{"id":"7c62e309-b044-4ac1-bf88-1aecb2d43ccb","arxiv_id":"1908.08251","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"For CNN-based liver segmentation from DCE-MRI, providing all contrast phases as input channels outperforms using a single phase or merging separate phase features.","lead":"This paper compares three ways to feed dynamic contrast enhanced MRI scans into a neural network for liver segmentation: one phase, all phases separately, or all phases stacked as input channels. It finds that stacking the phases as channels works best, but the evidence is partial and the study is small.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Configuration II uses shared weights across phases, so 'channels better than separate phases' is confounded: phase-specific first-layer weights in III, absent in II, may drive the result.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: configuration II's shared weights make it an unfair representative of 'separate phases as input.' My review agrees and sharpens the mechanism: III has not only early fusion but also phase-specific first-layer weights and slightly more parameters, while II has neither. This is a genuine confound, not a stylistic objection. The paper's own Section 5 explicitly concedes that separate per-phase weight optimization might improve II, which is equivalent to admitting the reported comparison is not clean. The central claim as stated in the conclusion is therefore underdetermined by the experiments. I do not think this requires rejecting the paper; the observed effect may well survive a fairer test, and the practical finding about channel input is plausible. But as it stands, the evidence supports only a conditional acceptance, exactly the reader's verdict. I would not move the verdict. A secondary issue, that for the U-net the DSC difference between I and III was not reported as significant (only II vs III was), further tempers the abstract's blanket claim, but the shared-weight confound is the more fundamental problem because it targets the main novel comparison. The concrete test above would settle whether the concern lands. If no code or data are released, this test cannot be independently run, which adds to the conditionality but does not change the verdict.","tokens_in":5724,"tokens_out":5838,"duration_ms":63510,"concrete_test":"Retrain configuration II on the same data with six independent convolutional streams, one per phase, no shared weights, merged at the same second-last layer; keep the total parameter count comparable to configuration III by reducing the number of kernels per stream (e.g., 16 kernels per stream in the dilated FCN instead of 32). Use the same loss, training schedule, post-processing, and the same 19-patient test set. If the non-shared configuration II matches or exceeds configuration III on DSC and HD, the original conclusion is confounded by weight sharing. If configuration III still outperforms, the concern is resolved and the channel-input advantage is real.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim depends on the comparison between configuration II (six phases processed separately) and configuration III (six phases as input channels). In configuration II, the same convolutional kernels are applied to every phase and the resulting feature maps are merged at the second-last layer. In configuration III, all phases enter as input channels, so the first convolution has six input channels and can learn separate weights for each phase. The comparison therefore conflates early versus late fusion with the ability to learn phase-specific features at the first layer. Additionally, configuration III has more parameters at that first layer: for the dilated FCN, 32*6*3*3 versus 32*1*3*3 weights, an extra 1,440 parameters; for the U-net the excess is 5*16*3*3 = 720. The authors acknowledge in Section 5 that shared weights 'might not be the most optimal setup' and that separately optimizing kernel weights for each phase 'might result in better segmentation results.' This is precisely the gap: the observed superiority of III over II could stem from the weight-sharing constraint or the added first-layer capacity, not from the channel-input representation itself. Since configuration II is the only 'separate phases' baseline, this confound directly undermines the conclusion that channel input performs better 'regardless of the architecture of the network.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper compares three ways of presenting six-phase DCE-MRI data to convolutional neural networks for liver segmentation: (I) a single selected late-arterial-phase image, (II) the six phases processed separately through shared-weight networks and merged late with an extra 1×1 convolution, and (III) the six phases stacked as input channels. Two architectures are considered, a dilated FCN and a small U-net. Using 16 training, 3 validation, and 19 test DCE-MR series, the authors report that configuration III gives significantly higher DSC than configuration II for both networks and than configuration I for the dilated FCN, and that III gives significantly better HD than II in the U-net. They conclude that channel-stacked DCE-MR input is superior to single-phase or separate-phase input regardless of network architecture.","tokens_in":6026,"tokens_out":7446,"duration_ms":70762,"significance":"The