Slice Interaction Module (SIM) Overview

• Slice Interaction Module (SIM) is a mechanism that explicitly models relationships between adjacent slices in medical imaging using attention-based techniques.
• It integrates channel, spatial, and slice-relation branches to refine features, offering an efficient alternative to heavy 3D processing.
• SIM improves segmentation accuracy by capturing anatomical continuity and addressing weak signal challenges in peripheral imaging slices.

A Slice Interaction Module (SIM) is an architectural block that explicitly models relationships among slices rather than treating a slice stack as a mere channel concatenation. In the most explicit formulation in the current literature, SIM is a lightweight attention-based module introduced for 2.5D PET/CT segmentation of internal gross tumor volume (IGTV) in lung cancer, where it operates on three consecutive axial slices and refines features through channel attention, spatial attention, and a slice-relation branch (Luo et al., 26 Sep 2025). Closely related mechanisms appear in MRI, OCT, semantic occupancy, and prompt-based segmentation systems, although several of those papers do not use the name “SIM” literally and instead present cross-slice attention, cross-slice feature fusion, or slice-residual attention modules that are functionally equivalent forms of slice interaction (Hung et al., 2022, Kim et al., 25 Mar 2026, Chen et al., 2019).

1. Definition and terminological scope

In volumetric medical imaging, SIM usually denotes a module that takes features from neighboring 2D slices and learns how they should influence one another within a 2.5D pipeline. The defining motivation is that pure 2D processing ignores anatomical continuity, whereas full 3D processing is often costly or poorly matched to anisotropic data. The lung PET/CT formulation defines SIM as a slice-aware interaction unit that, given a 3-slice stack, learns which slices to emphasize, which spatial regions across slices are salient, and how slice-wise features should be refined before deeper encoding and multimodal fusion (Luo et al., 26 Sep 2025).

A common source of ambiguity is that the word slice is not uniform across the literature. In imaging papers, a slice is usually an axial, sagittal, coronal, or B-scan image. In “slice-based learning,” by contrast, a slice is an application-critical subset of the data distribution specified by a slicing function $\lambda_i:\mathcal{X}\rightarrow\{0,1\}$, and the interaction mechanism combines slice experts rather than tomographic planes (Chen et al., 2019). This suggests that SIM is best understood as a broader design pattern: an explicit mechanism for modeling structured interactions among slices, where the meaning of slice depends on the domain.

Setting	Slice notion	Interaction mechanism
Lung PET/CT IGTV (Luo et al., 26 Sep 2025)	Three consecutive axial slices	Channel attention, spatial attention, depthwise slice relation
Prostate MRI (Hung et al., 2022)	Full stack of MRI slices	Cross-slice Transformer attention
OCT retinal layers (Kim et al., 25 Mar 2026)	Neighboring B-scans	$1\times1$ conv + softmax slice fusion
Slice-based learning (Chen et al., 2019)	Critical data subsets	Slice-residual attention over slice experts
3D occupancy (Huang et al., 4 Sep 2025)	Vertical height slices	Global–local cross-attention fusion

2. Canonical SIM design in multimodal PET/CT segmentation

The explicit SIM in lung cancer PET/CT is embedded in a 2.5D segmentation framework built on CIPA, a Cross-modal Interactive Perception Network with Mamba. The overall system uses dual PET and CT encoders with Mamba-based backbones, multimodal fusion through CRM and DCIM, and a decoder that predicts the segmentation map for the central slice $\mathbf{I}_t$. Instead of a single-slice input, the model consumes three consecutive slices per modality,

$$\{\mathbf{I}_{t-1}, \mathbf{I}_t, \mathbf{I}_{t+1}\},$$

so that each modality branch begins from a tensor $$\mathbf{X}^{(mod)} \in \mathbb{R}^{B \times 3 \times H \times W}$$ and the PET+CT input has six channels in total (Luo et al., 26 Sep 2025).

The SIM is positioned in the encoder after initial convolutional feature extraction and before deeper encoding and multimodal fusion. Its purpose is narrowly defined around two coupled difficulties in IGTV segmentation: weak PET intensities in peripheral cranial/caudal slices, and the inadequacy of naïve 2.5D channel stacking for learning how slice $t$ should interact with slices $t-1$ and $t+1$ (Luo et al., 26 Sep 2025). That design target is important: in this formulation, SIM is not merely an attention add-on, but a replacement for implicit cross-slice coupling by explicit slice-aware modeling.

