{"id":"7a35834f-d7b6-4d9b-b1c2-dbb5ab37d3c4","arxiv_id":"2506.10142","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A dual-branch frequency-domain brain tumor segmentation network, HFF-Net, reports large Dice improvements over prior methods on four public MRI datasets, especially for contrast-enhancing tumor regions.","lead":"This paper proposes HFF-Net, a brain tumor segmentation network that splits MRI images into low- and high-frequency parts to better capture tumor edges and textures. If the reported Dice gains hold under independent evaluation, it would be a strong practical improvement for an established medical imaging task.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Final output is chosen as the 'superior' of two branch predictions (§III-E), but no test-time selection criterion is given; if selection uses ground-truth Dice, the reported gains are inflated.","rationale":"The central claim is not merely that HFF-Net is accurate, but that the frequency-domain modules cause the large reported gains. That causal claim requires the reported predictions to be produced by the network under a fixed, non-oracle inference rule. The manuscript never specifies how the \"superior\" branch is selected at test time, which is a concrete internal omission that can invalidate HFF-Net's own numbers, not just the comparison to quoted baselines. The reader's weakest assumption correctly flags protocol mismatch and unverifiable baselines; my concern is narrower and more severe because a test-time oracle branch selector would inflate the method's own Dice scores. The paper does contain useful independent evidence, including ablations, cross-domain 2D results, and failure analysis, so rejection is not warranted. However, acceptance should remain conditional on releasing the code and demonstrating a fixed, non-oracle branch selection rule that reproduces the headline ET improvements.","tokens_in":34126,"tokens_out":8306,"duration_ms":94321,"concrete_test":"Inspect the released HFF-Net repository's inference path for how the LF and HF main outputs are combined. Specifically, check whether branch selection uses ground-truth labels, a validation split, a confidence/uncertainty heuristic, or a fixed rule. Then run the BraTS2023-MEN evaluation three ways: (a) select by ground-truth Dice, (b) average the two branch probability maps, and (c) always use the LF branch. If (a) reproduces the reported 96.16 ET Dice but (b) or (c) drop ET Dice by more than roughly 1-2 points, the central claim that the frequency-domain design causes the improvement is unsupported. Even without code, the authors should state the selection rule and ablate it.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section III-A says the final result y is \"selected from the optimal primary branch prediction,\" and Section III-E says it is \"determined by comparing and selecting the superior main output from the two branches.\" No inference-time criterion for \"superior\" is described: there is no confidence score, no uncertainty estimate, no fusion rule, and no ablation comparing selection strategies. Because L_unsup (Eq. 18-22) is a symmetric consistency term between branch outputs, it cannot decide which branch is better at test time. If the choice is made with held-out labels (e.g., choosing the branch with higher Dice), every reported number, including the 96.16 ET Dice on BraTS2023-MEN, becomes an oracle upper bound rather than a model prediction. This is more load-bearing than the quoted-baseline issue because it affects HFF-Net's own outputs, not just comparisons. The paper's claim of \"end-to-end inference without any post-processing\" is also in tension with a post-hoc branch selection. The provided GitHub URL was not verifiable in this review, so this step remains unverifiable from the manuscript.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HFF-Net, a dual-branch network for brain tumor segmentation that decomposes multimodal MRI into low-frequency (DTCWT) and high-frequency (NSCT) components, then processes them with an Adaptive Laplacian Convolution (ALC) layer, Frequency Domain Cross-Attention (FDCA), and a 3D Dynamic Focal Loss consistency term. The authors report large Dice improvements over prior methods, especially for enhancing tumor (ET), on BraTS2019, BraTS2020, BraTS2023-MEN, MSD-BTS, and additional cross-domain datasets, with ablations, visualizations, and failure-case analyses supporting the frequency-domain design.","tokens_in":34348,"tokens_out":6572,"duration_ms":76641,"significance":"If the reported gains survive a shared-protocol comparison, the frequency-decomposition design is a plausible and potentially useful direction for improving ET boundary segmentation. The paper is strong in breadth: it includes four 3D brain datasets, two additional 3D datasets, four 2D datasets, component ablations, operator ablations, decomposition-strategy ablations, Grad-CAM and t-SNE analyses, and a code link. However, the central numerical claim is not currently established because the main comparisons use baselines quoted from prior publications rather than re-run under