question is practically relevant: if the result holds, it provides simple, actionable guidance for designing CNN inputs for DCE-MRI segmentation. The study has genuine strengths: two different architectures are compared, both DSC and a boundary metric are reported, manual annotations are independently generated and checked, and the paper is transparent about the training setup and the acknowledged limitations of configuration II. However, the central comparison is confounded by an asymmetry in the way configuration II and III allocate first-layer weights, and the statistical evidence is weaker than the abstract's categorical conclusion. With an additional ablation or a suitably narrowed claim, this could be a useful feasibility result; in its current form the main conclusion is not fully established.","major_comments":[{"comment":"The load-bearing comparison between configuration II (separate phase streams) and configuration III (phase channels) is confounded by weight sharing and first-layer parameter capacity. In configuration II, the same convolutional kernels are applied to every phase and the streams are merged only by a late 1×1 convolution; in configuration III, the first convolutional layer has six input channels and can learn phase-specific kernels at the earliest layer. For the dilated FCN, the first layer in III has 32×6×3×3 weights versus 32×1×3×3 for each stream in II, an excess of 1,440 weights; for the U-net the excess is 5×16×3×3 = 720 weights. The authors themselves state in Section 5 that shared weights in configuration II 'might not be the most optimal setup' and that separately optimizing kernel weights for each phase 'might result in better segmentation results.' The observed advantage of III over II could therefore reflect the addition of early phase-specific capacity or the suboptimal weight-sharing constraint, rather than the channel-input representation itself. I request an additional ablation, for example a separate-weight variant of configuration II or a parameter-matched early-fusion comparison, before the central claim is accepted.","section":"Section 5"},{"comment":"The statistical support is weaker than the abstract's claim that channel input 'performs better' than the other configurations. The test set has only 19 series; many pairwise comparisons are performed without correction for multiple testing; and several comparisons do not reach significance, including DSC I versus III for the U-net, and most Hausdorff-distance comparisons for the dilated FCN. No p-values or effect sizes are reported, so the reader cannot assess the magnitude or reliability of the differences. The authors should either restrict the conclusion to the comparisons that are significant, pre-specify the primary comparisons, or apply a multiple-comparison correction and report exact p-values.","section":"Section 4"}],"minor_comments":[{"comment":"The loss formula is written as (2*X∩Y + s)/(X^2 + Y^2 + s); the denominator appears to be a typo for the sum of the volumes, X+Y, as used in the DSC definition. Please correct.","section":"Section 2.3"},{"comment":"The definition of the modified Hausdorff distance is garbled: 'K 95th ranked minimum of the minimum Euclidean distances' should be a directional 95th-percentile Hausdorff distance, and the formula has a misplaced subscript. Please rewrite.","section":"Section 3"},{"comment":"In the paragraph beginning 'The results did not show a significant difference', configuration III is referred to as configuration II; the text should say 'when using the six phases as channels input image (configuration III)'.","section":"Section 5"},{"comment":"The intensity normalization assumes that the 99.8th percentile corresponds to the contrast-agent peak in the aorta; this is an ad hoc assumption and should be explicitly treated as a free parameter whose influence on the comparison is untested.","section":"Section 2.1"},{"comment":"The boxplots in Figure 3 use asterisks without a table of p-values or a key identifying the exact pairs; a supplementary table with all pairwise p-values and effect sizes would improve the reproducibility of the statistical claims.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a clearly written feasibility study, but the main conclusion is not yet supported because configuration III is compared against a weight-shared configuration II, and the authors acknowledge that independent per-phase weights could change that baseline's performance. The small test set and uncorrected multiple testing further weaken the claims. I would not reject the paper: the confound is curable by adding one ablation or by reframing the conclusion, and the empirical comparison is otherwise informative. Please ask the authors to address the confound and the statistical reporting in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a reasonable, transparent feasibility study that tests a practical question — how to feed the six phases of DCE-MRI into a CNN for liver segmentation. The authors compare single-phase input, six phases processed separately with shared weights, and six phases stacked as input channels, in two network architectures. Their main finding, that channel-stacking outperforms the separate-phase configuration in both networks, is plausible and likely true for the configurations as implemented. But the paper does not fully earn the sweeping conclusion that channel input is better \"regardless of the architecture,\" because the comparison is confounded.