{"id":"800f90fd-329f-4a39-8f8c-6e3f7ce47590","arxiv_id":"1908.05599","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A marginal super-resolution, fusion, and refinement pipeline using only 2D CNNs outperforms linear and 2D/3D CNN baselines for interpolating anisotropic brain MRI.","lead":"Anisotropic MRI scans have sparse spacing between slices; this paper fills in the missing slices by super-resolving the two fully-sampled side views and then fusing and refining the estimates. The reported three-stage 2D CNN pipeline outperforms linear interpolation and 2D/3D CNN baselines on 120 ADNI brain scans at 4x and 8x sparsity.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 3D CNN baseline is capacity- and receptive-field-limited (one RDB, 144x144x256 patches), so the claimed advantage over 3D CNNs may reflect a weak baseline rather than the marginal 2D decomposition.","rationale":"The reader's CONDITIONAL verdict remains appropriate. My concern is a sharper version of the capacity-mismatch weakness already noted in the reader's rationale: the 3D CNN baseline is not matched in capacity or receptive field, so the central empirical claim 'our method consistently outperforms 3D CNN baselines' is not yet established against a fair 3D comparator. This does not move the verdict from CONDITIONAL; it specifies the condition under which the claim would be convincing. The reader's weakest_assumption identifies information sufficiency of the marginal views, and the proposed concrete test addresses that assumption indirectly by giving the 3D model full context. If the matched 3D model still loses, the sufficiency assumption gains support; if it wins, the assumption fails and the central claim weakens. I therefore agree only partially with the reader's framing: the underlying concern is the same, but the most direct threat is the confounded baseline comparison, not a missing formal proof of equivalence.","tokens_in":8609,"tokens_out":9005,"duration_ms":92978,"concrete_test":"Train a 3D CNN baseline with the same total parameter count as the proposed RDN (two RDBs) and with a receptive field covering the full 256x256x256 volume, using memory-efficient techniques such as gradient checkpointing, mixed precision, or a hierarchical/global 3D architecture. Evaluate it on the same central 144x144x256 region and the same 20 test volumes as in Table 1. If this matched-capacity 3D baseline reaches or exceeds PSNR 32.22/SSIM 0.9441 at 4x and 28.87/0.8808 at 8x, then the claimed advantage of the marginal 2D decomposition over a properly resourced 3D model is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest empirical claim is that the proposed pipeline 'consistently outperforms 2D CNN and 3D CNN baselines' (Section 5.2). This comparison is not controlled for model capacity or receptive field. In Section 5.1, the proposed RDN uses two RDBs, while the 3D CNN 'only uses one RDB' due to memory constraints, and the 3D CNN is trained and evaluated on limited local regions (144x144x256). The 2D pipeline operates on full 2D slices, so it sees the whole y-z plane, while the 3D baseline sees only small patches. The reported gain over the 3D CNN (0.88 dB at 4x, 1.49 dB at 8x in PSNR) could therefore be due to the baseline's reduced receptive field and parameter budget rather than to the marginal super-resolution decomposition itself. The information-sufficiency assumption that two orthogonal 2D views capture the 3D context a full 3D model would use remains untested because the only 3D comparator is deliberately weakened.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a deep-learning pipeline for interpolating missing axial slices in anisotropic brain MRI. The first stage, marginal super-resolution (MSR), applies a 2D residual dense network to slabs of sagittal and coronal slices and reformats the outputs into two estimates of the axial volume. The second stage, two-view fusion and refinement (TFR), fuses these two axial estimates and refines the fused volume along the axial direction. Experiments on 120 ADNI T1 volumes with 4x and 8x under-sampling compare the method with linear interpolation and with 2D and 3D CNN baselines, using PSNR/SSIM and gray/white-matter segmentation Dice and Hausdorff distances. The paper also reports an ablation at 4x sparsity. The reported results show the proposed method achieving the best numbers on all metrics.","tokens_in":8854,"tokens_out":7012,"duration_ms":64744,"significance":"If the main comparison were fully controlled, the paper would make a genuinely useful practical contribution: it shows a memory-efficient 2D-only strategy for slice interpolation, with a clear experimental protocol, a held-out split, an ablation, and downstream-segmentation evaluation. The MSR+TFR pipeline is intuitive and the ablation indicates that each