{"id":"be2235a2-7e2b-4d1c-8762-bdc5709770cf","arxiv_id":"2608.11335","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"DD-CMD combines text-guided spatial cross-attention with DCT-based spectral FiLM modulation and reports higher Dice/mIoU than prior text-guided models on QaTa-COV19 and MosMedData+.","lead":"A new deep learning model segments lung infections in X-rays and CT scans by combining text descriptions with both spatial attention and frequency analysis. It reports higher Dice and mIoU scores than prior text-guided models on two public COVID-19 datasets, though the gains are modest.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing frequency-domain baseline [22] and per-metric best-baseline averaging leave the 'strongest prior baselines' claim unsupported.","rationale":"I read the paper as a systems/empirical contribution: two decoder modules (TGSA, STAM) integrated into a coarse-to-fine text-guided segmentation pipeline, evaluated on two public benchmarks. The method is internally coherent: Eq. 1–5 are straightforward, the ablations in Table 2 show consistent incremental gains, and the code link is a positive. The equations do not contain an obvious circularity or hidden mathematical error. The load-bearing risk is entirely in the comparison protocol. The omission of [22] is especially important because the paper's motivation says prior work overlooks frequency content, yet [22] is a frequency-domain fusion method published in MICCAI 2025, and the paper explicitly cites [22] for its splits. If [22] were included and still fell below DD-CMD, the claim would be much stronger; as written, the SOTA claim is not fully substantiated. The per-metric cherry-picking in the average gains and the absence of error bars compound this. None of this is fatal to the architecture, and it is fixable by adding the missing baseline, reporting error bars, and specifying one consistent strongest-baseline comparator. The reader's CONDITIONAL verdict already captures these conditions, so I do not recommend moving the verdict; I agree with the reader's weakest-assumption analysis.","tokens_in":8007,"tokens_out":5933,"duration_ms":49178,"concrete_test":"Use the released code/weights of [22] (or reimplement per the paper) and run it on the exact QaTa-COV19 and MosMedData+ splits named in Section 3, using the same 224x224 resizing, augmentations, 160-epoch AdamW/cosine protocol, batch size 8, and Dice+CE loss. Report Dice and mIoU for [22] and for DD-CMD across 5 random seeds (mean±std). Then recompute the headline gains against one pre-specified consistent baseline (e.g., MMI-UNet for both datasets) rather than per-metric bests. If [22] matches/beats DD-CMD on either dataset, or if the QaTa +0.58 Dice gap is within one standard deviation, the SOTA claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DD-CMD outperforms the strongest prior text-guided baselines on QaTa-COV19 and MosMedData+. The weakest assumption is that Table 1 contains all relevant strong competitors and that the reported gains are computed against a meaningful, consistent comparator. Section 3 states 'Following [13,9,3,22], we adopt the exactly same train/val/test splits for fair comparison,' yet [22]—a frequency-domain multimodal fusion method for language-guided medical image segmentation—is absent from Table 1. This is the same niche the paper says prior work overlooks. If [22] is competitive on these splits, the motivating narrative and the SOTA claim both weaken. The headline average gains (+1.96 Dice, +2.67 mIoU) are also computed by picking the best prior baseline per dataset and per metric (MMI-UNet for QaTa, MAdapter for MosMed Dice, RecLMIS for MosMed mIoU), so no single 'strongest baseline' is compared against. Finally, no error bars are reported; the smallest gain, +0.58 Dice on QaTa, could be within seed-to-seed variance. These are comparison/reporting issues rather than internal flaws, but they are load-bearing for the empirical claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DD-CMD, a text-guided medical image segmentation model that combines spatial cross-attention (TGSA) with frequency-domain channel calibration (STAM) in a coarse-to-fine decoder, followed by a high-resolution refinement module. The method is evaluated on QaTa-COV19 and MosMedData+, reporting 91.46% Dice / 84.26% mIoU and 81.95% Dice / 69.42% mIoU respectively, with claimed average gains of +1.96 Dice and +2.67 mIoU over the strongest prior baselines. The architecture is specified with equations, ablations support the contribution of each component, and code is released. The main weaknesses concern the comparison protocol: the cited frequency-domain baseline [22] is absent from Table 1, the headline gains are computed against different baselines per dataset and metric, and no error bars are reported.","tokens_in":8242,"tokens_out":2862,"duration_ms":25073,"significance":"If the empirical claims hold after a corrected comparison, DD-CMD would be a useful contribution to text-guided medical image segmentation, particularly in demonstrating that frequency-domain conditioning and spatial attention are complementary. The paper is