Cascaded Data Consistent Imputation Network

• The paper introduces CDCIN as the ST-specific reconstruction module within S2S-ST, uniquely enforcing hard data-consistency to preserve observed gene-expression values.
• It employs a cascaded predict-and-refine architecture with RDHAN blocks and explicit DC operations to progressively improve imputation quality.
• Empirical evaluations demonstrate that integrating CDC significantly improves MAE, PCC, and SSIM metrics in single-sample spatial transcriptomics reconstruction.

Searching arXiv for the specified paper and any directly relevant contextual references. Cascaded Data Consistent Imputation Network (CDCIN) is the spatial-transcriptomics-specific reconstruction module within the S2S-ST framework introduced in "Sparser2Sparse: Single-shot Sparser-to-Sparse Learning for Spatial Transcriptomics Imputation with Natural Image Co-learning" (Fang et al., 22 Jul 2025). It is designed for a setting in which only a single, sparsely sampled high-resolution spatial transcriptomics (ST) sample is available, and the model must learn from that sample itself rather than from a large external ST corpus. CDCIN is intended not merely to super-resolve sparse ST maps, but to preserve measured gene-expression values exactly while progressively inferring missing values. Its defining characteristic is a cascaded restoration backbone in which every refinement stage is coupled to an explicit data-consistency operation that clamps observed entries to the true sampled measurements.

1. Placement within the S2S-ST framework

CDCIN is the ST-specific reconstruction module inside S2S-ST. The broader framework is presented as a single-shot sparser-to-sparse learning system for ST imputation that uses a single low-cost sparsely sampled ST dataset together with natural images for co-training (Fang et al., 22 Jul 2025). Within that system, CDCIN provides the reconstruction mechanism that maps sparse inputs to denser estimates while remaining constrained by the actual sampled data.

The motivation for CDCIN is tied directly to the target regime. The paper assumes only a single, sparsely sampled high-resolution ST sample is available. In that regime, a plain image-restoration network would tend to overfit or hallucinate values that conflict with observed spots. CDCIN is therefore formulated as a biologically constrained restoration backbone, inspired by cascaded image restoration ideas, but modified by an explicit data-consistency projection that reintroduces measured ST values at every refinement step.

The paper states that this design is what makes the self-supervised training setup viable in the unusually sparse single-sample setting. A plausible implication is that CDCIN should be understood less as a generic super-resolution model than as a constrained imputation network whose central requirement is fidelity to sampled gene-expression measurements.

2. Cascaded architecture and refinement dynamics

CDCIN is a multi-stage cascaded network with $K$ stages; the implementation default is $K=3$ (Fang et al., 22 Jul 2025). Each stage consists of a Data Consistency (DC) operation followed by a Residual Dense Hybrid Attention Network (RDHAN). The sparse-resolution ST patch $X_m^{st}$ is first upsampled to the higher-resolution grid to obtain the initialization

$\hat{X}^{st}_{h,0} = \mathcal{U}(X^{st}_m).$

The model then performs iterative predict-and-refine updates: $\hat{X}^{st}_{h,0} = \mathcal{U}(X^{st}_{m}),$

$$\hat{X}^{st}_{h,k} = \hat{X}^{st}_{h,k-1} + \text{RDHAN}(\text{DC}(\hat{X}^{st}_{h,k-1}, M_{h}^{st})), \quad k=1,2,\dots,K.$$

This update rule gives each cascade access to a repaired version of the current estimate, after which RDHAN predicts a residual correction that is added to the previous estimate. The paper explicitly characterizes this as a predict-and-refine architecture, in which accuracy and coherence improve progressively over the cascades, while cascade depth mediates a quality-versus-compute trade-off.

The default configuration uses three cascaded DC-RDHAN stages. Each stage contains 8 RDHAB blocks. Each residual dense block (RDB) uses 32 channels, growth rate 32, and 4 convolutional layers. In the hybrid attention block (HAB), the Swin-style window size is 8 and the CAB scaling coefficient is $\alpha=0.01$. The paper further reports that gains improve up to 3 stages and then saturate or slightly decline beyond that, making three cascades the preferred operating point.

3. Data consistency as hard projection

The data-consistency mechanism is the core feature that distinguishes CDCIN from a standard cascade. The paper defines a DC operator that re-injects known observed ST values into the current estimate so that they cannot drift during refinement (Fang et al., 22 Jul 2025). The equation is given as

$$\text{DC}(X^{st}_{h},M^{st}_{h}) = \hat{X}^{st}_{h,k} \odot (1 - M^{st}_{h,DC}) + X^{st}_{h} \odot M^{st}_{h}.$$

The notation is described as somewhat compressed, but the intended meaning is explicit: the current reconstruction is retained only at unobserved locations, while observed locations are overwritten by the true measured values. The binary mask $M_h^{st}$ indicates where the high-resolution ST data are actually sampled, and the complement mask selects missing regions. The paper also describes $M_h^{st}$ as “a binary mask indicating the positions of LR inputs corresponding to their locations in the HR data,” and describes DC as a regularization step that reintroduces known data during intermediate processes, “ensuring that the imputation results remain consistent and preventing information degradation during the cascading process.”

Accordingly, DC functions as a projection onto the affine constraint defined by the observed samples: observed entries are clamped to ground truth, and only missing entries are modifiable by the network. This is the sense in which CDCIN is biologically constrained. The model does not treat the measured ST values as soft guidance; it enforces them at every stage of refinement.

A common misunderstanding is to regard CDCIN as a standard restoration backbone with an auxiliary penalty term. The paper instead presents the DC component as an explicit in-network correction step, not merely as a loss-based regularizer.

