2.5D Architectures for Efficient Volumetric Medical Imaging

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2217 papers294 conceptsfrozen 2026-10-04

Medical imaging data, such as CT and MRI scans, are inherently three-dimensional. However, researchers in medical AI face a persistent "dimensionality dilemma." Processing these volumes as independent 2D slices is computationally efficient but "blind" to the vertical axis, leading to anatomical inconsistencies and "flickering" artifacts. Conversely, full 3D convolutional neural networks (CNNs) capture the entire spatial context but are notoriously memory-intensive and slow, often forcing researchers to sacrifice network depth or image resolution.

The research direction emerging across these 169 papers is the 2.5D architecture—a strategic middle ground that captures volumetric context while maintaining the efficiency of 2D processing.

The Dimensionality Gap

The core motivation for 2.5D approaches is the trade-off between spatial awareness and computational feasibility. Matthew Lai notes that while 3D convolutions offer the highest theoretical accuracy, they are often impractical for clinical deployment. Adam P. Harrison and Dakai Jin argue that 2D-based processing allows for significantly deeper network architectures and broader spatial context within the memory constraints of standard hardware. 2.5D methods bridge this gap by providing "just enough" 3D information to resolve ambiguities without the overhead of a full volumetric model.

Pseudo-3D via Channel Stacking

The simplest way to implement 2.5D is to treat the depth dimension (the Z-axis) like color channels in a standard RGB image. Instead of feeding a single slice into a network, researchers stack a target slice with its immediate cranial and caudal neighbors (e.g., slices $t-1, t, t+1$).

Yunan Wu and Jie Deng use this "pseudo-3D" approach for intracranial hemorrhage detection, allowing the model to see the continuity of a bleed across adjacent layers. Similarly, Xiao Han demonstrated in the Liver Tumor Segmentation (LiTS) challenge that a small stack of adjacent slices provides sufficient volumetric context to segment complex 3D structures accurately.

Feature-Level Slice Interaction

More advanced 2.5D frameworks move beyond simple stacking to integrate context at the feature level. Xinwei Xue and Gaoyu Wang introduced the Adjacent Slice Feature Fusion (ASF) module, which uses attention mechanisms to weigh the importance of neighboring features. This allows the network to selectively focus on inter-slice correlations that are most relevant to the target organ or lesion.

In cardiac imaging, researchers have employed parallel contracting paths to process adjacent slices simultaneously. This design helps the model distinguish between anatomically similar structures that appear ambiguous in an isolated 2D view but distinct when their 3D neighbors are considered.

Axial Decomposition and Sequential Modeling

A third facet of this research direction involves decomposing 3D operations into 2D components to save memory. Xiangyi Yan and Hao Tang 0010 developed AFTer-UNet, which splits 3D self-attention into separate intra-slice (spatial) and inter-slice (axial) operations. This "axial fusion" captures vital volumetric context without the prohibitive memory costs of standard 3D Transformers.

Alternatively, Jinkui Hao treats the stack of CT slices as a time sequence. By using a memory-capable ConvLSTM module in the SLP-Net architecture, the system maintains a "memory" of anatomical structures as it moves through the volume, ensuring spatial continuity along the axial dimension.

Why Local Context Suffices

The success of 2.5D architectures—from brain metastasis detection by Endre Grøvik to mitochondria segmentation by Zhengrong Luo and Shikun Liu—rests on the intuition that medical images exhibit high local correlation along the Z-axis.

Because anatomical structures do not change abruptly between 1mm or 5mm slices, the information contained in a small window of neighbors is usually enough to resolve 3D ambiguities, such as branching blood vessels or elongated tumors. By capturing this local volumetric context, 2.5D models achieve a "sweet spot": they provide the spatial awareness required for clinical accuracy while remaining lightweight enough for real-time deployment on standard hospital hardware.

Papers in this version

Showing 60 of 217 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 151 papers
  2. Version 1 · 2026-10-04 · 217 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_pseudo_3d_spatial_context_22f85fdf,
  title = {pseudo-3D spatial context},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/pseudo-3d-spatial-context/rdr_22f85fdf8e5c4cb0a6f382a8a24388ed},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_2_5d_architectures_for_efficient_volumet_22f85fdf_v1,
  title = {2.5D Architectures for Efficient Volumetric Medical Imaging},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/2-5d-architectures-for-efficient-volumetric-medical-imaging/rdr_22f85fdf8e5c4cb0a6f382a8a24388ed/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}