transparent about its architecture and ablations, and the release of code is a practical strength. However, the central claim of outperforming the strongest prior baselines is currently supported only by a comparison set that omits the most relevant frequency-domain competitor [22] and by a per-metric best-baseline averaging protocol. The significance is therefore conditional on strengthening the empirical evidence rather than on the architectural novelty, which is incremental.","major_comments":[{"comment":"The comparison in Table 1 omits reference [22], a frequency-domain multimodal fusion method for language-guided medical image segmentation, even though §3 states that the authors follow [13,9,3,22] for identical train/val/test splits. This omission is load-bearing because the paper's motivating narrative argues that prior work overlooks frequency content, yet [22] operates in exactly that niche. The authors should add [22] to Table 1 (using its reported numbers under the same splits, or by re-running it) or explicitly justify its exclusion. Without this, the claim that DD-CMD outperforms the strongest prior baselines is not fully supported.","section":"§3, Table 1"},{"comment":"The reported average gains of +1.96 Dice and +2.67 mIoU are computed by selecting a different strongest baseline per dataset and metric: MMI-UNet for QaTa-COV19, MAdapter for MosMedData+ Dice, and RecLMIS for MosMedData+ mIoU. This means no single baseline serves as the 'strongest prior baseline' across all four cells. The comparison protocol should be stated explicitly and consistently, for example by reporting gains against the best baseline per dataset or against a single overall strongest baseline, so the reader can verify the claimed margins.","section":"§3, Comparison with SOTA Methods"},{"comment":"No error bars or standard deviations are reported for any method. The smallest headline gain, +0.58 Dice on QaTa-COV19 over MMI-UNet, is within typical seed-to-seed variance for segmentation models at this scale. The authors should report mean and standard deviation over at least three random seeds for the proposed method and, where feasible, for the closest baselines, and indicate whether the observed differences are statistically significant.","section":"§3, Comparison with SOTA Methods"}],"minor_comments":[{"comment":"The sentence 'An overview of DD-CMD is shown in 1' is missing the figure number; it should read 'shown in Fig. 1.'","section":"§2.2"},{"comment":"The notation V(s) is used both for the input visual tokens and for the attended output in Eq. (2); please use separate symbols, e.g., V_in and V_att, to avoid ambiguity.","section":"§2.2, Eq. (2)"},{"comment":"The 'w/o FiLM-Style Params' row is ambiguous about whether the high-resolution refinement FiLM conditioning is also removed; please clarify what remains in this variant.","section":"§3, Table 2"},{"comment":"Several baselines report N/R for parameters and FLOPs, which makes the efficiency comparison incomplete; consider citing the original sources or recomputing these values for a fair comparison.","section":"§3, Table 1"},{"comment":"The qualitative comparison in Fig. 2 does not identify which baseline methods are shown; please label the columns or add a descriptive caption.","section":"§3, Fig. 2"},{"comment":"References [17] and [18] appear to be self-citations with only marginal relevance to the DCT and lightweight-conditioning discussions; please verify that they are necessary and, if possible, cite the original sources of these ideas.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript includes two self-citations ([17], [18]); [17] is a DCT watermarking paper that is only weakly related to the spectral method, and [18] is used for a generic 'lightweight conditioning' claim. This is not disqualifying, but the authors should carefully assess whether these citations are necessary or whether they should be replaced. The missing baseline [22] is the most significant concern; if the authors can add that comparison and clarify the evaluation protocol, the paper would be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes DD-CMD, a text-guided medical image segmentation model that combines spatial cross-attention (TGSA) with a frequency-domain branch (STAM) based on DCT band energies and FiLM conditioning. The architecture is clearly specified, the coarse-to-fine decoder with a separate high-resolution refinement path is sensible, and Table 2's ablations do support each component's contribution. This is a legitimate, if incremental, architectural contribution to the text-guided segmentation subfield.\n\nThe soft spots are real and mostly in the empirical reporting. The paper explicitly follows the splits of [13,9,3,22] but never includes [22] in Table 1, even though [22] is a frequency-domain fusion method for the same task. That is the closest related baseline, and its absence undercuts the motivating claim that prior work overlooked frequency content. The \"average gains\" of +1.96 Dice and +2.67 mIoU are computed by taking the best prior per dataset and per metric (MMI-UNet on QaTa, MAdapter on MosMed Dice, RecLMIS on MosMed mIoU), so there is no single strongest baseline being beaten. And there are no error bars; the +0.58 Dice gain on QaTa sits within normal seed variance for this kind of model. These are comparison/reporting issues rather than internal contradictions, but they are load-bearing for the main claim.