{"id":"7a4c0925-6798-4cde-8a2d-07f674251b7c","arxiv_id":"2411.16792","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A 2D diffusion model generates pseudo-high-resolution volumes from low-resolution vEM data, and a 3D network trained on those pseudo-volumes performs 8x axial super-resolution.","lead":"This paper presents D2R, a three-stage training framework that uses a 2D diffusion model to generate high-resolution training volumes from low-resolution vEM data, then trains a 3D network to perform 8x axial super-resolution without high-resolution 3D supervision. The authors report that the method matches supervised training on two FIB-SEM datasets and outperforms other unsupervised methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (1)'s cross-plane isotropy is the load-bearing assumption: both the pseudo-volume fidelity in Stage II and the lateral-to-axial transfer of DGEAN depend on it, and the paper never validates it directly.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the isotropic-distribution assumption in Eq. (1) is what makes it permissible to train a 2D diffusion model on XY slices and apply it to XZ/YZ slices, and then to train a 3D network in lateral directions and apply it axially. I agree that this assumption is unverified and that the paper's final metrics do not directly test it. The proposed diagnostic, comparing the lateral pseudo-volume to ground-truth HR volumes on held-out data, would settle whether the assumption holds on the actual datasets. If it fails, then the framework's success cannot be attributed to the diffusion-based pseudo-supervision as claimed. The paper's other issues, such as the overstated 'no high-resolution supervision' claim and the nonstandard FSC-based resolution metric, are secondary and do not change the conditional recommendation. The reader's conditional verdict, contingent on clarifying and validating the central assumption, remains appropriate.","tokens_in":16467,"tokens_out":14903,"duration_ms":265848,"concrete_test":"Run the Stage I diffusion model on the lateral (XZ and YZ) slices of the held-out LR test volumes to form the pseudo-volume V^H = (V^H_XZ + V^H_YZ)/2, then compare V^H slice-wise with the true HR test volumes. Report PSNR_XZ and PSNR_YZ of V^H against ground truth, alongside the diffusion model's PSNR on reconstructed XY slices from the same volumes. If the lateral PSNRs are materially lower than the XY PSNR (e.g., by more than 1-2 dB), Eq. (1) is falsified and the pseudo-supervision that D2R relies on is not faithful; if the lateral PSNRs are comparable, the assumption is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, that 3D super-resolution networks can be trained without high-resolution volume supervision, rests on the isotropic-distribution assumption in Eq. (1): p(IH_XZ|IL_XZ) = p(IH_YZ|IL_YZ) ≈ p(IH_XY|downsample(IH_XY)). This assumption is invoked twice. First, in Stage II (Sec. 3.3, Eqs. 7-10), a 2D diffusion model trained on XY slices is applied to XZ/YZ slices of the low-resolution volume to generate the pseudo-high-resolution volume V^H. Second, in Stage III (Sec. 3.4), a 3D network trained on lateral directions of V^H is applied in the axial direction at inference. If the statistical relationship between high-resolution and low-resolution slices is not equivalent across spatial dimensions, then the diffusion model is miscalibrated for lateral planes, V^H is not a faithful high-resolution approximation, and the Stage III supervision signal is biased. The paper asserts that biological samples exhibit a consistent data distribution across spatial dimensions, but it provides no direct measurement of this equivalence for the FIB-25 and EPFL datasets. The final PSNR/SSIM metrics are only indirect evidence; they could be produced largely by DGEAN's architecture and continuity losses even if the diffusion prior transfers poorly. The synthetic degradation used in Stage I, downsample(IH_XY), is also not obviously the same degradation that low-resolution lateral slices undergo (one-axis slice discarding), which makes the assumption in Eq. (1) even less secure. Because the entire 'without HR volumes' claim depends on this transfer, it is the most load-bearing unverified step.