component adds value. However, the headline claim of consistent superiority over 3D CNNs is weakened by an unbalanced 3D baseline and by the absence of any statistical uncertainty estimates, so the central comparison is not yet established at the level required by a journal.","major_comments":[{"comment":"The comparison against the 3D CNN baseline is not controlled for model capacity or receptive field. The paper states in Sec. 5.1 that \"due to memory constraint, 3D CNN only uses one RDB\", while the proposed pipeline uses two RDBs, and the 3D CNN is evaluated on a limited 144x144x256 region whereas the proposed method processes full 2D slices. The reported PSNR/SSIM advantages over the 3D CNN (0.88 dB at 4x and 1.49 dB at 8x in PSNR from Table 1) could therefore be attributed to the weaker baseline rather than to the marginal super-resolution decomposition. The authors should add a 3D baseline with matched parameter count and effective receptive field (e.g., via gradient accumulation or overlapping sub-volume inference), or explicitly state this as a limitation and provide a separate justification for the claimed advantage over 3D processing.","section":"Sec. 5.1 and Table 1"},{"comment":"All quantitative claims are based on single point estimates. With only 20 test volumes, the phrase \"consistently outperforms\" (Sec. 5.2) is not supported by a statistical test. The authors should report per-subject means with standard deviations or confidence intervals, and apply a paired significance test (e.g., Wilcoxon signed-rank or paired t-test) across the test volumes for PSNR, SSIM, DICE, and HD. Without this, the reported gains of roughly 0.9-1.5 dB PSNR and the DICE differences may not be significant.","section":"Sec. 5.2 and Table 1"},{"comment":"The paper assumes that 2D super-resolution of sagittal and coronal views is informationally sufficient for axial slice interpolation, but this assumption is never validated against a strong 3D model. Because the only 3D comparator is capacity-limited (see the first major comment), the experiments cannot distinguish between the merit of the marginal decomposition and the effect of using a more powerful 2D network. A concrete test would be to compare the MSR-only stage (Table 2) with a full-resolution 3D network of matched capacity on a subset, or to measure the information loss via a consistency check between the fused volume and the observed axial slices. At minimum, the paper should discuss this limitation explicitly.","section":"Sec. 4.1"}],"minor_comments":[{"comment":"Several figures are numbered Fig. 4 in Secs. 5.3 and 5.4, which makes cross-referencing confusing; the figures should be renumbered.","section":"Figure numbering"},{"comment":"There are typos such as \"super-resovled\" and inconsistent use of \"saggital\"/\"sagittal\"; the manuscript should be proofread.","section":"Sec. 4.2"},{"comment":"The test-set restriction to slices containing \"mostly brain tissues\" should be quantified (how many slices per volume) and its effect on the reported metrics should be discussed.","section":"Sec. 5.1"},{"comment":"Architectural details for the baseline 2D CNN and the fusion/refinement networks are underspecified; the text says only that the upsampling network is replaced by one convolutional layer, but the number of filters, kernel sizes, activations, and strides should be reported for reproducibility.","section":"Secs. 4.2 and 5.1"},{"comment":"The statement that at 8x sparsity linear interpolation's HD scores are \"comparable\" to the CNN baselines is not clearly supported by Table 1, where the WM HD values are 10.59 for LI versus 9.017 for 3D CNN.","section":"Sec. 5.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The MSR decomposition is the real thing here: instead of interpolating between adjacent axial slices or training a full 3D network, you super-resolve the sagittal and coronal views with a 2D network, then fuse and refine axially. That is a genuine departure from the prior slice-interpolation literature, and the practical memory savings are obvious. The ablation is well done, each component earns its place, and they report segmentation Dice and Hausdorff distance rather than only PSNR/SSIM. I also give them credit for trying GAN refinement and dropping it because it invented fake anatomy; that is the kind of negative result we do not see enough of.\n\nThe soft spots are real but not fatal. The headline comparison against the 3D CNN baseline is not controlled. Their 3D net uses one RDB instead of two and sees only 144x144x256 patches, while the proposed networks see full 2D slices. So the 0.88 and 1.49 dB gains over the 3D CNN could reflect capacity and receptive field differences, not the marginal decomposition itself. The ablation partially rescues the story because even single-view MSR3 beats both baselines at 4x, but that only shows the 2D approach is memory-practical, not that it is informationally superior to a well-parameterized 3D model. They also give no error bars or significance tests, and they evaluate only on slices that contain mostly brain tissue, which may inflate the numbers. No code or models are released, so reproducibility is hard to check.