\n\nMinor notes: self-citation [17] is a DCT watermarking paper of weak relevance, and [18] is fine. The code link is provided but not verified; I did not run anything.\n\nOverall, the architecture itself looks coherent and the ablations are honest. The empirical claim as stated is not supported until the missing baseline and error bars are addressed. This is fixable in revision.\n\nWho is this for? Researchers working on language-guided medical segmentation who want a well-ablated decoder module. It will not change clinical practice, but it is useful within the subfield.\n\nMy recommendation: send it to peer review. A serious referee should require comparison against [22], per-seed variance, and a consistent baseline definition for the headline numbers. The paper deserves that engagement.","headline":"A plausible dual-domain decoder with a solid ablation, but the headline SOTA claim is undercut by a missing closest baseline and per-metric cherry-picking of the comparator.","tokens_in":8746,"tokens_out":1616,"would_cite":false,"duration_ms":14908,"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":"Clinical text guides segmentation in both space and frequency, lifting Dice to 91.46%.","keywords":["medical image segmentation","text-guided segmentation","dual-domain cross-modal decoding","cross-attention","discrete cosine transform","FiLM conditioning","pulmonary infection segmentation","clinical text"],"falsifier":"Run the frequency-domain multimodal fusion method cited as [22] on the same QaTa-COV19 and MosMedData+ train/val/test splits with the same augmentation and preprocessing; if its Dice/mIoU reaches or exceeds 91.46/84.26 and 81.95/69.42 on those exact splits, the paper's central claim of beating the strongest prior baselines would be contradicted.","tokens_in":7802,"feed_emoji":"🩻","tokens_out":6216,"duration_ms":49613,"temperature":0.7,"pith_summary":"The paper sets out to show that clinical text can guide medical image segmentation more effectively when language influences decoding in two distinct ways at once: where to attend spatially, and how strongly each feature channel responds to different frequency content. Embedded in a coarse-to-fine decoder, it pairs a text-guided spatial cross-attention module with a spectral-text modulation module that uses 2D DCT band-energy statistics and FiLM-style conditioning. The result, applied to pulmonary infection segmentation, is reported as 91.46% Dice / 84.26% mIoU on QaTa-COV19 and 81.95% Dice / 69.42% mIoU on MosMedData+, with average gains of +1.96 Dice and +2.67 mIoU over the strongest baselines in its comparison table. A sympathetic reader would take the paper to establish that frequency-aware language conditioning is a real, additive source of improvement beyond spatial alignment, particularly for boundary-sensitive lesion masks.","feed_headline":"91.46% Dice: text guides masks in space and frequency","feed_subtitle":"Spatial cross-attention plus frequency recalibration beats prior baselines on two pulmonary infection benchmarks","key_machinery":"The load-bearing mechanism is the dual-domain decoder built from two named modules. TGSA (Text-Guided Spatial Cross-Attention) uses locally mixed visual tokens as queries and clinical text embeddings as keys and values, scaling the attention output by a learned image-text agreement gate. STAM (Spectral-Text Adaptive Modulation) applies a channel-wise 2D discrete cosine transform, computes learnably gated per-channel band energy, concatenates it with the pooled text vector, and predicts FiLM parameters $(\\gamma,\\beta)$ that are applied with a tanh bound to recalibrate channels. The coarse-to-fine schedule keeps cross-modal attention at low resolutions, while the two-stage refinement module restores full-resolution boundaries through shallow-feature fusion and FiLM conditioning from the text.","core_discovery":"At its core, the paper claims that clinical text should steer segmentation in two complementary channels during decoding: spatially, by aligning visual features with text semantics, and spectrally, by recalibrating feature channels according to text-conditioned frequency-band statistics. It embodies this in DD-CMD, whose decoder interleaves TGSA and STAM at three progressively finer resolutions from 7x7 to 56x56, followed by a lightweight refinement module that restores 224x224 masks with FiLM conditioning and shallow skip features. On QaTa-COV19 the method reaches 91.46% Dice and 84.26% mIoU, and on MosMedData+ it reaches 81.95% Dice and 69.42% mIoU, yielding average gains of +1.96 Dice and +2.67 mIoU over the strongest prior baselines in the comparison table. The paper interprets these results as evidence that spatial alignment and frequency calibration are complementary forms of language guidance.","pith_inferences":["Because STAM's DCT and FiLM conditioning do not depend on pulmonary texture