\n\nWhat's genuinely new: as far as I know this is the first direct comparison of these three input formats for DCE-MRI liver segmentation. It is a simple, useful experiment. The authors also report inter- and intra-observer variability, which gives context for what \"good\" looks like, and they are honest about the main weakness of their setup in Section 5 — the shared weights in the separate-phase arm. Credit where due: they flag the exact confound themselves.\n\nThe soft spots are real but not fatal. The biggest is the one they admit: in configuration II, the same convolutional kernels are applied to every phase, so the first layer cannot learn phase-specific features, whereas configuration III gives the network six input channels and more first-layer parameters (about 1,440 extra for the dilated FCN, 720 for the U-net). So the observed superiority of III over II could come from the weight sharing or the extra capacity, not from the \"channels\" representation per se. A fairer baseline would give each phase its own first-layer weights and then fuse later. That changes the claim from \"channels are better than separate phases\" to \"early fusion with phase-specific kernels beats late fusion with shared kernels.\"\n\nThe statistical support is also partial. For DSC, III beats II in both networks, but III beats I in only the dilated FCN, and I beats II in only the U-net. For HD, only one comparison is significant. With n=19 and no multiple-comparison correction, the abstract overstates the consistency. No code or data is provided, which limits reproducibility, though for a feasibility study that's a lesser sin.\n\nIn sum: this is a useful, honestly written paper for people working on DCE-MRI or multi-phase imaging segmentation. It gives a practical hint that stacking phases as channels is a good default, but the design does not isolate the cause. It deserves peer review, but with a request for a fairer baseline or a carefully reworded conclusion. I'd take it as a conference paper that needs a revision, not a desk reject.","headline":"Useful, honest feasibility study on input format for DCE-MRI liver segmentation; the channel-stacking result is plausible but the key comparison is confounded by weight sharing and the stats are thinner than the abstract claims.","tokens_in":6523,"tokens_out":2447,"would_cite":false,"duration_ms":24945,"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":"Stacking the six phases of a DCE-MR series as input channels improves CNN liver segmentation over single-phase or separate-phase inputs.","keywords":["dynamic contrast enhanced MRI","liver segmentation","deep learning","convolutional neural networks","input configuration","multi-channel input","Dice similarity coefficient"],"falsifier":"Retrain configuration II on the same data with independently initialized convolutional weights for each of the six phases and compare Dice on the same test set; if per-phase weights erase the gap to configuration III, the central claim about input representation is not supported.","tokens_in":5558,"feed_emoji":"🧲","tokens_out":6875,"duration_ms":62812,"temperature":0.7,"pith_summary":"This paper asks how best to feed a dynamic contrast enhanced MRI liver scan into a convolutional neural network for segmentation. It compares three input configurations: a single late-arterial-phase image, the six contrast phases processed as separate images merged late in the network, and the six phases stacked as channels of one input image. The central claim is that the channel-stacked configuration gives the best segmentation, significantly outperforming both alternatives for each of two network architectures. If the claim is right, a pure input-format change captures the dynamic information in DCE-MRI without modifying the network.","feed_headline":"Six DCE-MR phases as channels outperform single-phase liver inputs","feed_subtitle":"Both tested networks segment the liver better when the six contrast phases are combined in one input image.","key_machinery":"The load-bearing object is the input tensor: six co-registered DCE-MR phase slices stacked as channels of a single image, so early convolutional kernels compare phase intensities at the same spatial location before any nonlinear feature extraction. Configuration II feeds the same six phases through the network separately and merges feature maps only at the second-to-last layer, with convolutional weights shared across phases. The authors argue that the channel version lets the network learn cross-phase relationships from raw MR intensities, which they identify as the likely reason for its advantage. The two architectures, a dilated fully convolutional network and a compact U-net, serve to show that the effect is not tied to a single network design.","core_discovery":"On the paper's own terms, the discovery is that collapsing the six DCE-MR phases into the channel dimension of a single input image is a better way to exploit dynamic contrast information for liver segmentation than either ignoring the dynamics (one phase) or delaying fusion until late feature maps (separate phases). Both the Dice similarity coefficient and the 95th-percentile Hausdorff distance favor configuration