identical conditions, and the final-branch selection rule at inference is unspecified.","major_comments":[{"comment":"Section III-A states that 'the final segmentation result y is selected from the optimal primary branch prediction' and Section III-E states that it is 'determined by comparing and selecting the superior main output from the two branches.' The manuscript does not specify any inference-time criterion for this selection: there is no confidence score, uncertainty estimate, fusion rule, or ablation of selection strategies. If the selection uses ground-truth labels, then every reported number, including the 96.16% ET Dice on BraTS2023-MEN, is an oracle upper bound rather than a model output, and the claimed 'end-to-end inference without any post-processing' is not supported. Please specify the selection criterion used, or change the inference to a fixed branch or deterministic fusion, and report results under that protocol.","section":"§III-A, §III-E"},{"comment":"The main tables compare HFF-Net with baseline numbers taken from prior publications rather than results re-run under the same protocol. The implementation details in Section IV-B (350 epochs, 128³ patches, learning-rate schedule, λ_max=15, warm-up, DTCWT/NSCT settings) are specific to HFF-Net, so the reported gains of 4.48% mean Dice and 7.33% ET Dice cannot be attributed to the frequency-domain design without a shared experimental protocol. Please re-run all compared methods under identical data splits, preprocessing, patch extraction, training length, and inference settings, or provide a quantified analysis of protocol sensitivity.","section":"§IV-C, Tables I–IV"},{"comment":"Section IV-A describes BraTS2023-MEN as split into 80% training, 15% validation, and 5% testing, while Table V states that 'all experiments were conducted using five-fold cross-validation.' The HFF-Net numbers also differ between Table I and Table V (e.g., ET 96.16 vs 96.1, TC 96.34 vs 96.1). Please clarify which evaluation protocol produced which table and report a single consistent protocol with per-fold results and standard deviations for each dataset.","section":"§IV-A, §IV-C, Table V"},{"comment":"Section III-D says the DTCWT decomposition level is one and later says NSCT uses 'two levels for both pyramid and directional decomposition,' but Table IX identifies the best configuration as [1,4] (one level, four directions) and states that two-level setups perform worse. The exact FDD configuration is therefore ambiguous. Please align the method description with the code and report the precise filter banks and decomposition levels used for the main results.","section":"§III-D, Table IX"},{"comment":"The 3D Dynamic Focal Loss is not reproducible from the text as written: D in Eq. (20) is already a sum over u,v,w, and Eq. (22) then multiplies this sum by per-frequency weights and divides by N, which is not a weighted average of amplitude distance and dynamic weights. Please provide the exact tensor operations, with shapes and the role of α, and clarify whether the 'prediction from one branch as pseudo-label' is applied in the frequency or spatial domain.","section":"§III-E, Eqs. (19)–(22)"}],"minor_comments":[{"comment":"Table III reports 36.01M parameters for HFF-Net while Table VII reports 37.76M for the full model; please reconcile these numbers.","section":"Tables III and VII"},{"comment":"A throughput of 6.71 images/s corresponds to about 149 ms per image, not the reported 159.3 ms; please correct the inconsistency.","section":"Figure 10(b)"},{"comment":"Table VIII reports single-run results without standard deviations; the 0.58% ET gap between the Kirsch and discrete Laplace operators is small relative to the standard deviations in Table V, so statistical significance should be assessed.","section":"Table VIII"},{"comment":"Cross-dataset results on LiTS and LA are reported without error bars or number of runs; since the test sets are small, repeated-run statistics are needed to support the claimed improvements.","section":"Table VI"},{"comment":"The symbol ŷ^ft_i uses the superscript t before it is defined; please define t and the side-output subscript explicitly.","section":"Eq. (17)"}],"recommendation":"major_revision","confidential_remarks":"The GitHub repository was not verifiable during this review, so the branch-selection rule and exact FDD settings could not be checked from code. I recommend requesting the code and asking the authors to provide the exact selection criterion and re-run statistics for baseline methods under a shared protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this paper. The architecture is a real contribution: combining DTCWT for low-frequency and NSCT for high-frequency, with an EWC-style Laplacian-initialized convolution and an FFT-based attention module, is a combination I have not seen for 3D brain tumor segmentation. The ablations are genuinely informative, covering different transform pairs, decomposition levels, and directions, and the results match the stated rationale. The paper also goes beyond the usual single-dataset evaluation: transfer to LiTS and LA, a 2D variant on four datasets, failure-case analysis, and significance tests. That is more than most architecture papers in this space.