{"id":"6010d0d3-582d-44d7-a562-174f25a2b819","arxiv_id":"2505.10729","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"C2-STi generates missing spatial transcriptomics slices at arbitrary positions between two measured slices, using H&E images, gene co-expression graphs, and deformable convolution, and reports higher PSNR, SSIM, and PCC than five baselines on the HNSCC dataset.","lead":"This paper presents C2-STi, a deep learning method that fills in missing spatial transcriptomics slices using neighboring slices and stained tissue images. If it works, researchers could obtain 3D gene-expression views of tissue without profiling every physical slice.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 2.3 requires the gradient of the missing H&E slice to set p_i, but that slice is not an input; the arbitrary-position mechanism is either oracle-dependent or under-specified.","rationale":"The paper's headline claim is that C2-STi is the first method to interpolate ST slices at arbitrary intermediate positions from adjacent slices. The only mechanism that makes positions arbitrary is DLSM, and its position p_i is computed from the gradient of the missing H&E slice. This is a stronger and more immediate problem than the reader's biological-validity concern: even if H&E gradient magnitude were a perfect deformation proxy, the required H&E slice is not among the stated inputs. The issue is internal inconsistency, not disagreement with community consensus. I am not accusing the authors of deliberate cheating; the code may reveal a straightforward implementation and the text may simply be sloppy. That is why the concrete test is to inspect and rerun the publicly available code. The flaw sits at the exact point that distinguishes arbitrary-position interpolation from fixed-midpoint interpolation, so it is load-bearing. The paper has useful components and provides a code link, which is a real asset, but the current manuscript does not establish the headline claim. If the code confirms the use of H_i, the comparisons are against methods that do not have access to the target tissue, and 'first algorithm' is unsupported. If the code does not use H_i, the paper must be rewritten to specify the actual position encoding; the observed results could still stand after clarification. The numeric mismatch for P (s-1 positions instead of s, denominator i/s rather than i/(s+1)) reinforces that this part of the method was not carefully specified. Therefore I would reject the current version rather than accept it conditionally, because the central mechanism as written is not implementable in the stated setting; a revised manuscript that resolves the ambiguity could be reconsidered.","tokens_in":8750,"tokens_out":11864,"duration_ms":118457,"concrete_test":"Download the public repository (github.com/XiaofeiWang2018/C2-STi) and trace the computation of w_i and p_i in the DLSM module. Determine whether the forward pass accesses the intermediate H&E image H_i for the target section. If it does, rerun the HNSCC single- and multi-slice experiments with p_i obtained only from H0 and H1, e.g., p_i = i/(s+1) or with w_i estimated from interpolated H0/H1 gradients. If the reported PSNR, SSIM, PCC, and RMSE gaps in Tables 1 and 2 shrink materially, the published comparisons rely on target-slice information. If the code does not use H_i, check whether the formula in the paper matches the code; if it does not, the paper's description must be revised before the 'arbitrary position' claim can be accepted.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that C2-STi can interpolate an arbitrary number of missing ST slices from two adjacent slices. That claim rests on the DLSM position encoding in Section 2.3, where p_i = (i/s) * w_i and w_i = Norm(1 + alpha * |grad H_i|). This formula requires the gradient of the i-th H&E slice, i.e., the H&E image of the missing intermediate section. The stated model inputs are only the adjacent ST slices I0, I1 and their paired H&E images H0, H1; no mechanism is given for obtaining H_i at inference time. As written, the method cannot compute p_i for a genuinely missing slice. If the implementation actually loads the intermediate H&E image, then the model uses target-slice information, so the claimed 'interpolate missing ST slices from adjacent slices' setting is not what is being evaluated. If H_i is not used, the paper omits how p_i is derived from H0 and H1 alone. Either way, the arbitrary-position mechanism is not well-defined. There is also an internal numeric inconsistency: P is defined only for i = 