{"id":"7193251e-e8fb-4e37-9291-df7e38ca9dc9","arxiv_id":"2412.09231","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A learned volumetric medical image codec transfers inter-slice latent features across slices so a single compressed stream supports both image reconstruction and direct organ segmentation.","lead":"The paper proposes a learned compression system for 3D medical scans that uses one compressed file for both human viewing and machine segmentation. It reports lower bit rates than several standard codecs for reconstruction, and better segmentation accuracy than JPEG 2000 when analyzing compressed data directly.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The machine-vision comparison is not like-for-like: the segmentation head is retrained per codec and the latent input has far more channels than the JP2K pixel input, so reported DICE/HD95 gains may reflect input capacity/adaptation rather than bitstream information preservation.","rationale":"The reader's conditional verdict already hinges on the fairness of retraining TransUNet on VVMIC latent features. My stress-test sharpens that concern by isolating an even more concrete confound: the input dimensionality differs by a factor of roughly 192 between the latent branch and the JP2K pixel branch, and the segmentation head is adapted separately for each. This makes the DICE/HD95 comparison an unfair test of whether the bitstream itself preserves task-relevant information. The human-vision reconstruction results are not the problem; the machine-vision comparison is the load-bearing part of the abstract's claim, and it is currently uncontrolled. A channel-matched JP2K control would settle whether the advantage is real. Because the authors may be able to provide such a control or otherwise justify the protocol, the appropriate verdict remains conditional rather than rejection.","tokens_in":15551,"tokens_out":6714,"duration_ms":74455,"concrete_test":"Re-run the Section IV-D comparison on CHAOS with a channel-matched JP2K control: train a 1x1 conv that maps 1-channel JP2K-decoded pixels to 192 channels, insert it before the modified TransUNet first layer, and retrain with identical hyperparameters, BPP grid, and validation protocol. If the VVMIC-latent DICE/HD95 advantage over this control largely disappears, the reported gains are caused by input channel capacity and per-codec retraining rather than by better task-information preservation in the transmitted bitstream.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a single VVMIC bitstream supports both human reconstruction and machine segmentation, with segmentation performed directly on coded latent features. The only evidence for the machine half is the Section IV-D experiment comparing 'Our Latents' (input Mx) against JP2K pixels through TransUNet. This comparison is confounded in two ways. First, Mx is a multi-channel intermediate feature (the paper states latent and hyper-latent channels are 192), while the JP2K branch receives a 1-channel reconstructed image; Section IV-D says the input channel number is modified to fit each input type. Second, Section IV-E2 states that TransUNet is retrained for each feature type. Thus the segmenter is free to learn codec-specific, high-capacity features from Mx, and the reported DICE/HD95 advantage over JP2K may be due to the larger input dimensionality and per-codec retraining rather than to the bitstream preserving more task-relevant information. The abstract and conclusion nevertheless present the segmentation gains as a general property of the coding framework. This is the weakest load-bearing assumption because the human-reconstruction BD-Rate claims are internally consistent; only the machine-vision claim lacks a controlled baseline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VVMIC, a recurrent volumetric medical image compression framework that produces a single bitstream intended to support both human-oriented reconstruction and machine-oriented segmentation. The model augments a learned autoencoder with an inter-slice auxiliary feature buffer, multi-scale inter-slice analysis/synthesis transforms, and a multi-dimensional context model combining spatial, channel-wise, hyperprior, and inter-slice contexts. The authors report BD-Rate savings of 7.67%, 17.23%, and 8.84% against VVC on the MRNet axial, coronal, and sagittal views, and higher DICE/lower HD95 than JP2K on CHAOS, ACDC, and Synapse when segmentation is performed on the intermediate latent feature Mx rather than on decoded pixels. Ablations on MRNet-Axial attribute the reconstruction gains to the checkerboard context, channel-wise context, and auxiliary inter-slice modules.","tokens_in":15779,"tokens_out":6231,"duration_ms":61102,"significance":"If the claims are borne out, VVMIC would be a useful contribution to JPEG-AI-style coding for medical volumes: the idea of decoding only the intermediate features needed by a segmentation network, rather than full pixels, and the recurrent inter-slice latent context are both natural and potentially impactful for medical cloud storage and transmission. The reconstruction part is internally consistent: the ablation study in Table II shows each suggested module improves BD-Rate, and the rate-distortion curves in Figs. 5-6 are plausible. However, the machine-vision half of the paper is currently not established by a controlled experiment, and the reconstruction comparison lacks relevant volumetric neural baselines and any measure of variability. The paper does not ship code, weights, or error bars, so the empirical claims cannot be independently checked from the manuscript alone. The authors do, to their credit, include a limitations paragraph