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes D2R, a three-stage framework for 3D volume super-resolution in focused ion beam scanning electron microscopy (FIB-SEM). In Stage I, a 2D diffusion model is trained on high-resolution XY slices and their degraded versions to learn a slice restoration prior. In Stage II, this diffusion model is applied slice-by-slice along the XZ and YZ planes of a low-resolution volume, and the two resulting volumes are averaged to form a pseudo high-resolution volume. In Stage III, a 3D convolutional network (DGEAN) is trained on this pseudo-volume in the lateral directions; at inference, DGEAN is applied along the axial direction to super-resolve the low-resolution volume. The authors evaluate D2R-DGEAN on the FIB-25 and EPFL datasets, reporting PSNR/SSIM values close to those of the same network trained with full high-resolution volume supervision, as well as favorable downstream segmentation and neuron-reconstruction metrics. The central claims are that D2R trains a 3D super-resolution network without any high-resolution volume supervision and that D2R-DGEAN outperforms all unsupervised vEM super-resolution methods.","tokens_in":16802,"tokens_out":3885,"duration_ms":37753,"significance":"If the claims hold, the paper makes a useful practical contribution: it offers a way to train 3D volume super-resolution networks when isotropic high-resolution volumes are unavailable, which is a common situation in volume electron microscopy. The central empirical comparison is extensive, covering similarity metrics in three planes, resolution estimates, membrane segmentation, and neuron reconstruction, on two public datasets. The paper also provides a concrete network architecture (DGEAN) and demonstrates that its training framework can be applied to an existing network (SRUNet). The idea of using 2D diffusion priors to bootstrap training data for 3D networks is plausible and of interest. However, the significance is tempered by the fact that the method is not fully unsupervised (Stage I requires high-resolution XY slices for diffusion training), and by several load-bearing assumptions and metric-interpretation issues that need to be clarified and validated before the central claims can be accepted.","major_comments":[{"comment":"Equation (1) states the load-bearing assumption p(IH_XZ|IL_XZ) = p(IH_YZ|IL_YZ) ≈ p(IH_XY|downsample(IH_XY)), but the manuscript provides no direct validation that the statistical relation between high- and low-resolution slices is the same across the XY, XZ, and YZ planes for the FIB-25 and EPFL datasets. This assumption is used twice: first to apply the XY-trained diffusion model to lateral slices in Stage II, and second to justify transferring the laterally trained DGEAN to the axial direction in Stage III. The final PSNR/SSIM results are only indirect evidence, because DGEAN's architecture and continuity losses could yield plausible-looking volumes even if the diffusion prior transfers poorly. I request either a direct test (for example, report the lateral-plane accuracy of the Stage II diffusion output against ground-truth lateral slices on held-out volumes, which the experimental setup already has access to) or an ablation in which the diffusion model is trained on XZ/YZ pairs and applied to XY, to quantify the asymmetry.","section":"Sec. 3.2, Eq. (1)"},{"comment":"The 'estimated resolution' in Table 2 uses FSC-0.5 computed between the predicted volume V_pred and the ground-truth volume V_gt, as defined in Supplementary Eq. (1). This is a correlation measure between a prediction and its target, not the standard Fourier shell correlation between two independent half-maps that is used to estimate the intrinsic resolution of a reconstruction. As such, the values in Table 2 conflate reconstruction accuracy with resolution; a method that simply copies more information from the ground truth (e.g., through supervised training) will appear to have 'better resolution' for reasons unrelated to spatial frequency content. The claim that FSC 'strongly demonstrates the effectiveness of our method in volumetric super-resolution reconstruction' should be reworked: either use a proper split-half FSC estimator, or explicitly state that the reported numbers are prediction-vs-ground-truth correlations and interpret them accordingly.","section":"Sec. 4.4 and Supplementary Sec. 1"},{"comment":"The data flow of Stage III training versus axial inference is underspecified. The text says DGEAN is trained on the recovered volume V_hat^H in lateral directions to 'learn stable structure deformation between continuous slices,' and the task is to generate (r−1) slices between consecutive input slices from a sequence I_1:2n. However, the paper does not clearly state what the input and output tensors are during Stage III training: are the 2n input slices taken from V_hat^H and the target slices also from V_hat^H (i.e., interpolation learning), or is V_hat^H first downsampled to synthesize low-resolution inputs paired with high-resolution targets? At inference, DGEAN is applied directly to the low-resolution volume V_L in the axial direction, which implies the network must map low-resolution inputs to high-resolution outputs. If Stage III only trains on high-resolution lateral slices without synthesizing low-resolution inputs, the domain gap between training (lateral, high-res inputs) and inference (axial, low-res inputs) is a major concern that is not