1, ..., s-1, giving s-1 positions while the task states s intermediate slices, and p_i = i/s does not place s slices between endpoints at 0 and 1 (i/(s+1) would). These issues prevent the central claim from being evaluated as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes C2-STi, a deep learning method for interpolating missing spatial transcriptomics (ST) slices between two adjacent ST slices, using the paired hematoxylin and eosin (H&E) images as additional input. The method combines a cross-modal alignment module, a pyramid gene co-expression correlation module, and a distance-aware local structural modulation (DLSM) module intended to enable interpolation at arbitrary intermediate positions. On the HNSCC dataset, the authors report higher PSNR, SSIM, and PCC and lower RMSE than five baselines for single-slice interpolation and for 2-, 3-, and 4-slice interpolation, and they provide ablation studies for each of the three proposed modules.","tokens_in":9092,"tokens_out":4068,"duration_ms":39353,"significance":"If the claims are correct, C2-STi would be a practically valuable contribution to spatial transcriptomics analysis, potentially reducing the cost of dense slicing by generating missing ST slices from adjacent pairs. The consistent margin over five baselines, the availability of code, and the ablation experiments are notable strengths, and I found no circularity: the model is trained on training patches and evaluated on held-out test patches against external baselines. However, the evidence base is a single dataset with a single split and no error bars or significance tests, and the arbitrary-position mechanism is under-specified. The significance of the claimed contribution is therefore conditional on resolving the methodological issues described below.","major_comments":[{"comment":"The position encoding is defined as p_i = (i/s) * w_i with w_i = Norm(1 + alpha * |\\nabla H_i|), which requires the gradient of the H&E image of the missing i-th slice. However, the inputs listed in Section 2 are only the adjacent ST slices I_0, I_1 and their paired H&E images H_0, H_1; no procedure is given for obtaining \\nabla H_i from H_0 and H_1 alone. If the intermediate H&E image is available at inference time, then the claimed setting of interpolating missing slices from adjacent slices is violated and target information leaks into p_i; if it is not available, the module is undefined. This issue blocks evaluation of the central claim of arbitrary-position interpolation.","section":"Section 2.3, Cross-section distance modeling"},{"comment":"The text states that the model generates s imputed intermediate ST slices, but the position set is defined as P = {p_i | p_i = (i/s) * w_i, i values from 1 to s-1}, which contains only s-1 positions. Moreover, with i ranging from 1 to s-1, the values i/s produce positions 1/s, ..., (s-1)/s, which does not correspond to s intermediate locations between the endpoints 0 and 1; a spacing of i/(s+1) would be the natural choice. This internal inconsistency makes it unclear how the s output slices are generated and how the loss terms in Section 2.4 are computed.","section":"Section 2.3, position set P"},{"comment":"All multi-slice experiments use s = 2, 3, or 4 evenly spaced intermediate slices, but no experiment varies the relative positions of the missing slices. The claim that C2-STi supports interpolation at arbitrary positions is therefore not empirically demonstrated. In addition, the paper reports results on a single split of a single dataset without error bars or significance tests, and several margins over baselines are small (e.g., SSIM 0.78 vs. 0.77 in the 2-slice row of Table 2), so the robustness of the headline empirical claim is unclear.","section":"Section 3.1, Table 2"},{"comment":"The GCN update C_{L,graph}^0 = delta(P C_L^0 W_n) uses a propagation operator P that is never defined. It should be stated explicitly (for example, as a normalized adjacency matrix derived from the correlation matrix A in Eq. (2)) for the method to be reproducible and for the role of A to be understood.","section":"Section 2.2, Eq. (3)"}],"minor_comments":[{"comment":"There is a duplicated word in 'to to quantify biological structure similarity', and the sentence defining d_{i,j} = (s+1)Delta_d is confusing because the right-hand side does not depend on i and j; the distance between slices I_i and I_j should presumably be (j-i)Delta_d.","section":"Section 2.3, first paragraph"},{"comment":"The dataset description says '19 consecutive slices spanning 350 mm'; this is likely a unit error and should be