in Section IV-E4 that acknowledges the fixed update stride and the difficulty of per-volume optimization.","major_comments":[{"comment":"The machine-vision comparison is confounded in two ways. First, the 'Our Latents' input is the multi-channel (192-channel) intermediate feature Mx, whereas the JP2K branch receives a single-channel reconstructed image; Section IV-D states that the input channel number is modified 'to fit the latent decoding feature.' Second, Section IV-E2 states that TransUNet is retrained for each feature type ('we retrain the TransUnet network with different decoding features as input'). Consequently, the reported DICE/HD95 advantage of 'Our Latents' over JP2K can be explained by higher input dimensionality and by per-codec adaptation of the segmenter, rather than by the bitstream preserving more task-relevant information. This is load-bearing because the abstract's machine-vision claim and the JPEG-AI-style 'single bitstream without full pixel decoding' contribution rest on this comparison. I request a controlled evaluation: a fixed segmentation model trained once on original images and then applied to all coded inputs, or a dimensionality-matched comparison (e.g., projecting Mx to one channel, or applying a comparable feature extraction to JP2K reconstructions), plus bitrate-matched operating points and multiple random seeds.","section":"§IV-D and §IV-E2"},{"comment":"The reconstruction comparison is missing the volumetric neural codecs that the paper itself cites as related work. Section II-A discusses aiWave, aiWave-lite, and DBVC as 3D medical image compression methods, but Tables I and Figs. 5-6 compare only 2D image codecs (Minnen2018, Cheng2020, Zou2022) applied to slices, plus traditional codecs. The conclusion's statement that VVMIC 'outperforms existing learning-based networks' is therefore not supported for volumetric learned codecs. In addition, all BD-Rate, DICE, and HD95 numbers are single-run point estimates with no error bars, significance tests, or seeds, so it is impossible to tell whether the reported margins (e.g., 2.19% vs. -7.67% in Table I) are stable. Please add the volumetric baselines or explicitly restrict the claim to 2D slice-wise learned codecs, and report variability across at least three runs.","section":"§IV-C and §II-A"}],"minor_comments":[{"comment":"The qualitative comparison in Fig. 7 is not at matched rates: the VVC point carries BPP 0.209 while the 'Ours' point carries BPP 0.747, so the visual edge-preservation claim is uninterpretable as a compression advantage. The paper should compare at the same or similar BPP, or show rate-distortion curves with markers for each method at all four λ values.","section":"§IV-D"},{"comment":"There are several typos and notation inconsistencies: 'Multi-Dimention' in the Section III-B2 header, 'Segmentatioon' in the Section IV-D header, 'V olumetric' in the running header, and the text in Section IV-E1 refers to 'spatial context priors Ψsp' while Eq. (15) uses Φk_sp for spatial context and Ψ for hyperparameters.","section":"§III-B2 and §IV-D"},{"comment":"The caption of Table I does not state the sign convention; since negative BD-Rate values denote bitrate savings relative to the VVC anchor, the caption should explicitly say 'negative is better' to avoid ambiguity.","section":"Table I"},{"comment":"The transfer-learning protocol is described only briefly. The paper should state explicitly which patient/volume split is used for fine-tuning VVMIC and for training the segmentation network on each dataset, so that the reader can rule out any overlap between training and test volumes.","section":"§IV-B3"},{"comment":"No code, trained weights, or evaluation protocol details are provided. Since the paper reports a new end-to-end framework with many learned components, an anonymous code/weights release would substantially support verification of the reported numbers.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a medical imaging or image coding venue. The reconstruction results are plausible, but the machine-vision evidence is not yet controlled enough to support the headline claim. I would not reject on the reconstruction results alone; rather, the authors should be asked to add a controlled segmentation experiment and the missing volumetric baselines. The lack of error bars and the absence of code are additional barriers to assessing the stability and reproducibility of the results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read on VVMIC. It's a genuine first: no earlier paper in the cited literature combines a single-bitstream volumetric medical codec with direct machine segmentation from intermediate latent features. The architecture is sensible — a recurrent inter-slice latent buffer plus a multi-dimensional context model that folds spatial, channel, and inter-slice priors into the entropy model. The ablation in Table II shows each module earns its keep, and the reconstruction gains on MRNet against VVC (−7.67% / −17.23% / −8.84% BD-Rate) are respectable if the rates hold up. Transfer to CHAOS, ACDC, and Synapse adds breadth.\n\nNow the soft spots. The stress-test note lands: the segmentation comparison against JP2K is not like-for-like. Section IV-D says TransUNet's input channels are modified to fit the latent feature (192 channels) while JP2K gets a one-channel image, and IV-E2 says the segmenter is retrained for each feature type. So the DICE/HD95 advantage could simply reflect the segmenter adapting to a much higher-dimensional, codec-specific representation, not the bitstream preserving task-critical information. The abstract and conclusion state the segmentation result as a general property, which overreaches. This is the weakest load-bearing claim.