Local Spatial Consistency Decoder (LSCD)

• LSCD is a method that enforces local spatial coherence by integrating localized convolutional processing with belief-propagation, ensuring smooth and consistent signal recovery.
• It is applied in both high-resolution medical image reconstruction and spatially coupled error-correcting codes, balancing fine-detail preservation with computational efficiency.
• LSCD refines decoded outputs via dedicated post-upsampling modules or semi-global decoding, reducing complexity while maintaining high fidelity in practical signal recovery pipelines.

The Local Spatial Consistency Decoder (LSCD) is an architectural and algorithmic concept that arises in two distinct research domains: high-resolution medical image reconstruction and spatially coupled error-correcting codes with sub-block decoding. Despite divergent domains, in both settings LSCD serves to enforce spatial coherence under local constraints, facilitating improved signal recovery while maintaining locality and computational tractability.

1. Motivation for Local Spatial Consistency

In arbitrary-scale medical slice super-resolution, coordinate-wise decoding via point-wise networks (e.g., MLPs) neglects inter-pixel or inter-voxel correlations, increasing the risk of micro-scale artifacts such as discontinuities or unnatural anatomical features. For medical applications, small inconsistencies can compromise interpretability by introducing spurious anatomy or loss of structural fidelity. LSCD mitigates this risk by injecting a local inductive bias through localized convolutional processing, ensuring that the super-resolved output honors spatial coherence among neighboring volumetric elements (Song et al., 25 Jun 2026).

In the coding-theoretic context, spatially coupled LDPC codes with sub-block locality (SC-LDPCL) partition codewords into sub-blocks that can be decoded in isolation or with the aid of a few neighboring blocks. Here, LSCD refers to the semi-global belief-propagation (BP) decoding mode, which leverages local check nodes and controlled coupling with adjacent sub-blocks to optimize the trade-off between locality, failure probability, and overall code threshold (Ram et al., 2019).

2. Architectural Realization in Medical Image Super-Resolution

Within the Dual-Prior Null-space Learning (DP-NSL) framework for arbitrary medical slice super-resolution, the LSCD is integrated as a dedicated post-upsampling decoder module that refines high-resolution volumetric features generated by the Mixture-of-Splines (MoS) upsampler:

• Input: High-resolution feature field $$F_{\mathrm{SR}} \in\mathbb{R}^{C\times \hat S\times H\times W}$$ from the MoS module.
• Split-transform-merge Block: Channels are divided into five groups—$f_{\mathrm{id}}, f_{\mathrm{conv}}, f_1, f_2, f_3$—and processed in parallel by: identity, standard 3×3×3 convolution, and depthwise convolutions with 3×3×3, 5×5×5, and 7×7×7 kernels, respectively.
• Concatenation: The outputs from each transform are concatenated along the channel axis, and a nonlinearity such as ReLU is applied.
• Stacking and Projection: Multiple such inception-style blocks are stacked, followed by a $1\times1\times1$ convolution to produce the scalar null-space prediction $\hat V_{\mathcal N}$.
• No Batch-norm: Batch normalization is omitted to preserve high-frequency signals.

This architectural design enables the LSCD to aggregate fine-to-coarse neighborhood information, capturing both high-frequency detail and broader anatomical context. By refining only the null-space component (through projection after decoding), LSCD ensures that all generated details are locally coherent yet fundamentally orthogonal to the acquired measurements (Song et al., 25 Jun 2026).

3. Algorithmic Formulation and Semi-Global Decoding in Coding Theory

In the SC-LDPCL framework, LSCD implements semi-global decoding (SGD). Here, the code's protograph is partitioned into $M$ sub-blocks, with controlled numbers of local and coupling checks parameterized by $t$. LSCD (in this context synonymous with SGD) decodes a selected sub-block with possible assistance from $d$ adjacent helper sub-blocks on either side:

• Local Decoding (LD): Each sub-block can be decoded using only its internal local checks, resulting in a BP threshold $\epsilon_m^*$ determined by the local subgraph.
• Semi-Global Decoding (LSCD/SGD): To improve reliability, the LSCD uses BP not only on the target sub-block but also with incoming messages (CC erasures) from up to $d$ neighbors, propagating helper information according to specified density-evolution (DE) recursions.
• Boundary Terms: The coupling with adjacent helpers is formalized as boundary conditions injecting erasure-probabilities $\delta_L$ and $f_{\mathrm{id}}, f_{\mathrm{conv}}, f_1, f_2, f_3$0 into the BP.
• Fixed-point Analysis: Success is decided by the convergence of a coupled fixed-point system derived from the protograph DE equations under the imposed boundary conditions.

