2.5D” denotes the largest lesion slice in each canonical direction. “2.5D+” extends this with neighboring slices in each canonical direction. OmniSlicer includes the canonical directions but systematically augments them with many non-canonical directions.

2. Spherical sampling and slice extraction

Let the preprocessed 3D volume be centered at the origin and enclosed by a bounding sphere

$\mathcal{S} = \{ \mathbf{x} \in \mathbb{R}^3 : \|\mathbf{x}\|_2 \leq r \}.$

Three points are fixed on the sphere surface at $(r,0,0)$ , $(0,r,0)$ , and $(0,0,r)$ , corresponding, up to coordinate convention, to axial, coronal, and sagittal view normals (Kiechle et al., 12 Nov 2025). The remaining sampling points are optimized so that all points are approximately equally distributed on the sphere surface.

The optimization is formulated as a Thomson-like repulsion problem. If $\{x_i\}_{i=1}^N$ are unit-norm points on the sphere, the pairwise Coulomb energy is

$E = \sum_{i \neq j} \frac{1}{\|x_i - x_j\|^2}.$

The gradient of $N-3$ 0 defines repulsive forces; non-fixed points are iteratively moved along these forces and re-projected to the unit sphere. The paper summarizes this process as repulsion-based point optimization with fixed canonical points, randomly sampled non-fixed points, iterative force computation, position updates $N-3$ 1, and reprojection $N-3$ 2.

Each optimized point $N-3$ 3 defines a view direction. For each spherical point, the slicing plane is aligned so that its normal vector points to the respective point on the sphere. Slice extraction then proceeds by identifying the largest lesion slice from the respective canonical or non-canonical plane. The paper does not write this optimization in closed form, but states the operational procedure: for each orientation, scan along the normal direction within the volume bounds, intersect the segmentation mask, and select the plane whose intersection with the mask has maximal area.

The slices are planar cross-sections in 3D space. They are neither curved surfaces nor MIP-like projections. This distinction is operationally important because the representation remains a collection of true 2D cross-sections sampled from the 3D voxel grid. Volumes are resampled to isotropic $N-3$ 4 mm spacing. For omnidirectional slicing, scalar image volumes use linear interpolation, whereas segmentation masks use nearest-neighbor interpolation to preserve labels. The mapping from 3D coordinates to 2D slice coordinates is not given in closed form, but the paper describes the standard orthonormal-basis construction with $N-3$ 5 and in-plane coordinates defined along $N-3$ 6 and $N-3$ 7 (Kiechle et al., 12 Nov 2025).

A common misconception is to treat OmniSlicer as a generic projection operator. In the TomoGraphView formulation, it is specifically a directional plane-selection mechanism that searches for lesion-centered planar slices. Its objective is not volumetric rendering, but a view set that better intersects the morphology of structures that are poorly summarized by canonical orientations.

3. Coupling with 2D foundation models and graph aggregation

OmniSlicer is designed to produce slices that can be passed directly to a frozen 2D ViT encoder, specifically DINOv2-small. DINOv2 is described as a self-supervised ViT trained on 142M natural images; the small variant has approximately 21M parameters. Input slices are resized to $N-3$ 8 pixels, “obtained by linearly up-sampling the input images if not matched initially,” and each slice is encoded into a 384-dimensional feature vector $N-3$ 9. “The encoder is kept frozen during all experiments” (Kiechle et al., 12 Nov 2025).

This slice-wise embedding stage places OmniSlicer in a larger architectural pipeline. All volumes are reoriented to RAS+ and resampled to $N$ 0 mm. Symmetric cropping is applied around the lesion to reduce distortion when resizing to $N$ 1. The resulting slice embeddings can be used directly by simple aggregation heads such as an MLP, LSTM, or MST baseline, but the full TomoGraphView system instead treats them as node features in a spherical graph.

The graph uses the same sampling points $N$ 2 that define slice orientations. A Delaunay triangulation on the spherical surface establishes local adjacency, yielding an undirected graph $N$ 3 with one node per view and node attributes given by the DINOv2 slice features. Local triangulation edges initially carry weight $N$ 4. To enrich connectivity, the method adds cross-connections between all node pairs, producing a complete graph with edge weights determined by inverse hop distance on the local triangulated graph: $N$ 5

Message passing is implemented with GraphSAGE-type layers using mean and max aggregation. After $N$ 6 layers, node embeddings are mean-pooled to produce a graph-level representation $N$ 7, followed by a linear classifier for binary prediction. In this context, OmniSlicer supplies the view geometry and the slice embeddings, while the spherical graph and GNN are responsible for reintroducing volumetric structure that would otherwise be lost in a purely slice-wise “bag of views” model.

4. Empirical behavior across oncology classification tasks

The evaluation comprises six oncological classification datasets: brain tumors (MRI T1c, low vs high grade), head-neck tumors (CT, HPV status), breast tumors (DCE MRI, Nottingham grade high vs low/intermediate), liver tumors (multi-phase CT, HCC vs cHCC-CCA), kidney tumors (contrast CT, WHO/ISUP high vs low grade), and soft-tissue sarcomas (MRI T2 fat-sat, high vs low grade). All tasks are binary classification, and the principal metric is AUROC (Kiechle et al., 12 Nov 2025).

The slicing-strategy benchmark shows that omnidirectional sampling improves over canonical slicing baselines when paired with an MLP head of 100k parameters. The paper reports mean AUROC across the six datasets as follows.

Strategy	Views	Mean AUROC
2D-axial+	24	0.7701
2.5D+	24	0.7885
Omnidirectional	8	0.7876
Omnidirectional	16	0.8012
Omnidirectional	24	0.8154

At the dataset level, the 24-view omnidirectional configuration improves over 2.5D+ on several tasks, including brain tumors $N$ 8, head-neck tumors $N$ 9, and soft-tissue sarcomas $N$ 0. These results are presented as evidence that omnidirectional sampling captures discriminative information beyond what is available from canonical slicing alone.