\n\nThe writing is clear, the claims are mostly proportionate, and the one overreach is that sentence about consistently outperforming the 3D CNN. A revision that makes the baseline comparison fairer, adds variance estimates, and clarifies the test-slice selection would make this a solid paper. I send it to referees, and I would tell those referees to focus on the baseline issue and the missing statistics.\n\nThe paper is for people working on MRI slice interpolation or any volumetric reconstruction where 3D memory is the bottleneck. The MSR idea is worth knowing, and the module is easy enough to reimplement. With modest revisions it becomes a dependable reference for the two-view fusion approach.","headline":"Genuinely novel marginal super-resolution decomposition, but the 3D CNN comparison is capacity-uncontrolled; still worth refereeing after addressing that and missing statistics.","tokens_in":9345,"tokens_out":3306,"would_cite":true,"duration_ms":32640,"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":"Deep slice interpolation can be done with 2D networks alone, and the 2D pipeline beats both 2D and 3D CNN baselines on anisotropic brain MRI.","keywords":["slice interpolation","marginal super-resolution","2D convolutional neural network","anisotropic MRI","two-view fusion","brain volume estimation","semantic segmentation","image refinement"],"falsifier":"Take isotropic brain volumes, keep every $k$-th axial slice and discard the rest, then run the pipeline. If for some volumes the reconstruction PSNR falls below that of a 3D CNN while the 3D CNN stays high, and the failures concentrate where sagittal and coronal views are smooth but axial anatomy changes sharply, the marginal-super-resolution sufficiency assumption would be violated.","tokens_in":8446,"feed_emoji":"🧠","tokens_out":6531,"duration_ms":55317,"temperature":0.7,"pith_summary":"This paper argues that interpolating missing axial slices in an anisotropic brain MR scan does not require a 3D neural network. The proposed pipeline treats the undersampled volume as two stacks of 2D slices, sagittal and coronal, and super-resolves each stack with a 2D convolutional network, producing line-by-line estimates of the missing axial slices. A fusion network combines the two estimates, and a refinement network improves consistency along the axial axis. The payoff is that the fully 2D method beats 2D and 3D CNN baselines on reconstruction quality and gray/white-matter segmentation while needing less GPU memory and fewer training examples.","feed_headline":"Two 2D views beat 3D CNNs at MRI slice interpolation","feed_subtitle":"Orthogonal-view super-resolution plus fusion and refinement rebuilds missing brain slices without 3D memory costs.","key_machinery":"The load-bearing object is marginal super-resolution (MSR): super-resolving the downsampled volume independently from the sagittal and coronal directions, using a residual dense network that takes a slab of $s=3$ slices and outputs a single high-resolution slice. The second component is two-view fusion, a small network that takes the re-sliced sagittal and coronal super-resolved estimates of each axial slice and combines them into one slice, removing the horizontal and vertical streak artifacts each view leaves behind. The third component is a refinement network that processes a slab of $k+1$ slices along the axial direction to enforce between-slice consistency; the paper uses an $L^1$ loss at every stage and reports that adversarial loss at refinement produces fake anatomy.","core_discovery":"The central claim is that 3D slice interpolation can be decomposed into two 2D super-resolution problems followed by fusion and refinement, and that this decomposition is not just a memory-saving approximation but a better predictor. On 120 T1 brain scans downsampled by factors of 4 and 8, the full pipeline reports the best PSNR/SSIM (32.22 dB/0.9441 at 4x, 28.87 dB/0.8808 at 8x) and the best gray/white-matter Dice coefficients (0.9021/0.8593 at 4x, 0.8189/0.7828 at 8x) among the compared methods. The paper interprets the consistent margin over the 3D CNN as evidence that exploiting full 2D context through orthogonal views is more valuable than a direct 3D mapping trained on small patches.","pith_inferences":["If marginal super-resolution is as sufficient as the paper's results suggest, the same decompose-fuse-refine recipe should transfer to other anisotropic volumetric modalities, such