The LSCD algorithm achieves substantial complexity reduction compared to full-chain BP decoding, requiring only $f_{\mathrm{id}}, f_{\mathrm{conv}}, f_1, f_2, f_3$1 per targeted sub-block (where $f_{\mathrm{id}}, f_{\mathrm{conv}}, f_1, f_2, f_3$2, $f_{\mathrm{id}}, f_{\mathrm{conv}}, f_1, f_2, f_3$3 is VN degree, $f_{\mathrm{id}}, f_{\mathrm{conv}}, f_1, f_2, f_3$4 the iteration count) (Ram et al., 2019).

4. Implicit Mathematical Role and Loss Formulation

In DP-NSL for medical imaging, LSCD implicitly enforces a local smoothness prior without requiring an explicit loss term. Through overlapping three-dimensional convolutions, LSCD encourages locally consistent null-space representations:

$f_{\mathrm{id}}, f_{\mathrm{conv}}, f_1, f_2, f_3$5

where $f_{\mathrm{id}}, f_{\mathrm{conv}}, f_1, f_2, f_3$6 are learnable convolution weights and $f_{\mathrm{id}}, f_{\mathrm{conv}}, f_1, f_2, f_3$7 is the neighborhood of voxel $f_{\mathrm{id}}, f_{\mathrm{conv}}, f_1, f_2, f_3$8.

The overall loss is the $f_{\mathrm{id}}, f_{\mathrm{conv}}, f_1, f_2, f_3$9 discrepancy between the measurement-consistent super-resolution and the ground-truth high-resolution volume:

$1\times1\times1$0

where $1\times1\times1$1 is the orthogonal projection enforcing measurement consistency.

In the coding setting, the LSCD decoding threshold and success/failure criteria are specified via DE fixed-point analysis, which considers both local and semi-global scenarios (Ram et al., 2019).

5. Empirical Performance and Complexity Trade-Offs

Medical Slice Super-Resolution

Ablation studies demonstrate that LSCD achieves nearly the same PSNR as point-wise MLP decoders but with only one-third of the computational cost. On the colon CT dataset, adding LSCD increases PSNR@×2 from 41.82 to 42.06 dB and SSIM from 0.9810 to 0.9817, representing $1\times1\times1$2 dB. When combined with both MCP and MoS, the full DP-NSL pipeline reaches PSNR@×2 of 42.54 dB and SSIM of 0.9832 (Song et al., 25 Jun 2026).

Method	PSNR@×2	FLOPs (M)	Params (K)
Pixel-wise MLP	42.16	98	214
Slice-wise conv	41.79	52	112
LSCD (ours)	42.05	32	71

LSCD thus offers competitive fidelity with marked efficiency.

SC-LDPCL Decoders

For SC-LDPCL, LSCD (realizing SGD) enables decoding a single sub-block with $1\times1\times1$3 helpers, achieving a significant reduction in decoding complexity (as low as 73% of global complexity with $1\times1\times1$4) and minimal loss in BP threshold (a drop of only 3.7%). The trade-off between locality and global coupling parameter $1\times1\times1$5 allows for tailoring the code to application needs: lower $1\times1\times1$6 for better locality, higher $1\times1\times1$7 for stronger global guarantees. Latency and memory footprint are substantially improved under LSCD/SGD, most of the gain observed with just $1\times1\times1$8 adjacent helpers (Ram et al., 2019).

6. Integration in Broader Signal Recovery Pipelines

In DP-NSL, LSCD is integrated after the MoS-based upsampler and before the measurement-consistent projection:

1. $1\times1\times1$9 3D encoder $\hat V_{\mathcal N}$0.
2. $\hat V_{\mathcal N}$1 MoS $\hat V_{\mathcal N}$2.
3. $\hat V_{\mathcal N}$3 LSCD $\hat V_{\mathcal N}$4.
4. $\hat V_{\mathcal N}$5.
5. $\hat V_{\mathcal N}$6.

In the SC-LDPCL pipeline, LSCD/SGD is invoked as needed, decoding only the sub-block of interest with optional neighbor assistance, while global BP decoding remains available for batch correction or archival reliability.

7. Significance and Application Scope

LSCD operationalizes local spatial coherence within broader probabilistic or deterministic recovery frameworks. In medical imaging, this maintains anatomically plausible super-resolved reconstructions at high computational efficiency, directly addressing the limitations of point-wise decoders in clinical contexts (Song et al., 25 Jun 2026). In coding theory, LSCD enables low-latency, memory-efficient block access suitable for modern storage systems, achieving locality without sacrificing most of the global error-correcting capacity (Ram et al., 2019). The underlying principle of enforcing local consistency by localized computation—in either networked or message-passing architectures—is extensible to a variety of high-dimensional signal processing and communications applications.