2.5D Context-Aware Modeling

• 2.5D context-aware modeling is a computational approach that integrates within-slice and neighboring slice context to efficiently process volumetric data.
• It employs strategies like stacking, attention-based fusion, and early channel fusion to balance 2D network efficiency with essential 3D spatial sensitivity.
• Empirical results demonstrate improvements in segmentation, classification, and thermal modeling, highlighting its broad applicability in medical imaging and engineering.

A 2.5D context-aware model is a class of computational methods and neural network architectures designed to explicitly utilize both within-slice (intra-slice) and near-slice (inter-slice) context from volumetric data, while avoiding the computational and representational burdens of full 3D modeling. The 2.5D approach is prevalent in diverse domains such as medical image analysis, computer-aided diagnosis, chiplet-aware system design, and thermal modeling, where data are inherently volumetric but exhibit discontinuities, significant anisotropy, or efficiency constraints that render naively 2D or 3D strategies suboptimal. In 2.5D models, context is aggregated using mechanisms such as stacking spatially adjacent slices, attention-based fusion across neighboring slices, or early fusion along the channel dimension, marrying the computational efficiency of 2D networks with the spatial sensitivity necessary for accuracy in 3D environments.

1. Definition and Core Principles of 2.5D Context-Aware Modeling

The term “2.5D” denotes models that operate between strict 2D and 3D: they process a target 2D slice (or its features) while incorporating explicit context from a small surrounding neighborhood of adjacent slices, typically through stacking, attention, or context fusion strategies. This approach leverages local 3D spatial cues while minimizing the inefficiencies and data requirements of full 3D convolutions, making it well-suited for data with high in-plane but low through-plane resolution or with significant cross-slice discontinuity (Ou et al., 2021, Kumar et al., 30 Apr 2024).

In medical imaging, this context-awareness captures anatomical continuity and structure that would be missed with pure 2D segmentation, improving delineation at ambiguous or partially observed boundaries (Kumar et al., 30 Apr 2024, Ghouse et al., 15 May 2025, Kim et al., 18 Nov 2025). In chiplet floorplanning or thermal modeling, 2.5D models allow accurate accounting for heat flow or stress distribution across multiple but sparsely interacting layers (Wang et al., 21 Nov 2025, Parekh et al., 29 Apr 2025).

Key architectural motifs for 2.5D modeling include:

• Early-fusion stacking of nearby slices along a designated axis as multi-channel inputs to a 2D CNN (Kim et al., 18 Nov 2025, Ghouse et al., 15 May 2025).
• Hierarchical attention or pooling across both spatial and slice dimensions (Gao et al., 11 Jun 2024, Kumar et al., 30 Apr 2024).
• Pixel-level or feature-level cross-slice attention mechanisms enabling detailed inter-slice correlation (Kumar et al., 30 Apr 2024).
• Hybrid neural–analytical surrogates for physical systems (e.g., thermal models) where context from both local and remote regions is analytically encoded (Wang et al., 21 Nov 2025, Parekh et al., 29 Apr 2025).

2. Representative Architectures and Context Fusion Strategies

2.5D context-aware models exhibit considerable diversity in their design to address specific application challenges:

LambdaUNet (Ou et al., 2021): Implements Lambda+ layers that compute separate linear functions (lambdas) over three disjoint spatial contexts: (i) global intra-slice, (ii) local intra-slice, and (iii) sparse inter-slice (pillar-shaped), then fuse these with pixel-wise queries. This enables accurate segmentation of discontinuous lesions in diffusion-weighted imaging.

CSA-Net (Kumar et al., 30 Apr 2024): Employs two explicit attention modules—cross-slice attention (CSA) and in-slice attention (ISA)—to integrate information for the center slice from adjacent slices and to reason about long-range within-slice structure. Attention is applied at the pixel level, followed by feature aggregation and further processing with a vision transformer.

CARP3D (Gao et al., 11 Jun 2024): Proposes a two-stage attention framework for pathology triage: patch-wise attention identifies salient regions within a slice, while inter-slice attention fuses context-aware information from neighboring slices into the final risk prediction.

MOSAIC (Ghouse et al., 15 May 2025): Builds a 2.5D fused tri-slice representation by concatenating triplets of neighboring slices from axial, coronal, and sagittal views (total of nine channels), followed by separate encoders and cross-attentional fusion across views, supporting organ slice selection and anatomical localization.

2.5D Plane Classifiers (Kim et al., 18 Nov 2025): Use an early-fusion strategy by stacking adjacent slices as channels fed into a conventional 2D CNN backbone (AlexNet or ResNet-18), boosting orientation classification accuracy and enabling downstream uncertainty-aware metadata fusion.

In non-image domains, context-aware 2.5D models include:

• ATMPlace (Wang et al., 21 Nov 2025): Physics-informed surrogate models for wirelength, thermal field, and mechanical warpage, with analytical coupling across chiplets and layers.
• FSA-Heat (Zhang et al., 19 Apr 2025): Neural frequency-spatial domain aware architectures for fast thermal prediction in layered ICs, integrating high-to-low frequency information and spatial context via frequency-domain transformers.

3. Mathematical Formulation and Losses

Common operations in 2.5D context-aware models include:

• Extraction of query, key, and value features from spatial and slice-wise neighborhoods (e.g., Lambda+ layers in (Ou et al., 2021)).
• Attention weight computation via scaled dot-product and softmax normalization, optionally gated or with position-encoding for inter-slice relationships (Kumar et al., 30 Apr 2024, Gao et al., 11 Jun 2024).
• Fusion of context via weighted sums, linear projections, or trainable convolutional blocks.
• Multi-scale cross-feature interaction, e.g., frequency-domain transformers (Zhang et al., 19 Apr 2025) or vision-language fusion with CLIP encoders (Ghouse et al., 15 May 2025).