micrometers, since 350 mm would be anatomically implausible for HNSCC tissue sections.","section":"Section 3, Dataset"},{"comment":"The description of the MGC-Graph construction is ambiguous: it says the output features C_L^0 and C_L^1 are used as input nodes, but it is not clear how individual genes are identified as nodes in a graph of size N, nor how the same adjacency A is applied to both the forward and backward branches.","section":"Section 2.2, MGC-Graph"},{"comment":"The experimental protocol for the baselines is underdescribed; please state whether the same training/validation split, the same number of input gene channels, and the same optimizer settings were used for RIFE, RBI, DAIN, IFRNet, and U-Net, and clarify how image-quality metrics are computed over genes.","section":"Section 3.1, Table 1"},{"comment":"Several references cite only arXiv versions without DOIs; adding published versions where available would improve the reference list.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The core empirical finding is a consistent improvement over five baselines on one dataset, and the code is promised to be available. The major concern is that the arbitrary-position mechanism in Section 2.3, which is central to the claimed contribution, is either under-specified or potentially uses the target slice's H&E gradient. This is fixable in revision if the authors clarify the actual implementation and add experiments with nonuniform positions, but as written it prevents evaluation of the central claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: this is a plausible engineering contribution with a new task, but the paper needs a significant fix before the arbitrary-position claim is testable.\n\nWhat is new: the task of generating multiple spatial transcriptomics slices between two measured slices, with position flexibility, is not in the prior work cited. Monjo et al. only does the middle slice. The architecture borrows deformable convolution, GCN, and gated attention, but the packaging for ST interpolation is new. They report consistent improvements over five baselines on HNSCC, and they ship code.\n\nCredit: the experimental setup is standard, the metrics are clear, and the ablations show all three modules contribute. The baseline comparison is reasonable for the field.\n\nSoft spots: the biggest one is in Section 2.3. The position encoding p_i = (i/s)*w_i uses w_i = Norm(1 + alpha |grad H_i|), which requires the H&E gradient at the missing slice. The declared inputs are only H0 and H1. There is no description of how H_i is synthesized or approximated. Either the method uses the unseen slice's H&E (making it not a purely adjacent-slice interpolation), or that part is under-specified. This is load-bearing for the central claim. Also, P has s-1 positions while the task states s intermediate slices, and p_i = i/s does not place s positions between 0 and 1. These are small edits, but they need to be made.\n\nOther soft spots: only one dataset, no error bars or significance tests; the \"arbitrary positions\" are actually uniform 2/3/4-slice settings. The GCN equation (3) is also a bit opaque. None of these are fatal on their own, but together they mean the headline \"first to interpolate arbitrary number\" is not yet substantiated.\n\nWho it's for: people working on ST reconstruction or multimodal medical image interpolation. It deserves a serious referee, because the task is real and the baselines are standard. I would send it back for major revision rather than desk-reject.\n\nRecommendation: engage, but require the authors to fix the H_i dependence, correct the indexing, add variance estimates, and test a non-uniform position set.","headline":"Genuinely new ST interpolation task, but the arbitrary-position mechanism is under-specified in a load-bearing way and the single-dataset evidence needs error bars.","tokens_in":9580,"tokens_out":2939,"would_cite":true,"duration_ms":29976,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Two adjacent gene-expression tissue maps and their paired stained images suffice to synthesize any missing slice in between.","keywords":["spatial transcriptomics","slice interpolation","cross-modal alignment","gene co-expression graph","deformable convolution","H&E-stained images","arbitrary-position interpolation","deep learning"],"falsifier":"Hold out real middle sections from HNSCC triples, train the model once with the proposed gradient-weighted position and once with alpha=0, and compare per-gene PSNR and Pearson correlation; if the gradient-weighted version is not consistently