specifically, the same dual-domain decoder could transfer to other lesion types or imaging modalities where boundary fidelity is frequency-sensitive; the paper does not test this.","The learned frequency gate inside STAM could be visualized to show which DCT bands clinical text emphasizes, offering a direct check on whether the gains come from semantically meaningful frequency channels; the paper does not report such an analysis.","The frozen text encoder and L=24 truncation suggest that the reported gains rely on clinical descriptors appearing within the first 24 tokens, a dependency the paper's own text-length ablation supports."],"forward_implications":["DD-CMD reports the highest Dice and mIoU on both benchmarks among all compared text-free and text-guided methods.","Ablations show that adding STAM to TGSA improves overlap and lowers HD95 on both datasets, so frequency-aware channel recalibration contributes beyond spatial alignment.","Ablations show that removing either TGSA or STAM degrades performance, with TGSA providing the larger individual jump.","The high-resolution refinement module mainly sharpens boundaries and closes small gaps rather than changing coarse localization, consistent with the coarse-to-fine design.","Text length near L=24 is needed for the reported performance; shorter prompts truncate lesion descriptors and longer prompts add noise."],"supporting_citations":[{"why":"Introduces text-guided medical segmentation with vision-language pairing and supplies the datasets and evaluation protocol used for comparisons.","marker":"[13]"},{"why":"MAdapter is the strongest prior Dice baseline on MosMedData+ in the comparison table, so its numbers anchor one of the claimed improvements.","marker":"[24]"},{"why":"RecLMIS provides the strongest prior mIoU baseline on MosMedData+, setting the reference for one of the reported gains.","marker":"[11]"},{"why":"MMI-UNet is the strongest prior baseline on QaTa-COV19 in the table, the direct competitor DD-CMD must beat.","marker":"[4]"},{"why":"ConvNeXt-Tiny supplies the four-scale visual feature pyramid that the decoder consumes.","marker":"[14]"},{"why":"PubMedBERT provides the frozen clinical text encoder whose token embeddings and pooled vector condition TGSA and STAM.","marker":"[8]"},{"why":"Defines the 2D discrete cosine transform used to compute band-energy statistics in STAM.","marker":"[1]"},{"why":"Introduces FiLM conditioning, the mechanism STAM and the refinement module use to recalibrate features from text.","marker":"[16]"}],"fun_headline_variants":["Text steers segmentation in space and frequency","Dual-domain text guidance bests prior baselines","Spatial and spectral text cues both sharpen segmentation","Text-conditioned frequency bands boost Dice by +1.96"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"That DD-CMD beats the strongest prior baselines assumes the methods in Table 1 are the strongest relevant ones under the same splits; in particular, the cited frequency-domain fusion method [22] is never compared, so it could tie or beat the reported numbers under identical conditions.","fun_headline_variants_meta":{"raw":{"variants":["Text steers segmentation in space and frequency","Dual-domain text guidance bests prior baselines","Spatial and spectral text cues both sharpen segmentation","Text-conditioned frequency bands boost Dice by +1.96"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000209,"raw_usage":{"total_tokens":1431,"prompt_tokens":992,"completion_tokens":439,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":376}},"tokens_in":608,"tokens_out":439,"duration_ms":41710,"temperature":1.0,"reasoning_tokens":376,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:12:44.107363+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the frequency-domain multimodal fusion method cited as [22] on the same QaTa-COV19 and MosMedData+ train/val/test splits with the same augmentation and preprocessing; if its Dice/mIoU reaches or exceeds 91.46/84.26 and 81.95/69.42 on those exact splits, the paper's central claim of beating the strongest prior baselines would be contradicted.","supporting_citations":[{"cited_title":"IEEE transactions on medical imaging43(1), 96–107 (2023) 10 M","cited_arxiv_id":null,"evidence_quote":"Introduces text-guided medical segmentation with vision-language pairing and supplies the datasets and evaluation protocol used for comparisons."},{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"MAdapter is the strongest prior Dice baseline on MosMedData+ in the comparison table, so its numbers anchor one of the claimed improvements."},{"cited_title":"IEEE Transactions on Medical Imaging44(4), 1821–1835 (2024)","cited_arxiv_id":null,"evidence_quote":"RecLMIS provides the strongest prior mIoU baseline on MosMedData+, setting the reference for one of the reported gains."},{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"MMI-UNet is the strongest prior baseline on QaTa-COV19 in the table, the direct competitor DD-CMD must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the 2D discrete cosine transform used to compute band-energy statistics in STAM."}],"review_version":1}