\n\nTwo more concerns. First, the reconstruction comparison omits the volumetric neural codecs they cite — aiWave and DBVC are in the references but not in the tables. The learned baselines used are 2D slice-wise codecs, so the claim of beating 'existing learning-based networks' on volumetric data is under-supported. Second, the paper never mentions arithmetic coding or an actual entropy coder; the reported BPP is presumably the model's estimated rate. That should be stated explicitly.\n\nNone of these are fatal. The architecture is plausible, the paper is honest about its own limitations (Section IV-E4), and the missing comparisons are easy to enumerate. It deserves a serious referee. I'd ask the authors for a controlled segmentation experiment — same input dimensionality or a fixed task network, plus at least one learned codec's latent features as a baseline — and for a head-to-head with aiWave or DBVC on reconstruction. Also a sentence on how rates are computed.\n\nFor who: this is for people working on medical image compression or human-machine coding standards. It's not a field-reorganizing paper, but it's a useful new application. I'd bring it to a reading group if one of us is working in that area. Send it out; the conditional verdict is right.","headline":"A genuine first for volumetric medical human-machine coding, but the segmentation evidence is confounded and the reconstruction comparison skips the volumetric neural baselines.","tokens_in":16373,"tokens_out":3911,"would_cite":false,"duration_ms":39088,"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":"A single compressed bitstream can support both human viewing of volumetric medical images and direct machine segmentation from latent features, without decoding to pixels.","keywords":["neural image compression","volumetric medical images","human-machine vision","medical image segmentation","entropy coding","inter-slice redundancy","learned autoencoder"],"falsifier":"Run a single, fixed segmentation model (no per-codec retraining) on either VVMIC latent features or decoded pixels from VVMIC, JPEG2000, and a third codec at matched bitrates on the CHAOS, ACDC, and Synapse datasets; if DICE and HD95 for the latent input are not consistently better than for the pixel inputs, the claim that the coded representation directly supports accurate analytics is not established.","tokens_in":15287,"feed_emoji":"🩻","tokens_out":11487,"duration_ms":91626,"temperature":0.7,"pith_summary":"This paper proposes VVMIC, a neural compression codec for volumetric medical images (CT and MRI volumes) that aims to serve two audiences at once: humans who need to look at reconstructed images, and algorithms that need to segment organs. The central claim is that a single bitstream can support both, so a machine-vision model can work directly on the codec's intermediate latent features without ever reconstructing pixels. On knee MRI scans, the codec reports bitrate savings over VVC of about 8% to 17% at equal reconstruction quality. On abdominal and cardiac datasets, segmentations computed from the latent features achieve higher DICE scores and lower boundary errors (HD95) than segmentations computed from JPEG2000-compressed pixels at comparable bitrates.","feed_headline":"Volumetric medical codec feeds segmentation without pixel decoding","feed_subtitle":"Skipping pixel decoding for analytics could cut compute and latency in medical cloud workflows.","key_machinery":"The load-bearing component is the Versatile Volumetric Autoencoder (VVAE), a recurrent module that reads a buffer of auxiliary features $F_{t-1}$ from previous slices, extracts multi-scale inter-slice contexts $E_1, E_2, E_3$ and an inter-slice latent prior $L_F$, and injects them into the current slice's analysis and synthesis transforms. The same auxiliary path produces intermediate decoding features $M_x$ (current-slice) and $M_F$ (inter-slice), which are fused into $F_t$ for the next slice and also handed to the downstream networks. This design is paired with a multi-dimensional context model that combines checkerboard spatial context, channel-wise context, the inter-slice prior $L_F$, and hierarchical hyperparameters $\\Psi$ to estimate the entropy parameters of the quantized latents, which is what makes the rate savings possible.","core_discovery":"The paper's central discovery is that the same volumetric autoencoder that produces high-quality reconstructions also yields partially decoded intermediate features that are themselves suitable for organ segmentation, eliminating the need to decode to pixels for the machine task. The recurrent VVAE module carries inter-slice context forward slice by slice, producing current-slice latent decoding features $M_x$; these features are fed directly into a TransUNet segmentation network, and this 'Our Latents' scheme beats both decoded-pixel input from the same codec and JPEG2000-compressed pixel input in DICE and HD95 across the CHAOS, ACDC, and Synapse datasets. The authors interpret this as evidence that the coded representation preserves the structural information the segmenter needs, even at low bitrates where the reconstruction network cannot produce a clear image.","pith_inferences":["The machine-vision comparison uses JP2K as the only anchor and retrains the segmentation network to each input format; a fairer test of whether the codec's representation genuinely supports analytics would freeze one segmentation model and feed it either latent features or decoded pixels from several codecs at matched bitrates.","The 16-slice update stride implies the inter-slice benefit comes from short-range correlation; volumes with longer-range