Technically, SIM takes $$\mathbf{X} \in \mathbb{R}^{B \times C \times H \times W}$$ and decomposes the refinement into three branches. The channel-attention branch uses a squeeze–excitation style mechanism,

$$\mathbf{A}_{ch} = \sigma\big(\mathbf{W}_2 \cdot \text{ReLU}(\mathbf{W}_1 \cdot \text{GAP}(\mathbf{X}))\big), \qquad \mathbf{X}_{ca} = \mathbf{X} \odot \mathbf{A}_{ch},$$

with reduction ratio $1\times1$0. The spatial-attention branch computes a shared spatial mask on the stacked tensor,

$1\times1$1

The slice-relation branch applies depthwise $1\times1$2 and $1\times1$3 convolutions,

$1\times1$4

The three outputs are fused through a weighted residual connection,

$1\times1$5

with $1\times1$6, $1\times1$7, and $1\times1$8 (Luo et al., 26 Sep 2025).

A central conceptual point is that inter-slice modeling is achieved without heavy 3D convolutions. Channel attention operates directly over slice channels, spatial attention is computed from the stacked three-slice representation, and the depthwise branch performs slice-specific refinement conditioned by those earlier attention signals. The PET/CT paper explicitly frames this as consistent with a 2.5D philosophy: use a small number of slices, exploit their relationships through attention and stacked representations, and avoid the memory burden and data demands of full 3D processing (Luo et al., 26 Sep 2025).

3. Alternative SIM formulations across domains

Cross-slice interaction has been implemented in several distinct ways. In CAT-Net for prostate zonal segmentation, the analogous module is a cross-slice attention Transformer inserted on every skip connection of a 2D encoder–decoder. If $1\times1$9 denotes the stack of slice features at one scale, pooled queries and keys are used to build an $\mathbf{I}_t$0 attention matrix

$\mathbf{I}_t$1

followed by cross-slice aggregation

$\mathbf{I}_t$2

This design allows each slice to attend to every other slice in the volume, rather than only its immediate neighbors, and was introduced specifically to improve difficult apex and base slices in anisotropic prostate MRI (Hung et al., 2022).

In OCT retinal layer segmentation, the Cross-slice Feature Fusion module provides a different SIM archetype. At each encoder level, feature maps from $\mathbf{I}_t$3 consecutive B-scans are concatenated, a $\mathbf{I}_t$4 convolution predicts slice-specific weights, and a softmax across the slice axis yields normalized spatially varying slice weights. The fused representation is

$\mathbf{I}_t$5

This combines attention-style weighted fusion with a residual average that stabilizes training and prevents collapse to extreme slice selection (Kim et al., 25 Mar 2026). A plausible implication is that SIMs need not rely on self-attention or Transformers; explicit slice weighting with 2D operators can suffice when local continuity is the primary requirement.

CSAM defines yet another 2.5D pattern for anisotropic volumetric segmentation: semantic, positional, and slice attention on deep feature maps at different scales, with the stated goal of capturing information across all slices in the volume while keeping trainable parameters minimal (Hung et al., 2023). Slide-SAM adopts a simpler but still recognizably slice-interactive approach: a window of three adjacent slices is treated as a three-channel input, prompts are placed only on the central slice, masks are predicted for all three slices, and the predicted masks at the top and bottom are converted into prompts for adjacent windows during sliding inference (Quan et al., 2023). In 3D semantic occupancy, SliceSemOcc uses vertical slices along the height axis rather than axial medical slices; global and local vertical slice features interact through bidirectional cross-attention, and SEAttention3D assigns channel weights per height layer rather than uniformly across height (Huang et al., 4 Sep 2025).

The non-imaging lineage in slice-based learning is structurally related but semantically different. There, a slice-residual-attention mechanism combines slice expert representations $\mathbf{I}_t$6 with attention weights

$\mathbf{I}_t$7

where $\mathbf{I}_t$8 are learned slice-indicator logits and $\mathbf{I}_t$9 are slice prediction logits (Chen et al., 2019). This formulation preserves the core SIM idea—explicitly weighting specialized slice representations—but applies it to critical subsets of a data distribution rather than neighboring image planes.