better on slice pairs with large H&E edge magnitudes, the distance model's core assumption fails.","tokens_in":8562,"feed_emoji":"🧬","tokens_out":10287,"duration_ms":93822,"temperature":0.7,"pith_summary":"The paper proposes C2-STi, a deep network that takes two adjacent spatial transcriptomics (ST) slices and their paired hematoxylin-and-eosin (H&E) images and generates the ST slices that would lie at any requested intermediate position. The authors claim this is the first method to interpolate an arbitrary number of missing sections between adjacent ST slices, rather than only a single middle slice. The network couples three mechanisms: a cross-modal alignment gate that lets ST features borrow cellular detail from H&E, a pyramid graph that tracks co-expressed genes across scales, and a distance-aware modulation that turns a slice-position scalar into deformable-convolution adjustments. On the public HNSCC dataset they report consistently higher PSNR, SSIM, and Pearson correlation and lower RMSE than the tested frame-interpolation, medical-interpolation, and U-Net baselines in both single-slice and multi-slice settings. If the claim holds, a laboratory could produce dense three-dimensional gene-expression volumes from a few physically sectioned and stained slices.","feed_headline":"Fills missing gene-expression slices at any position","feed_subtitle":"Pairs of adjacent spatial transcriptomics maps plus H&E images yield any requested intermediate slice.","key_machinery":"Distance-aware local structural modulation (DLSM) is the mechanism that makes arbitrary-position interpolation possible. It expresses each requested slice as a scalar $p_i = (i/s)\\cdot\\mathrm{Norm}(1+\\alpha|\\nabla H_i|)$, where the Sobel gradient of the H&E image reweights the plain slice index; that scalar is then embedded and split into a channel-adaptation branch and a spatial-adaptation branch whose outputs predict kernels, offsets, and masks for a deformable convolution. The deformable convolution reshapes the coarse forward and backward deformation features into features at position $p_i$, so one network can generate any intermediate slice rather than only the midpoint. The cross-modal alignment gate and the multi-gene co-expression graph are supporting modules: the former filters H&E features with gated attention before they touch ST features, and the latter defines a graph whose adjacency is Pearson correlation between gene-expression profiles and applies a graph convolution at each pyramid level.","core_discovery":"C2-STi is presented as the first algorithm to interpolate missing spatial transcriptomics (ST) slices at arbitrary intermediate positions between two adjacent ST slices. Given the two slices I0 and I1 and their paired H&E images H0 and H1, it outputs any requested number s of imputed slices {Iout^i}. Three modules carry the work: a cross-modal alignment gate that refines ST features with H&E-derived cellular structure; a pyramid gene co-expression correlation module that builds a Pearson-correlation adjacency matrix over genes and propagates it through a graph-convolution layer at three scales; and a distance-aware local structural modulation module that embeds a position scalar $p_i=(i/s)\\mathrm{Norm}(1+\\alpha|\\nabla H_i|)$ into channel and spatial branches, producing kernels, offsets, and masks for deformable convolution. The network is trained with an L1 similarity loss against the ground-truth slice plus a gradient smoothness regularizer. On the HNSCC dataset the reported single-slice results beat the compared methods by at least 6.0% PSNR, 6.6% SSIM, 29.0% PCC, and 21.4% RMSE, with consistent improvements across 2-, 3-, and 4-slice interpolation, and the ablations show all three modules contribute.","pith_inferences":["A testable extension the paper does not run is applying the same position-scalar idea to temporal ST experiments, treating time between two observations as the interpolation coordinate; the gradient weight would then need to be a temporal-change proxy rather than an H&E edge map.","If the distance model works as claimed, interpolation error on a held-out middle section should correlate with the H&E gradient magnitude: sections with stronger edge structure should require larger learned offsets, a prediction one could check per region.","Since the paper evaluates on a single cancer dataset, the natural next question is cross-dataset transfer to a different tissue type; the authors do not claim such transfer."],"forward_implications":["A laboratory with only two successfully collected adjacent ST slices