repetition might benefit from an adaptive stride or hierarchical grouping, which the paper does not test.","If latent features prove to be lossy but task-sufficient, the same approach could apply to other 3D medical analytics like detection or classification, but the paper only demonstrates segmentation, leaving this an open extension rather than an established claim.","The ablation showing every context component contributes suggests the multi-dimensional context model is doing real work; one could test whether a learned transformer-based context model replaces or surpasses this hand-designed aggregation."],"forward_implications":["Medical cloud pipelines could run segmentation on the latent features directly, skipping full pixel decoding and cutting the compute and latency of analytics.","Hospitals could store one bitstream per volume that serves both archival/review and automated triage, instead of maintaining separate compressed copies.","Because the inter-slice prior $L_F$ is not coded as side information, the framework shows a way to exploit inter-slice redundancy at negligible bitrate cost in learned volumetric codecs.","The reported BD-Rate gains over VVC on MRNet suggest learned volumetric codecs are now competitive with advanced video codecs on medical data.","The authors state the framework can be extended to other machine-vision tasks such as classification and detection, since the latent features are claimed to carry expressive intra- and inter-slice information."],"supporting_citations":[{"why":"Defines the JPEG-AI single-stream human-machine vision requirement that this paper targets as its goal.","marker":"[8]"},{"why":"VVC is the anchor codec for the BD-Rate reconstruction comparisons on MRNet.","marker":"[15]"},{"why":"Supplies the joint autoregressive and hierarchical prior entropy-model framework on which the codec is built.","marker":"[18]"},{"why":"Learned image compression baseline that VVMIC must beat in the MRNet rate-distortion comparisons.","marker":"[23]"},{"why":"Earlier machine-vision-only 3D medical image compression method that motivates the move to a dual human-machine codec.","marker":"[31]"},{"why":"Checkerboard two-pass parallel context model used as a component of the multi-dimensional context model.","marker":"[36]"},{"why":"Channel-wise autoregressive context model combined with spatial and inter-slice priors in the entropy model.","marker":"[37]"},{"why":"Prior learning-based volumetric image compression work that establishes the inter-slice redundancy approach this paper extends.","marker":"[39]"},{"why":"TransUNet is the segmentation network that consumes the intermediate latent features in the machine-vision experiments.","marker":"[51]"}],"fun_headline_variants":["Codec enables direct organ segmentation from compressed volumes","Medical image codec skips pixel decode for AI analytics","Compressed medical volumes directly drive segmentation models","VVMIC: one codec for human viewing and machine segmentation","Neural codec outputs latent features for both vision and AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The machine-vision comparison assumes that retraining the segmentation network on each codec's latent features is a fair measure of the codec's information preservation; if the segmenter merely adapts to the VVMIC feature format, the reported DICE and HD95 advantages over JPEG2000 may not hold for a fixed, deployment-time model.","fun_headline_variants_meta":{"raw":{"variants":["Codec enables direct organ segmentation from compressed volumes","Medical image codec skips pixel decode for AI analytics","Compressed medical volumes directly drive segmentation models","VVMIC: one codec for human viewing and machine segmentation","Neural codec outputs latent features for both vision and AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000592,"raw_usage":{"total_tokens":2764,"prompt_tokens":923,"completion_tokens":1841,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":1762}},"tokens_in":539,"tokens_out":1841,"duration_ms":11479,"temperature":1.0,"reasoning_tokens":1762,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:12:40.756680+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a single, fixed segmentation model (no per-codec retraining) on either VVMIC latent features or decoded pixels from VVMIC, JPEG2000, and a third codec at matched bitrates on the CHAOS, ACDC, and Synapse datasets; if DICE and HD95 for the latent input are not consistently better than for the pixel inputs, the claim that the coded representation directly supports accurate analytics is not established.","supporting_citations":[{"cited_title":"The jpeg ai standard: Providing efficient human and machine visual data consumption,","cited_arxiv_id":null,"evidence_quote":"Defines the JPEG-AI single-stream human-machine vision requirement that this paper targets as its goal."},{"cited_title":"The devil is in the details: Window- based attention for image compression,","cited_arxiv_id":null,"evidence_quote":"Learned image compression baseline that VVMIC must beat in the MRNet rate-distortion comparisons."},{"cited_title":"Machine vision guided 3d medical image compression for efficient transmission and accurate segmentation in the clouds,","cited_arxiv_id":null,"evidence_quote":"Earlier machine-vision-only 3D medical image compression method that motivates the move to a dual human-machine codec."},{"cited_title":"aiwave: V olumetric image compression with 3-d trained affine wavelet-like transform,","cited_arxiv_id":null,"evidence_quote":"Prior learning-based volumetric image compression work that establishes the inter-slice redundancy approach this paper extends."}],"review_version":1}