addressed in the paper. Please specify the exact paired inputs/targets used in Stage III, including how the relative depth parameter d is encoded in training and whether it is consistent with the r=8 factor used in the experiments.","section":"Sec. 3.3 and Sec. 3.4"},{"comment":"The paper repeatedly describes the method as training 'without any high-resolution volumes as supervision' (which is technically true) and also as an 'unsupervised' method (which is misleading, since Stage I trains the 2D diffusion model on paired high-resolution and degraded XY slices). The supervision column in Table 1 marks D2R-SRUNet and D2R-DGEAN as unsupervised (×), but these methods rely on high-resolution 2D supervision in Stage I. This distinction matters because the comparison to other 'unsupervised' baselines, such as IsoRecon and IsoVEM, is presented as apples-to-apples; those methods do not use any high-resolution slices at all. The claims should be reworded to state precisely that the method avoids high-resolution volume supervision but still requires high-resolution XY slices, and the comparison tables should make this difference explicit.","section":"Abstract, Sec. 1, and Table 1 caption"}],"minor_comments":[{"comment":"The heading 'Degredataion Process' contains a typo; it should be 'Degradation Process'.","section":"Sec. 4.1 heading"},{"comment":"'refered' should be 'referred' in the sentence about STDIN.","section":"Sec. 2.1"},{"comment":"The discussion of IsoVEM says it 'assumes that the distribution of low-resolution lateral slices is the same as that of high-resolution axial slices,' but in the proposed method the assumption in Eq. (1) is about conditional distributions p(IH|IL), not about the marginal distribution of high-resolution axial slices. The distinction should be clarified to avoid conflating the two assumptions.","section":"Sec. 2.1"},{"comment":"In the Sup-DGEAN row, the Dice value for Neuron 2 (0.8185) is lower than the corresponding IoU value (0.8109), which is theoretically impossible for binary overlap measures (Dice ≥ IoU always). This suggests a typographical error and should be corrected.","section":"Table 4"},{"comment":"The arbitrary-scale experiment is interesting, but the text says 'the performance of models trained with high-resolution volume as supervision (Sup-DGEAN) and the D2R training framework (D2R-DGEAN) have same performance.' If this is a qualitative statement based on the figure, it should be stated as such; if quantitative, the numbers should be provided.","section":"Supplementary Sec. 4 and Fig. 5"},{"comment":"The training details do not report the number of diffusion sampling steps used in Stage II, the size of the Gaussian noise δ, or the total number of lateral slices processed. These details would help reproducibility.","section":"Sec. 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper makes a useful practical contribution, but the current form overstates the 'unsupervised' nature of the method and uses an FSC metric that is not a true resolution estimator. The most important fix is to address the three major points: validate the isotropy assumption directly, correct the FSC interpretation, and fully specify the Stage III training/inference data flow. If those points are addressed, the paper could be suitable for publication in a computer vision or medical imaging venue. I would also recommend asking the authors to soften the claim of outperforming 'all unsupervised vEM super-resolution methods' given that their method uses high-resolution 2D supervision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I've read the paper and mostly agree with your conditional read. The D2R pipeline is a real contribution: instead of using diffusion directly for inference on every slice, they train a 2D diffusion model on XY slices, apply it slice-by-slice to XZ/YZ planes of the low-res volume, average the two lateral reconstructions to form a pseudo-HR volume, and then train a 3D network (DGEAN) on that volume for axial inference. The cross-plane averaging and the choice to train a fast 3D network rather than sampling diffusion at test time are sensible and address the discontinuity and cost problems of direct diffusion-based VSR. The experimental campaign is unusually thorough: two FIB-SEM datasets, supervised and unsupervised baselines, PSNR/SSIM across three planes, FSC curves, downstream membrane segmentation and neuron reconstruction, plus arbitrary-scale evaluations. The numbers show D2R-DGEAN essentially closes the gap to supervised DGEAN, which is a strong empirical result.