4. Self-supervised objectives and shared co-learning

CDCIN is tightly coupled to the self-supervised learning strategy of S2S-ST. The framework constructs two ST views from the same sparse sample: a medium-resolution sparse patch $K=3$0 and a more aggressively downsampled “sparser” patch $K=3$1. Training uses two simultaneous ST tasks:

• sparser-to-sparse reconstruction, where $K=3$2 is used to reconstruct $K=3$3;
• sparse-to-dense consistency, where $K=3$4 is used to reconstruct $K=3$5, but supervision is applied after downsampling the prediction back to the sampled pattern.

The corresponding outputs are written as

$K=3$6

$K=3$7

The ST losses supervise all cascades with increasing weights toward deeper stages: $K=3$8

$K=3$9

$X_m^{st}$0

Here $X_m^{st}$1 denotes the downsampling or subsampling operator under the sampling pattern defined by $X_m^{st}$2. The paper states that CDCIN is the engine that makes these objectives meaningful, because the network learns mappings between sparse resolutions while never violating the observed ST measurements at the clamped positions (Fang et al., 22 Jul 2025).

The same shared reconstruction network $X_m^{st}$3 is also trained on a grayscale natural image (GNI) branch with the same cascade structure and the same sampling masks: $X_m^{st}$4

$X_m^{st}$5

with losses

$X_m^{st}$6

$X_m^{st}$7

$X_m^{st}$8

The total objective is

$X_m^{st}$9

with $\hat{X}^{st}_{h,0} = \mathcal{U}(X^{st}_m).$0.

The paper explicitly states that GNI co-training improves representation learning, while the CDC layers keep the reconstruction biologically grounded. This suggests that CDCIN operates as a shared restoration prior whose domain adaptation to ST is enforced structurally through masking and clamping, rather than through an ST-only parameterization.

5. Training regime, patch construction, and inference procedure

The procedure is single-shot and gene-specific: the model is trained on one gene at a time rather than on full multi-gene panels (Fang et al., 22 Jul 2025). ST patches are cropped at size $\hat{X}^{st}_{h,0} = \mathcal{U}(X^{st}_m).$1 with $\hat{X}^{st}_{h,0} = \mathcal{U}(X^{st}_m).$2, using stride $\hat{X}^{st}_{h,0} = \mathcal{U}(X^{st}_m).$3. Downsampling is performed by selecting the top-left pixel in each $\hat{X}^{st}_{h,0} = \mathcal{U}(X^{st}_m).$4 block. Only spots inside the tissue region are included in the loss.

Training uses Adam for 3000 epochs with an initial learning rate of $\hat{X}^{st}_{h,0} = \mathcal{U}(X^{st}_m).$5 on a single NVIDIA A6000 GPU. The binary sampling masks used for the ST branch are reused in the GNI branch, so the natural-image co-learning is structurally aligned with the ST sampling pattern.

At inference time, the trained network is applied over the whole slide in a sliding-window manner, and overlapping patch predictions are merged by weighted averaging. Because the architecture is cascaded, each pass refines the previous output rather than generating an independent map.

The paper emphasizes a computational trade-off inherent in this sample-specific procedure: because the method retrains per dataset, a full training run takes about six hours. A plausible implication is that CDCIN prioritizes fidelity and sample adaptation over throughput or cross-dataset amortization.

6. Empirical effects, ablations, and stated limitations

The ablation study reported in the paper identifies CDC as a major contributor rather than a cosmetic modification. On TENX94 ERBB2, a base RDHAN without both GNI co-training and CDC yields MAE 0.4226, PCC 0.8216, and SSIM 0.7212. Adding only GNI improves these to 0.4030 / 0.8308 / 0.7380. Adding only CDC improves them to 0.3959 / 0.8426 / 0.7508. Combining both gives 0.3840 / 0.8539 / 0.7613 (Fang et al., 22 Jul 2025).

The paper reports similar patterns on TENX95, TENX96, and TENX97, where CDC consistently improves MAE by about 6–11% over the base, and the full CDCIN gives the best or near-best metrics. A more localized analysis is also reported: on TENX95, CDC alone improved 66.27% of spots and yielded a net MAE gain of 0.0266; on TENX94, GNI co-training improved 54.85% of locations, but the strongest gains occurred when CDC was present.

The cascade-depth ablation shows that MAE decreases as the number of cascades increases up to 3, then saturates. This directly motivates the default three-stage design. The HAB ablation also shows that the hybrid attention block reduces MAE across datasets, although the paper treats this as a supporting architectural refinement rather than the central mechanism.

The authors also identify several limitations. CDCIN is trained gene-by-gene rather than on full multi-gene panels, which simplifies optimization but limits direct multi-gene context modeling. Validation is reported using technical metrics rather than downstream biological analyses. The method currently operates on a selective gene panel. In addition, the DC formulation is written tersely: conceptually it is a hard projection or clamping of observed entries, but the printed equation uses $\hat{X}^{st}_{h,0} = \mathcal{U}(X^{st}_m).$6, $\hat{X}^{st}_{h,0} = \mathcal{U}(X^{st}_m).$7, and $\hat{X}^{st}_{h,0} = \mathcal{U}(X^{st}_m).$8 in a way that presumes observed locations are always copied from the true input.

Taken together, these results position CDCIN as the mechanism that converts S2S-ST from a generic self-supervised restoration strategy into a data-consistent ST imputer. Its cascaded RDHAN backbone supplies iterative refinement capacity, while the DC layer enforces strict preservation of sampled ST values. The paper’s overall conclusion is that this combination enables learning from a single sparse sample, integration of natural-image co-learning, and reconstruction that remains faithful to the original sampled gene-expression pattern.