\n\nBut the evaluation has a load-bearing problem. Sections III-A and III-E state that the final prediction is 'selected from the optimal primary branch' and 'determined by comparing and selecting the superior main output' of the two branches, but no inference-time criterion is ever described. The unsupervised loss is symmetric, so at test time there is no principled way to know which branch is 'superior.' If the choice uses ground-truth Dice, then every reported number, including the 96.16 ET on BraTS2023-MEN, is an oracle upper bound. This affects HFF-Net's own outputs, not just the comparison to baselines, and it is more serious than the second issue: baseline numbers in Tables I–IV are quoted from other papers rather than re-run under a shared protocol. Part of the gain could come from training details like patch size, epoch count, or full-volume inference rather than the frequency design.\n\nThere are also minor internal inconsistencies: the ablation text says ALC adds 1.96% ET, but the table shows a 3.0-point jump, and Table V's numbers differ slightly from Tables I/II. Those are fixable.\n\nBottom line: the frequency-domain direction is worth pursuing and the paper deserves a serious referee. But the referee should demand a precise description of the branch-selection mechanism and a re-run of at least the strongest baselines on one dataset. If the selection is label-based, the paper needs major revision. I would not cite the reported numbers in my own work until that is resolved, but I would bring it to a reading group to discuss the architecture and the evaluation pitfall.","headline":"Interesting frequency-domain architecture; the reported Dice gains look inflated until the branch-selection rule and baseline protocol are clarified.","tokens_in":34901,"tokens_out":3044,"would_cite":false,"duration_ms":35591,"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":"HFF-Net claims that splitting brain MRI into low- and high-frequency branches lifts enhancing-tumor Dice by 5–8 points over recent state-of-the-art baselines.","keywords":["brain tumor segmentation","frequency domain","contrast-enhancing tumor","dual-tree complex wavelet transform","nonsubsampled contourlet transform","cross-attention","BraTS"],"falsifier":"Re-run S2CA-Net and nnUNet on BraTS2020 under HFF-Net's exact data splits, patch size (128×128×128), training length (350 epochs), and inference protocol; if their ET Dice rises to within the reported margin, the frequency-domain attribution is not supported.","tokens_in":33953,"feed_emoji":"🧠","tokens_out":4470,"duration_ms":49052,"temperature":0.7,"pith_summary":"The paper proposes HFF-Net, a brain tumor segmentation network that explicitly decomposes multi-modal MRI into low- and high-frequency components before learning to segment. It claims that treating these components separately—smooth contours in the low-frequency branch, directional textures and edges in the high-frequency branch—remedies the persistent degradation of contrast-enhancing tumor (ET) segmentation. On four public datasets, the paper reports consistent Dice gains, averaging 4.48% relative on mean tumor-region Dice and 7.33% relative on enhancing-tumor Dice. The contribution is not just a new architecture but a claim about where the difficulty lies: spatial-domain-only features miss the textural and directional cues that distinguish enhancing tumor boundaries.","feed_headline":"Frequency-split MRI network tops strong baselines on enhancing-tumor Dice","feed_subtitle":"By separating smooth low-frequency contours from directional high-frequency texture, HFF-Net sharpens contrast-enhancing region delineation.","key_machinery":"The argument stands on three modules. FDD (Frequency Domain Decomposition) uses DTCWT for the low-frequency branch, which provides approximate shift invariance, and NSCT for the high-frequency branch, which gives aliasing-free directional sub-bands; the key identity is the Hilbert-pair energy condition |φ̂r(ω)|² + |φ̂i(ω)|² ≈ 1 that makes the low-frequency representation stable under shifts. ALC (Adaptive Laplacian Convolution) initializes convolution kernels with a discrete Laplacian operator and uses Fisher-information Z-score thresholding to freeze important weights, preserving a high-pass filter while adapting the rest. FDCA (Frequency Domain Cross-Attention) transforms features via FFT, applies semantic, positional, and slice attention in the frequency domain, and returns via IFFT. All three modules are jointly trained under a loss that sums supervised Dice on each branch with an unsupervised 3D Dynamic Focal Loss aligning the two branches.","core_discovery":"On the paper's own terms, HFF-Net achieves large, consistent Dice improvements over strong baselines by harmonizing low- and high-frequency information from two classical transforms—DTCWT for low-frequency structure and NSCT for multi-directional high-frequency texture—through three coupled modules: Frequency Domain Decomposition, Adaptive Laplacian