\n\nThe main problems are three. First, the abstract and intro claim training without any high-resolution volumes as supervision, but Stage I needs HR XY slices. That's a real overstatement. A precise phrasing would be 'without HR volume supervision' or 'without full-volume ground truth.' Second, Eq. (1) is the load-bearing isotropy assumption, and the paper asserts it with no direct test. The synthetic degradation in Stage I is a 2D downsample of an XY slice, but the lateral slices in the LR volume are high-res in one axis and low-res in the other (due to slice discarding along Z). Those are not the same degradation, so the diffusion model may be miscalibrated for lateral planes. The fact that downstream results are close to supervised is encouraging, but it's indirect evidence. A straightforward fix would be to compare the generated V^H to the actual HR volume on lateral slices (or at least report PSNR/FSC of the pseudo-volumes) to validate the transfer. Third, calling the FSC-0.5 threshold between prediction and ground truth a 'resolution estimate' is nonstandard; it's a frequency-domain similarity metric. The paper should rename it or qualify it.\n\nThe Stage III training details are also thin. They say DGEAN is trained on lateral directions of V^H but don't specify how the lateral training sequences are sampled, the target slices, or patch counts. That needs clarification for reproducibility. No code is released either, which hurts given how much the value of this framework depends on the exact pipeline.\n\nNone of these are fatal. The central idea holds up: a 3D SR network can be trained on pseudo-volumes produced by a 2D diffusion prior, and the results are credible. The paper deserves a serious referee, and with a corrected claim about supervision, a direct test of the isotropy assumption, and released code I would be comfortable with it. Send it to review, but the authors should be asked to address these points.","headline":"A useful diffusion-based training framework for 3D vEM SR that nearly closes the supervised gap, but the 'no HR supervision' claim is overstated and the isotropy assumption goes untested.","tokens_in":17345,"tokens_out":3385,"would_cite":true,"duration_ms":32451,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that 3D volume super-resolution networks can be trained from 2D diffusion-generated pseudo-volumes, with no high-resolution 3D ground truth, and still land within a few tenths of a dB of a supervised network.","keywords":["3D volume super-resolution","diffusion models","volume electron microscopy","FIB-SEM","unsupervised super-resolution","isotropic distribution assumption","lateral continuity"],"falsifier":"Take a FIB-SEM volume whose structures are strongly oriented along the axial direction, downsample it anisotropically, and train the Stage I diffusion model only on XY slices. Then apply it to XZ and YZ slices and compare the per-plane PSNR and FSC of the recovered lateral slices against a model trained on each plane with paired ground truth. A substantially lower lateral-plane accuracy than XY-plane accuracy would falsify the isotropic posterior assumption and invalidate the pseudo-volume training signal.","tokens_in":16283,"feed_emoji":"🔬","tokens_out":9380,"duration_ms":83419,"temperature":0.7,"pith_summary":"The paper introduces D2R, a training framework that claims to train a 3D volume super-resolution network for volume electron microscopy with no high-resolution volume as supervision. Its load-bearing idea is that biological structures have an equivalent distribution across spatial dimensions: a 2D diffusion model trained to restore degraded XY slices learns a posterior that can be reused for the XZ and YZ lateral planes. Those lateral reconstructions are averaged into a pseudo-high-resolution volume that supervises DGEAN, a 3D convolutional network; at inference DGEAN performs the same super-resolution along the axial direction. The authors report D2R-DGEAN reaches PSNR 27.57 dB on FIB-25 in the XY plane and 26.27 dB on a second FIB-SEM dataset, close to the supervised 27.69 dB and 26.40 dB, and it surpasses all unsupervised baselines on similarity, FSC-0.5 resolution, membrane segmentation, and neuron reconstruction. If true, the framework removes the need for paired isotropic 3D training data and turns a slow diffusion process into a finite training-phase generator rather than an at-inference sampler.","feed_headline":"3D super-resolution with no high-res volumes nears supervised quality","feed_subtitle":"A 2D diffusion model creates pseudo-volumes to train the 3D network; reported PSNR gap to supervised is 0.1–0.2 dB.","key_machinery":"The machinery has two connected parts. The first is the isotropic conditional-distribution identity in Eq. (1): the posterior of a high-resolution lateral slice given its low-resolution observation is taken equal to the posterior of a high-resolution XY slice given its downsampled version. This identity allows a mean-reverting stochastic-differential-equation diffusion model, trained with the IRSDE maximum-likelihood objective, to act as a pseudo-label