The gain is not restricted to MLP aggregation. With LSTM, the best 24-view 2D-axial+ configuration yields AUROC 0.7446, whereas the best 24-view omnidirectional configuration yields 0.7863, an absolute increase of 4.17 points. With MST, the best 16-view 2D-axial+ configuration yields 0.7768, whereas the best 24-view omnidirectional configuration yields 0.8198, an absolute increase of 4.30 points. When OmniSlicer is combined with the spherical GNN in full TomoGraphView, performance increases further: omnidirectional+MLP reaches 0.8154, whereas omnidirectional+GNN reaches 0.8372, an average AUROC increase of 2.18 points (Kiechle et al., 12 Nov 2025).

The graph-topology ablation further indicates that the choice of spherical connectivity matters. The best mean AUROC, 0.8372, is obtained with a complete graph plus inverse weighting, that is, cross-edges weighted by $N$ 1. The paper characterizes the local spherical mesh alone as competitive and efficient, but reports that distance-weighted cross-connections improve performance.

5. Software realization and operational constraints

The paper states: “We publicly share our code base and provide a user-friendly library for omnidirectional volume slicing at https://pypi.org/project/OmniSlicer.” It also cites the accompanying code base at http://github.com/compai-lab/2025-MedIA-kiechle (Kiechle et al., 12 Nov 2025). As described there, OmniSlicer is intended as an accessible and reusable implementation of omnidirectional slicing for arbitrary 3D medical volumes.

The design goals that are explicit in the technical description are easy, reusable omnidirectional slicing; a parameterizable number of views $N$ 2; stable inclusion of canonical planes; and equidistant distribution of non-canonical directions via Thomson-like repulsion. The paper experiments with $N$ 3, and the library permits selection of this parameter. The repulsion optimization has a step size $N$ 4 and a convergence criterion, although no numerical defaults are given. The sphere radius $N$ 5 is set so that the volume is fully contained, and the center is typically the volume center or a lesion-centered subvolume.

Several practical constraints follow directly from the formulation. First, slice selection by maximal lesion area requires a segmentation mask of the target structure. The paper notes, however, that in workflows based on bounding boxes, the central slice along each orientation may be a reasonable proxy, and identifies bounding-box-based evaluation and integration of detection as future work. Second, omnidirectional slicing introduces computational overhead because the volume must be re-interpolated $N$ 6 times, once per direction. The paper characterizes this overhead as manageable for $N$ 7 and modest volumes. Third, non-canonical planes are interpolated from the 3D grid, so anisotropic volumes may incur artifacts. The reported anisotropy experiment, in which the brain dataset is artificially resampled to 6 mm $N$ 8-spacing, shows an overall AUROC drop but also a shift toward larger benefits from using more views.

The number of views is therefore not presented as a fixed optimum. For isotropic or near-isotropic datasets, 8–16 views may be sufficient and sometimes preferable for efficiency. For strongly anisotropic volumes or coarse $N$ 9-spacing, 24 views confer consistent gains. The paper does not impose modality-specific choices, but CT and MRI are both handled after reorientation to RAS+ and resampling to isotropic spacing.

Within the cited literature, “OmniSlicer” refers specifically to the omnidirectional volume slicing component and Python library released with TomoGraphView, rather than to a domain-general software stack spanning all uses of “slicing” (Kiechle et al., 12 Nov 2025). At the same time, related papers exhibit an analogous design motif: a complex high-dimensional input is reorganized into smaller, more information-dense units that can be processed by a downstream model or execution engine.

In multimodal LLM systems, Nemotron 3 Nano Omni uses dynamic resolution, pixel shuffle with $\mathcal{S} = \{ \mathbf{x} \in \mathbb{R}^3 : \|\mathbf{x}\|_2 \leq r \}.$ 0 downsampling, Conv3D temporal compression, Efficient Video Sampling, and audio subsampling to reduce the token sequence seen by the core decoder. Its authors do not name this subsystem OmniSlicer as a formal product, but the technical note explicitly states that the model “can be read as a concrete instantiation of an ‘OmniSlicer’” in the sense of modality-aware compression and reordering before fusion (NVIDIA et al., 27 Apr 2026). In interactive segmentation, TomoSAM precomputes SAM image embeddings along the three Cartesian directions, then combines prompt-based 2D masks with Slicer’s “Fill Between Slices” to reconstruct 3D labels, illustrating another slice-centric bridge between 2D foundation models and volumetric data (Semeraro et al., 2023). In distributed linear algebra, the phrase is used only interpretively: a universal one-sided matrix multiplication algorithm derives overlapping local tile multiplications through index slicing and one-sided get/accumulate operations, and the accompanying note describes this as what “we can interpret as an ‘OmniSlicer’” for arbitrary partitionings and replication factors (Brock et al., 10 Oct 2025).

This suggests that “OmniSlicer” functions both as a specific proper noun in medical imaging and as a broader design pattern centered on systematic slicing, compression, or decomposition of high-dimensional objects. In the TomoGraphView setting, the immediate future directions are explicitly stated: reducing dependence on segmentation masks, evaluating bounding-box-based selection, integrating a detection stage, learning adaptive or task-specific omnidirectional sampling instead of fixed uniform directions, and combining OmniSlicer with future 3D foundation models (Kiechle et al., 12 Nov 2025). The central methodological claim remains narrower and more concrete: omnidirectional planar sampling, when coupled with frozen 2D foundation models and a spherical graph, provides a structured alternative to canonical slice stacks for 3D medical image classification.