Convolution, and Frequency Domain Cross-Attention, trained with a 3D Dynamic Focal Loss that aligns the two branches. The strongest reported result is 96.16 enhancing-tumor Dice on BraTS2023-MEN versus 88.87 for S2CA-Net, and 87.36 versus 80.41 on BraTS2020.","pith_inferences":["A substantial part of the reported margin may come from the dual-branch capacity, patch size (128³), and training protocol (350 epochs, warm-up) rather than from frequency decomposition per se; the paper does not ablate the architecture keeping the total parameter count fixed.","The Fisher-information freezing in ALC is a single-task application of a continual-learning mechanism; its effect could be replicated by a fixed Laplacian filter, so an ablation replacing EWC with a plain frozen kernel would isolate its contribution.","We would expect the ET gains to shrink if baseline methods were re-run under HFF-Net's exact patch and epoch settings; a shared-protocol benchmark is the natural next test.","The reported frequency-domain entropy and shift-invariance metrics are diagnostic, not predictive; they do not yet establish a causal link between decomposition quality and segmentation outcome."],"forward_implications":["If the reported gains hold under a shared protocol, frequency decomposition becomes a cheap, architecture-agnostic preprocessing that any segmentation backbone could adopt.","The 2D variant's success on four non-brain datasets suggests the frequency-decoupling principle transfers to other imaging modalities with low-contrast boundaries.","Clinical ET segmentation at Dice above 95% on meningioma could reduce manual correction in radiotherapy target delineation.","The dual-branch consistency loss offers a route to enforce robustness when one input modality is corrupted, though the paper's own failure cases show this is not yet solved."],"supporting_citations":[{"why":"Supplies the DTCWT for low-frequency decomposition, providing shift invariance and noise robustness.","marker":"[31]"},{"why":"Supplies the NSCT for high-frequency multi-directional decomposition without downsampling aliasing.","marker":"[80]"},{"why":"Strong baseline (nnUNet) whose reported scores the method must beat on BraTS datasets.","marker":"[52]"},{"why":"Recent state-of-the-art baseline (S2CA-Net) used for the headline ET Dice comparisons.","marker":"[97]"},{"why":"Source of the Dice loss used for supervised branch supervision.","marker":"[88]"},{"why":"Provides the elastic weight consolidation (EWC) strategy adopted in the ALC layer to freeze important weights.","marker":"[85]"},{"why":"Defines the BraTS benchmark and its evaluation protocol for brain tumor segmentation.","marker":"[2]"},{"why":"Introduces the BraTS2023-MEN meningioma dataset used for the largest reported gains.","marker":"[89]"}],"fun_headline_variants":["Frequency-domain fusion boosts enhancing-tumor Dice by up to 8.6%","HFF-Net: frequency-split MRI lifts brain tumor segmentation","Two transforms, one network: sharper MRI tumor edges","Adaptive frequency fusion improves enhancing-tumor Dice by 7.3% avg"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported superiority over baselines rests on comparing HFF-Net's numbers against baseline numbers taken from other papers, without a shared, re-run experimental protocol; the reader must assume that protocol differences in patches, epochs, and inference settings do not account for the gain.","fun_headline_variants_meta":{"raw":{"variants":["Frequency-domain fusion boosts enhancing-tumor Dice by up to 8.6%","HFF-Net: frequency-split MRI lifts brain tumor segmentation","Two transforms, one network: sharper MRI tumor edges","Adaptive frequency fusion improves enhancing-tumor Dice by 7.3% avg"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000582,"raw_usage":{"total_tokens":2769,"prompt_tokens":1002,"completion_tokens":1767,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":1689}},"tokens_in":618,"tokens_out":1767,"duration_ms":17367,"temperature":1.0,"reasoning_tokens":1689,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:31:47.203915+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run S2CA-Net and nnUNet on BraTS2020 under HFF-Net's exact data splits, patch size (128×128×128), training length (350 epochs), and inference protocol; if their ET Dice rises to within the reported margin, the frequency-domain attribution is not supported.","supporting_citations":[{"cited_title":"The nonsubsampled con- tourlet transform: theory, design, and applications,","cited_arxiv_id":null,"evidence_quote":"Supplies the NSCT for high-frequency multi-directional decomposition without downsampling aliasing."},{"cited_title":"Shape-scale co- awareness network for 3d brain tumor segmentation,","cited_arxiv_id":null,"evidence_quote":"Recent state-of-the-art baseline (S2CA-Net) used for the headline ET Dice comparisons."},{"cited_title":"A comprehensive survey of continual learning: Theory, method and application,","cited_arxiv_id":null,"evidence_quote":"Provides the elastic weight consolidation (EWC) strategy adopted in the ALC layer to freeze important weights."}],"review_version":1}