generator for the XZ and YZ planes. The second is DGEAN, a 3D convolutional network with Gaussian Embedding Attention Blocks that inject relative depth and high-frequency position encodings, a lightweight feature attention layer after each encoder, and a final $7\\times7$ 2D convolution whose output is added to a traditional deconvolution result; its loss combines $L_1$, SSIM, focal frequency, and slice-continuity terms so it can learn smooth inter-slice transitions from the pseudo-volumes.","core_discovery":"The paper’s central claim is that the conditional distribution of a high-resolution slice given its low-resolution observation is approximately identical across the three spatial planes, $p(I^H_{XZ}|I^L_{XZ}) = p(I^H_{YZ}|I^L_{YZ}) \\approx p(I^H_{XY}|\\mathrm{downsample}(I^H_{XY}))$. Under this assumption, a 2D diffusion model trained only on synthetic low/high-resolution XY pairs can be applied slice-by-slice to the XZ and YZ planes of a low-resolution volume; averaging the two resulting reconstructions produces a full pseudo-high-resolution volume. That volume is then used as training data for DGEAN, an 18-layer 3D ResNet with Gaussian embedding attention blocks, which learns inter-slice feature transformations and is finally applied along the axial direction to yield the recovered high-resolution volume. The authors argue that the gap between this D2R-trained network and one trained with real high-resolution volumes is small (about 0.1–0.2 dB PSNR in the XY plane on their test sets), while both outperform earlier unsupervised approaches, including in estimated resolution and downstream segmentation.","pith_inferences":["The isotropic assumption is the main transfer risk: if a volume contains structures preferentially oriented along one axis, the XY-trained diffusion prior may generate lateral slices that are statistically implausible, and the 3D student could memorize those hallucinations.","Because the pseudo-volumes are generated by a stochastic sampler and then averaged across XZ and YZ, the averaging step likely suppresses some of the diffusion model's noise but may also blur directional structure; ablating the averaging and keeping XZ-only supervision would reveal how much cross-plane fusion contributes.","A direct stress test would be to run D2R on volumes with deliberately anisotropic structure and known ground truth in three planes; the resulting per-plane FSC curves would show whether the axial-resolution gains come from true recovery or from smoothing.","The framework's design implies that any future improvement to the 2D diffusion prior would automatically improve the pseudo-volume training data, so D2R can be combined with newer diffusion backbones without changing the 3D network."],"forward_implications":["Training 3D super-resolution networks no longer requires paired high-resolution 3D ground truth; any volume-super-resolution architecture that normally needs such supervision can be plugged into the D2R framework.","Because D2R only needs low-resolution volumes plus a 2D diffusion model trained on axial slices, it lowers the data barrier for isotropic volume electron microscopy reconstruction where high-resolution volumes are unavailable.","The same DGEAN model, trained once at scale factor 8, can be evaluated at other upsampling factors without fine-tuning, with the authors reporting consistent qualitative results across scales.","Downstream analyses such as membrane segmentation and neuron reconstruction inherit the benefit; reconstructions from D2R-DGEAN yield segmentation and tracing results close to those from ground-truth volumes.","The framework converts a slow generative procedure into a one-time data augmentation step: once pseudo-volumes are generated, inference is a deterministic 3D convolutional network."],"supporting_citations":[{"why":"Supplies the core observation that a 2D super-resolution model can be transferred to 3D volume reconstruction through the assumption of consistent structure across dimensions.","marker":"[18]"},{"why":"Provides the mean-reverting SDE diffusion model and its maximum-likelihood training objective used in Stage I to estimate the slice restoration posterior.","marker":"[32]"},{"why":"Defines the slice-discarding degradation simulation and the axial-super-resolution baseline against which D2R-DGEAN is compared.","marker":"[30]"},{"why":"Earlier work that also assumes the low-resolution lateral slice distribution matches the high-resolution axial distribution; D2R adopts and extends this assumption.","marker":"[9]"},{"why":"Unsupervised isotropic reconstruction baseline for volume EM that inspires the unsupervised setting and is compared as a benchmark.","marker":"[6]"},{"why":"Lateral-slice diffusion reconstruction baseline that motivates applying a 2D diffusion model plane-by-plane to fill in missing volume