III, with significant paired differences over configuration II for both tested networks and over configuration I for the dilated FCN. The two networks did not differ significantly under configuration III, which the authors take as evidence that input configuration matters more than architecture. The authors also report that automatic segmentations remain significantly below inter- and intra-observer agreement, although a radiologist judged the best configuration's outlines satisfactory for further analysis.","pith_inferences":["A decisive test of the paper's interpretation would be to retrain configuration II with separate, per-phase convolutional weights; if it then matches configuration III, the reported advantage is a weight-sharing artifact rather than a property of channel input.","Because the networks were trained on 2D slices and the authors note the concept extends to 3D, a 3D multi-channel version with all phases as channels of a volume is a natural next test that could sharpen or overturn the advantage.","The same channel-stacking logic plausibly transfers to other multi-phase or multi-sequence imaging tasks, such as DCE-MRI kidney or tumor segmentation or CT perfusion, where early fusion of temporal phases may be broadly beneficial."],"forward_implications":["On the tested data, reformatting existing DCE-MR series as multi-channel input improves liver segmentation without any change to network architecture or training procedure.","Because the two architectures behaved similarly under configuration III, input-format choice should be treated as a first-order design variable in similar segmentation studies.","The channel-stacked configuration included most liver lesions, showing that the early-fusion input preserves functional contrast information useful beyond the organ boundary.","Both networks still fall significantly short of human observer agreement, so the input-format gain alone does not close the gap to manual segmentation."],"supporting_citations":[{"why":"Supplies the U-net architecture used as one of the two networks in the comparison.","marker":"[5]"},{"why":"Supplies the dilated fully convolutional network used as the other network.","marker":"[6]"},{"why":"Supplies the motion-correction registration that aligns the DCE-MR phases before they are stacked as channels.","marker":"[8]"},{"why":"Supplies the Dice-based loss function used to train both networks and underlies the evaluation metric.","marker":"[9]"},{"why":"Supports the argument that comparable parameter counts keep overfitting risk similar across configurations, making the comparison fair.","marker":"[11]"}],"fun_headline_variants":["Liver segmentation improves when DCE-MRI phases are input as channels","Channel-combined DCE-MRI phases beat single-phase liver segmentation","DCE-MRI phases as channels boost CNN liver segmentation","Optimal DCE-MRI input: phases as channels for liver CNNs","Stacked DCE-MRI phases improve liver segmentation in CNNs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The main comparison assumes that running the six phases through shared convolutional weights is a fair way to implement 'separate phase input'; if the six phases instead had their own weights, separate-phase input might perform as well as channel-stacked input.","fun_headline_variants_meta":{"raw":{"variants":["Liver segmentation improves when DCE-MRI phases are input as channels","Channel-combined DCE-MRI phases beat single-phase liver segmentation","DCE-MRI phases as channels boost CNN liver segmentation","Optimal DCE-MRI input: phases as channels for liver CNNs","Stacked DCE-MRI phases improve liver segmentation in CNNs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000807,"raw_usage":{"total_tokens":3556,"prompt_tokens":972,"completion_tokens":2584,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":2492}},"tokens_in":588,"tokens_out":2584,"duration_ms":15488,"temperature":1.0,"reasoning_tokens":2492,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:44:46.211765+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain configuration II on the same data with independently initialized convolutional weights for each of the six phases and compare Dice on the same test set; if per-phase weights erase the gap to configuration III, the central claim about input representation is not supported.","supporting_citations":[{"cited_title":"(eds) Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015","cited_arxiv_id":null,"evidence_quote":"Supplies the U-net architecture used as one of the two networks in the comparison."},{"cited_title":"and Koltun, V., `` Multi-Scale Context Aggregation by Dilated Convolutions ,'' in [ Fourth International Conference on Learning Representations 0.1em ] , (2016)","cited_arxiv_id":null,"evidence_quote":"Supplies the dilated fully convolutional network used as the other network."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the motion-correction registration that aligns the DCE-MR phases before they are stacked as channels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Dice-based loss function used to train both networks and underlies the evaluation metric."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the argument that comparable parameter counts keep overfitting risk similar across configurations, making the comparison fair."}],"review_version":1}