as CT or fetal ultrasound, where orthogonal views carry complementary anatomy.","A direct stress test would be to run the pipeline on slices whose gaps are much larger than 8x, or on volumes where the under-sampled axis is not axis-aligned, since the orthogonality of the two marginal views weakens in that setting.","The improved segmentation dice on interpolated slices implies that volumetric biomarkers, such as gray-matter volume, could be estimated reliably from heavily under-sampled scans without acquiring an isotropic volume.","The reported failure of adversarial loss at the refinement stage suggests a natural follow-up: replace the GAN objective with a structure-preserving or consistency-based loss that sharpens details without inventing anatomy."],"forward_implications":["At 4x sparsity the full pipeline reaches 32.22 dB PSNR and 0.9441 SSIM, above the 2D CNN (31.24 dB, 0.9313) and the 3D CNN (31.34 dB, 0.9292).","At 8x sparsity the gap grows: 28.87 dB and 0.8808 SSIM versus 27.88 dB/0.8444 for the 2D CNN and 27.38 dB/0.8390 for the 3D CNN.","Interpolated slices segment gray and white matter more accurately, with GM/WM Dice of 0.9021/0.8593 at 4x and 0.8189/0.7828 at 8x.","The two-view fusion removes the horizontal and vertical streak artifacts visible in single-view MSR, and the refinement step adds between-slice consistency.","Because every operation is 2D, models can be trained on full slices rather than small patches, which the paper credits for the edge over the 3D CNN."],"supporting_citations":[{"why":"supplies the residual dense network used as the feature extractor for marginal super-resolution.","marker":"[18]"},{"why":"establishes convolutional super-resolution as the baseline approach that MSR extends to multi-view volumes.","marker":"[11]"},{"why":"defines the traditional intensity-based slice interpolation baseline that linear interpolation instantiates.","marker":"[5]"},{"why":"provides the U-Net architecture used to segment gray and white matter in the evaluation.","marker":"[9]"},{"why":"supplies the public brain MRI dataset used to train the segmentation network.","marker":"[23]"},{"why":"cited as the source for the U-Net winning model in the MRBrainS challenge that the segmentation setup follows.","marker":"[22]"}],"fun_headline_variants":["2D views outperform 3D CNNs for MRI slice interpolation","Marginal super-resolution beats 3D CNNs on brain MRI interpolation","Orthogonal 2D super-resolution improves MRI slice interpolation","Fusing two 2D views beats 3D CNNs for missing MRI slices","Brain MRI interpolation: two 2D steps beat full 3D CNNs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes that 2D super-resolution of the sagittal and coronal views captures enough 3D anatomical context to reconstruct the missing axial slices, even when the anatomy changes sharply between observed slices.","fun_headline_variants_meta":{"raw":{"variants":["2D views outperform 3D CNNs for MRI slice interpolation","Marginal super-resolution beats 3D CNNs on brain MRI interpolation","Orthogonal 2D super-resolution improves MRI slice interpolation","Fusing two 2D views beats 3D CNNs for missing MRI slices","Brain MRI interpolation: two 2D steps beat full 3D CNNs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000622,"raw_usage":{"total_tokens":2866,"prompt_tokens":914,"completion_tokens":1952,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":1855}},"tokens_in":530,"tokens_out":1952,"duration_ms":11749,"temperature":1.0,"reasoning_tokens":1855,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:07:35.773160+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take isotropic brain volumes, keep every $k$-th axial slice and discard the rest, then run the pipeline. If for some volumes the reconstruction PSNR falls below that of a 3D CNN while the 3D CNN stays high, and the failures concentrate where sagittal and coronal views are smooth but axial anatomy changes sharply, the marginal-super-resolution sufficiency assumption would be violated.","supporting_citations":[{"cited_title":"IEEE Trans","cited_arxiv_id":null,"evidence_quote":"defines the traditional intensity-based slice interpolation baseline that linear interpolation instantiates."},{"cited_title":"In: Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015 - 18th International Conference Munich, Germany, October 5 - 9, 2015, Proceedings, Part III","cited_arxiv_id":null,"evidence_quote":"provides the U-Net architecture used to segment gray and white matter in the evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the public brain MRI dataset used to train the segmentation network."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"cited as the source for the U-Net winning model in the MRBrainS challenge that the segmentation setup follows."}],"review_version":1}