Loss functions often combine cross-entropy, Dice coefficient, or hybrid (e.g., frequency–spatial) objectives: $$\mathcal{L} = \tfrac12 \mathcal{L}_{\mathrm{CE}} + \tfrac12 \mathcal{L}_{\mathrm{DSC}}$$ (Kumar et al., 30 Apr 2024), or,

$\mathrm{FSL} = L_s + \alpha L_f$

where $L_s$ is spatial MSE and $L_f$ is a frequency-domain loss (Zhang et al., 19 Apr 2025).

In engineering contexts, objectives may further include physically motivated penalties,

$$\mathcal{J}(x, y, \theta) = WL + \lambda_T \left[ T - T_{\text{th}} \right]_{+}^\gamma + \lambda_W \left[ Wpg - W_{\text{th}} \right]_{+}^\gamma$$

(Wang et al., 21 Nov 2025), or stress, thermal, and wirelength joint minimization (Parekh et al., 29 Apr 2025).

4. Evaluation Metrics and Empirical Performance

Empirical evaluations consistently report that 2.5D context-aware models improve over both 2D and 3D baselines in segmentation, classification, triage, and thermal prediction tasks.

• Segmentation (LambdaUNet): Dice scores of 86.5% (vs. 82.2% for 2D UNet and 78.2% for 3D UNet) (Ou et al., 2021).
• Slice Selection (MOSAIC): F1 score of 0.943 and Slice Localization Concordance (SLC) of 0.956, outperforming EfficientNet and Swin-T (Ghouse et al., 15 May 2025).
• Pathology Triage (CARP3D): AUC rises from 81.3% (2D) to 90.4% (2.5D), a ∼9% absolute gain (Gao et al., 11 Jun 2024).
• Thermal Modeling (FSA-Heat): RMSE reduction over 99% versus GCN+PNA, and 4.23× inference speedup (Zhang et al., 19 Apr 2025).
• Floorplanning (STAMP-2.5D, ATMPlace): Achieve up to 20% stress reduction, <1% increase in peak temperature, and 10% reduction in wirelength (Parekh et al., 29 Apr 2025, Wang et al., 21 Nov 2025).
• Plane Orientation (MRI): Accuracy improved from 98.74% to 99.49%, error reduction by 60% (Kim et al., 18 Nov 2025).

Tasks often introduce specialized metrics, such as SLC for slice localization (Ghouse et al., 15 May 2025), in addition to standard segmentation and classification metrics.

5. Computational Advantages and Applicability

2.5D context-aware models offer several computational and practical benefits:

• Efficiency: By processing a small number of slices (typically 3–9) per forward pass, 2.5D models drastically reduce memory and compute compared to volumetric 3D networks, while retaining relevant cross-slice context (Kumar et al., 30 Apr 2024, Kim et al., 18 Nov 2025).
• Flexibility: Incremental modification allows adaptation to volumes of varying depth and anisotropy without architectural overhaul or retraining (Kumar et al., 30 Apr 2024).
• Generalization: The explicit modeling of inter-slice continuity, without enforcing strong 3D priors, provides robustness to modality-specific artifacts such as slice discontinuity, variable layer thickness, or anisotropic resolution (Ou et al., 2021, Zhang et al., 19 Apr 2025).
• Physical interpretability: In chiplet-aware models, the matching of analytical surrogates for thermal and mechanical coupling yields both predictive accuracy and immediate design interpretability (Wang et al., 21 Nov 2025).

Major application domains include: medical imaging (lesion segmentation, plane orientation, multi-organ localization, pathology triage), chiplet-based IC design and floorplanning, and physics-informed surrogate modeling for real-time system monitoring.

6. Limitations, Trade-Offs, and Future Directions

Despite their advantages, 2.5D approaches have important tradeoffs:

• Loss of full 3D modeling: 2.5D models only ingest limited depth context, which may restrict performance in tasks requiring holistic 3D reasoning, particularly with thick slices or highly irregular spatial relations.
• Boundary ambiguity: The definition of context neighborhood (number, ordering, and selection of slices) directly affects the balance between spatial fidelity and computational budget (Kim et al., 18 Nov 2025, Ghouse et al., 15 May 2025).
• Design of context fusion: The effectiveness of attention, stacking, or pooling schemes needs empirical validation for each application, with potentially large gaps between naive early-fusion and advanced attention-based reasoning (Gao et al., 11 Jun 2024, Kumar et al., 30 Apr 2024).
• Domain generalization: Robustness to previously unseen context, slice thickness, or material heterogeneity depends on model design—approaches like frequency-domain fusion (Zhang et al., 19 Apr 2025) and physics-informed surrogates (Wang et al., 21 Nov 2025) yield strong generalization, but may require retraining or further development for dynamic environments.

Future research directions include adaptive or attention-based context neighborhood selection (Kim et al., 18 Nov 2025), multi-fidelity surrogate modeling for efficient physical simulation (Parekh et al., 29 Apr 2025), unsupervised or self-supervised context modeling to reduce annotation cost, and integration with uncertainty-aware inference for downstream clinical or operational decision support (Wang et al., 21 Nov 2025, Kim et al., 18 Nov 2025).