4. Training regimes and integration strategies

The most detailed training protocol for an explicit SIM appears in the lung PET/CT study. Training is divided into two stages. First, a 2D CIPA network is pre-trained on 21,930 PET/CT pairs with GTV labels using

$$\{\mathbf{I}_{t-1}, \mathbf{I}_t, \mathbf{I}_{t+1}\},$$0

Second, the pre-trained model is converted to 2.5D by changing the input to three consecutive slices and integrating SIM into the encoder, then fine-tuned on the private IGTV cohort from the PET/CT subset of LUCID with

$$\{\mathbf{I}_{t-1}, \mathbf{I}_t, \mathbf{I}_{t+1}\},$$1

All inherited CIPA weights are initialized from the GTV model, while SIM and new 2.5D-specific layers are randomly initialized; the fine-tuning learning rate is $$\{\mathbf{I}_{t-1}, \mathbf{I}_t, \mathbf{I}_{t+1}\},$$2 (Luo et al., 26 Sep 2025). This suggests a practical pattern for SIM deployment: attach the interaction block to a strong pre-trained 2D backbone, then learn slice interaction with a smaller task-specific dataset.

Other papers integrate slice interaction at different architectural locations. CAT-Net inserts cross-slice modules on skip connections at multiple scales and uses learnable slice-wise positional encodings initialized by sin–cos functions, with Adam at learning rate $$\{\mathbf{I}_{t-1}, \mathbf{I}_t, \mathbf{I}_{t+1}\},$$3, weight decay $$\{\mathbf{I}_{t-1}, \mathbf{I}_t, \mathbf{I}_{t+1}\},$$4, and 150 epochs (Hung et al., 2022). The OCT CFF module replaces each skip connection in an FCRN-like encoder–decoder, keeps the backbone fully 2D, and is trained with a combined loss over pixel-wise masks, column-wise boundary positions, and smooth $$\{\mathbf{I}_{t-1}, \mathbf{I}_t, \mathbf{I}_{t+1}\},$$5 surface regression (Kim et al., 25 Mar 2026). Slide-SAM, by contrast, integrates slice interaction into a promptable segmentation pipeline by combining a three-slice input window with a SAM image encoder, prompt encoder, and slice-specific MLP heads, and trains with a hybrid cross-entropy, Dice, and IoU-based objective (Quan et al., 2023).

Across these training strategies, SIM is not tied to one optimization protocol. It appears as a modular augmentation: an encoder block, a skip-connection block, a decoder-side fusion mechanism, or a prompt-conditioned multi-slice head.

5. Empirical effects and ablation evidence

The clearest quantitative evidence for an explicit SIM comes from IGTV segmentation in lung PET/CT. On the LUCID-PET/CT test set, the baseline 2D model without fine-tuning achieved IoU $$\{\mathbf{I}_{t-1}, \mathbf{I}_t, \mathbf{I}_{t+1}\},$$6, Dice $$\{\mathbf{I}_{t-1}, \mathbf{I}_t, \mathbf{I}_{t+1}\},$$7, Acc $$\{\mathbf{I}_{t-1}, \mathbf{I}_t, \mathbf{I}_{t+1}\},$$8, and HD95 $$\{\mathbf{I}_{t-1}, \mathbf{I}_t, \mathbf{I}_{t+1}\},$$9. Naïve 2.5D without fine-tuning changed performance only marginally to IoU $$\mathbf{X}^{(mod)} \in \mathbb{R}^{B \times 3 \times H \times W}$$0, Dice $$\mathbf{X}^{(mod)} \in \mathbb{R}^{B \times 3 \times H \times W}$$1, Acc $$\mathbf{X}^{(mod)} \in \mathbb{R}^{B \times 3 \times H \times W}$$2, and HD95 $$\mathbf{X}^{(mod)} \in \mathbb{R}^{B \times 3 \times H \times W}$$3. After fine-tuning, baseline 2.5D improved to IoU $$\mathbf{X}^{(mod)} \in \mathbb{R}^{B \times 3 \times H \times W}$$4, Dice $$\mathbf{X}^{(mod)} \in \mathbb{R}^{B \times 3 \times H \times W}$$5, Acc $$\mathbf{X}^{(mod)} \in \mathbb{R}^{B \times 3 \times H \times W}$$6, and HD95 $$\mathbf{X}^{(mod)} \in \mathbb{R}^{B \times 3 \times H \times W}$$7. Adding SIM yielded IoU $$\mathbf{X}^{(mod)} \in \mathbb{R}^{B \times 3 \times H \times W}$$8, Dice $$\mathbf{X}^{(mod)} \in \mathbb{R}^{B \times 3 \times H \times W}$$9, Acc $t$0, and HD95 $t$1, while CBAM reached Dice $t$2 and a non-local block Dice $t$3 (Luo et al., 26 Sep 2025). A common misconception—that stacking adjacent slices already constitutes effective inter-slice modeling—is directly contradicted by this ablation, where Dice changes only from $t$4 to $t$5 before explicit interaction is introduced.