could reconstruct any number of intermediate expression maps in a single forward pass, replacing repeated high-cost sectioning for 3D analysis.","Output slices can be requested at arbitrary fractional positions, so a user can choose the density of the reconstructed volume after the model is trained.","The ablations show that removing the H&E cross-modal alignment hurts most, indicating that histology information is the largest contributor to the reported gains.","Interpolation quality is reported with standard image metrics, so downstream quality control could screen synthesized slices before using them in analysis."],"supporting_citations":[{"why":"supplies the HNSCC dataset of aligned H&E-ST pairs used to train and evaluate the model.","marker":"[16]"},{"why":"supplies the deformable convolution mechanism used to fuse forward and backward features.","marker":"[17]"},{"why":"supplies the pre-trained ResNet-50 used to extract features from both modalities.","marker":"[21]"},{"why":"supplies the Sobel operator used to compute the H&E gradient magnitude weight w_i.","marker":"[4]"},{"why":"is the prior single-middle-slice interpolation approach this work extends to arbitrary positions.","marker":"[11]"},{"why":"one of the state-of-the-art video frame interpolation baselines compared in the experiments.","marker":"[5]"},{"why":"a depth-aware video frame interpolation baseline the model is compared against.","marker":"[1]"},{"why":"a real-time flow-based video interpolation baseline used in the comparison.","marker":"[3]"},{"why":"a registration-based medical image interpolation baseline used in the comparison.","marker":"[14]"}],"fun_headline_variants":["AI fills any missing gene-expression slice between ST maps","Fill any missing spatial transcriptomics slice with H&E aid","Interpolate missing ST slices at any position using H&E and gene co-expression","Cross-modal method creates any missing gene-expression slice between ST maps","Generate arbitrary missing ST slices with cross-modal alignment"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's central premise is that one scalar per intended slice, the slice index scaled by the H&E image's edge strength, correctly encodes where the missing tissue section lies between the two measured slices; if that scalar does not track how the tissue actually changes between sections, the synthesized intermediate slice is not biologically valid.","fun_headline_variants_meta":{"raw":{"variants":["AI fills any missing gene-expression slice between ST maps","Fill any missing spatial transcriptomics slice with H&E aid","Interpolate missing ST slices at any position using H&E and gene co-expression","Cross-modal method creates any missing gene-expression slice between ST maps","Generate arbitrary missing ST slices with cross-modal alignment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00058,"raw_usage":{"total_tokens":2792,"prompt_tokens":1062,"completion_tokens":1730,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":678,"completion_tokens_details":{"reasoning_tokens":1646}},"tokens_in":678,"tokens_out":1730,"duration_ms":12446,"temperature":1.0,"reasoning_tokens":1646,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:04:14.167897+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out real middle sections from HNSCC triples, train the model once with the proposed gradient-weighted position and once with alpha=0, and compare per-gene PSNR and Pearson correlation; if the gradient-weighted version is not consistently better on slice pairs with large H&E edge magnitudes, the distance model's core assumption fails.","supporting_citations":[{"cited_title":"IEEE Journal of solid-state circuits23(2), 358–367 (1988)","cited_arxiv_id":null,"evidence_quote":"supplies the Sobel operator used to compute the H&E gradient magnitude weight w_i."},{"cited_title":"Scientific Reports12(1), 4133 (2022)","cited_arxiv_id":null,"evidence_quote":"is the prior single-middle-slice interpolation approach this work extends to arbitrary positions."},{"cited_title":"IFRNet: Intermediate Feature Refine Network for Efficient Frame Interpolation","cited_arxiv_id":"2205.14620","evidence_quote":"one of the state-of-the-art video frame interpolation baselines compared in the experiments."},{"cited_title":"Depth-Aware Video Frame Interpolation","cited_arxiv_id":"1904.00830","evidence_quote":"a depth-aware video frame interpolation baseline the model is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"a real-time flow-based video interpolation baseline used in the comparison."}],"review_version":1}