content.","marker":"[37]"},{"why":"Reference-free isotropic 3D EM reconstruction using a 2D diffusion prior, one of the prior works whose lateral-diffusion strategy D2R extends.","marker":"[26]"},{"why":"Reference-free axial super-resolution of 3D microscopy using a 2D diffusion prior, supporting the cross-plane transfer idea.","marker":"[27]"},{"why":"Supervised 3D SR baseline that D2R is applied to and compared against to show the framework approaches supervised quality.","marker":"[11]"}],"fun_headline_variants":["2D diffusion model achieves 3D SR without 3D high-res data","3D SR via 2D diffusion, no 3D labels","Slice-by-slice 2D diffusion yields near-supervised 3D SR","Diffusion-based 3D SR trained on 2D pairs only","Leveraging 2D diffusion for 3D volume super-resolution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline rests on the assumption that a slice taken perpendicular to the imaging axis looks statistically like a slice taken along the imaging axis, so a model trained on one direction works for the other two; if biological structure is not equivalent across spatial dimensions, the diffusion model cannot recover the lateral slices and the 3D network has no valid training signal.","fun_headline_variants_meta":{"raw":{"variants":["2D diffusion model achieves 3D SR without 3D high-res data","3D SR via 2D diffusion, no 3D labels","Slice-by-slice 2D diffusion yields near-supervised 3D SR","Diffusion-based 3D SR trained on 2D pairs only","Leveraging 2D diffusion for 3D volume super-resolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.009926,"raw_usage":{"total_tokens":36160,"prompt_tokens":1043,"completion_tokens":35117,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":659,"completion_tokens_details":{"reasoning_tokens":35017}},"tokens_in":659,"tokens_out":35117,"duration_ms":147876,"temperature":1.0,"reasoning_tokens":35017,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:23:06.976793+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a FIB-SEM volume whose structures are strongly oriented along the axial direction, downsample it anisotropically, and train the Stage I diffusion model only on XY slices. Then apply it to XZ and YZ slices and compare the per-plane PSNR and FSC of the recovered lateral slices against a model trained on each plane with paired ground truth. A substantially lower lateral-plane accuracy than XY-plane accuracy would falsify the isotropic posterior assumption and invalidate the pseudo-volume training signal.","supporting_citations":[{"cited_title":"Super-resolution of biomedical volumes with 2d supervision","cited_arxiv_id":null,"evidence_quote":"Supplies the core observation that a 2D super-resolution model can be transferred to 3D volume reconstruction through the assumption of consistent structure across dimensions."},{"cited_title":"Diffusion-based deep learning method for augmenting ultrastructural imaging and volume electron microscopy","cited_arxiv_id":null,"evidence_quote":"Defines the slice-discarding degradation simulation and the axial-super-resolution baseline against which D2R-DGEAN is compared."},{"cited_title":"Isovem: Isotropic reconstruction for volume electron mi- croscopy based on transformer","cited_arxiv_id":null,"evidence_quote":"Earlier work that also assumes the low-resolution lateral slice distribution matches the high-resolution axial distribution; D2R adopts and extends this assumption."},{"cited_title":"Isotropic re- construction of 3d em images with unsupervised degradation learning","cited_arxiv_id":null,"evidence_quote":"Unsupervised isotropic reconstruction baseline for volume EM that inspires the unsupervised setting and is compared as a benchmark."},{"cited_title":"Dif- fuseir: Diffusion models for isotropic reconstruction of 3d microscopic images","cited_arxiv_id":null,"evidence_quote":"Lateral-slice diffusion reconstruction baseline that motivates applying a 2D diffusion model plane-by-plane to fill in missing volume content."},{"cited_title":"Reference-free isotropic 3d em reconstruction using diffusion models","cited_arxiv_id":null,"evidence_quote":"Reference-free isotropic 3D EM reconstruction using a 2D diffusion prior, one of the prior works whose lateral-diffusion strategy D2R extends."},{"cited_title":"Reference-free ax- ial super-resolution of 3d microscopy images using implicit neural representation with a 2d diffusion prior","cited_arxiv_id":null,"evidence_quote":"Reference-free axial super-resolution of 3D microscopy using a 2D diffusion prior, supporting the cross-plane transfer idea."},{"cited_title":"Deep learning for isotropic super-resolution from non- isotropic 3d electron microscopy","cited_arxiv_id":null,"evidence_quote":"Supervised 3D SR baseline that D2R is applied to and compared against to show the framework approaches supervised quality."}],"review_version":1}