The PET/CT paper further reports qualitative effects tightly aligned with the intended mechanism: SIM-based predictions better delineate weak PET boundary slices, are less fooled by adjacent high-uptake organs, and produce more complete and consistent IGTV volumes at cranial and caudal margins (Luo et al., 26 Sep 2025). Because the main failure mode motivating SIM was low-SUV peripheral ambiguity, this is more informative than overall Dice alone.

Independent evidence from other domains points in the same direction. CAT-Net improved prostate zonal segmentation over 2D, 2.5D, and 3D baselines; on the internal dataset with an nnU-Net backbone, CAT-nnU-Net achieved TZ Dice $t$6 and PZ Dice $t$7, compared with $t$8/$t$9 for 2D nnU-Net, $t-1$0/$t-1$1 for the best 2.5D nnU-Net, and $t-1$2/$t-1$3 for 3D U-Net (Hung et al., 2022). In OCT, adding the CFF module to FCRN reduced average MAD from $t-1$4 to $t-1$5 and average RMSE from $t-1$6 to $t-1$7 on the clinical dataset, and from $t-1$8 to $t-1$9 MAD and $t+1$0 to $t+1$1 RMSE on the DUKE DME dataset (Kim et al., 25 Mar 2026). In 3D semantic occupancy, global+local vertical slices with cross-attention achieved $t+1$2 mIoU, versus $t+1$3 for simple concatenation and $t+1$4 with no slicing (Huang et al., 4 Sep 2025). These results do not prove a single best SIM design, but they consistently indicate that explicit slice interaction is more effective than purely implicit fusion.

6. Limitations, assumptions, and broader significance

SIM designs inherit several recurring constraints. The explicit lung PET/CT SIM uses a fixed three-slice window and therefore captures only local context; the authors note that this may miss long-range relationships for tumors extending over many slices and suggest larger or adaptive slice windows or full 3D architectures for long-range context (Luo et al., 26 Sep 2025). The same study also notes sensitivity to slice thickness, scan consistency, PET motion blur, registration quality, and correct ordering of slices, because any slice-based interaction mechanism presumes meaningful adjacency and ordering (Luo et al., 26 Sep 2025).

Other SIM-like systems expose complementary trade-offs. CAT-Net obtains global slice-axis attention but with substantially increased parameters and memory relative to a plain nnU-Net (Hung et al., 2022). SliceSemOcc adds approximately $t+1$5 GB of GPU memory in validation because of the extra global and local vertical paths and cross-attention (Huang et al., 4 Sep 2025). Slide-SAM is limited by its local three-slice window and shows decreasing Dice as z-spacing increases, indicating that a fixed narrow context can become inadequate when neighboring slices differ strongly (Quan et al., 2023). In OCT, the CFF framework is explicitly positioned as a compromise between 2D inconsistency and 3D computational burden rather than a substitute for full volumetric reasoning (Kim et al., 25 Mar 2026).

The broader significance of SIM is therefore methodological rather than purely definitional. Across the cited literature, SIM-like modules are used when the slice axis contains indispensable but structured context: ambiguous prostate apex and base slices, weak peripheral PET boundaries, noisy OCT B-scans, vertically stratified occupancy patterns, or application-critical data subsets in non-imaging learning. The unifying principle is explicit slice-aware interaction. Whether implemented through attention over slice channels, slice-wise softmax fusion, cross-slice Transformers, vertical slice cross-attention, or residual mixing over slice experts, SIM marks a departure from the assumption that stacked slices can be treated as ordinary channels without dedicated relational modeling (Luo et al., 26 Sep 2025, Chen et al., 2019).