{"id":"a659a3ef-13cf-4320-9c24-1e29f0f317f8","arxiv_id":"2412.07804","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"XLSTM-HVED combines a heteromodal variational encoder-decoder with Vision XLSTM and attention-based fusion to improve brain tumor segmentation under missing MRI modalities, reporting state-of-the-art Dice and HD95 on BraTS 2024.","lead":"A new deep learning model, XLSTM-HVED, is proposed for brain tumor segmentation on MRI when some scan types are unavailable. The authors report improved segmentation scores on the BraTS 2024 dataset, but the experimental evidence is incomplete and the reconstruction capability is not measured.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline comparison is not demonstrably fair: the 'same backbone' claim is incompatible with mmformer/RMBTS, no missing-modality training protocol is given, and reported margins are small and statistically untested.","rationale":"The reader's weakest assumption is that the comparison against RA-HVED, RMBTS, and mmformer was fair, and that is exactly the load-bearing concern. I agree, and I sharpened the issue with concrete internal evidence: the paper's 'same backbone network' statement cannot hold for mmformer and RMBTS, no missing-modality training protocol is described, and the reported margins over mmformer are small and inconsistent across configurations. The proposed controlled re-benchmark with modality dropout, identical training conditions, and confidence intervals would settle whether the advantage is real. This concern is addressable and does not invalidate the architectural plausibility of the method, so the existing CONDITIONAL verdict remains appropriate; no verdict adjustment is needed.","tokens_in":6965,"tokens_out":5616,"duration_ms":57222,"concrete_test":"Run a controlled re-benchmark using the same BraTS 2024 80/20 split: train XLSTM-HVED, RA-HVED, RMBTS, and mmformer with identical random modality-dropout during training (each mini-batch randomly masks one or more modalities so every model sees missing-modality patterns), identical augmentation, loss weights, optimizer schedule, and compute budget across at least 3 seeds; report Dice and HD95 with paired 95% confidence intervals for each of the 15 availability masks. If XLSTM-HVED's average advantage over mmformer shrinks to within the confidence intervals or falls below roughly 1 Dice point, the claim of significant superiority is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that XLSTM-HVED 'significantly outperforms existing advanced methods' on missing-modality BraTS 2024 segmentation. Table 1 supports this only if the baselines were trained and evaluated under the same missing-modality protocol as the proposed model. The paper never states how RA-HVED, RMBTS, or mmformer were trained for missing inputs: no modality-dropout schedule, no loss functions, no augmentation, no hyperparameter tuning, no number of seeds, and no confidence intervals. The only stated control is 'All the baselines and our model were trained and tested using the same backbone network,' followed by 'We selected U-HVED [3] as the backbone network.' This is internally inconsistent because mmformer is a transformer and RMBTS is a feature-disentanglement and gated-fusion architecture; neither can literally use the U-HVED backbone. If the baselines were trained on complete four-modality data and then evaluated under missing-modality masks, or simply run for the same 72-hour wall-clock without ensuring convergence, the reported advantage would be an artifact of the evaluation setup rather than of the method. Moreover, the table itself shows that the average Dice advantage over mmformer is small (WT +0.37, TC +3.14, ET +0.84 percentage points), and in several single-modality configurations mmformer outperforms XLSTM-HVED by large margins, e.g., TC Dice on T2-only: 47.24 vs. 36.60, and on T1-only: 44.39 vs. 23.23. Without a controlled comparison, the abstract's 'significantly outperforms' claim is not secured.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes XLSTM-HVED, a multi-task model for brain tumor segmentation and MRI modality reconstruction under missing-modality conditions. The architecture extends the heteromodal variational encoder-decoder with a Self-Attention Variational Encoder (SAVE), a Vision-LSTM Attention (ViLA) module, and a Squeeze-Fusion-Excitation Cross Awareness (SFECA) module. Experiments on BraTS 2024 compare Dice and HD95 against RA-HVED, RMBTS, and mmformer across all 15 subsets of available modalities, and an ablation study examines the three proposed components. The authors claim that their model significantly outperforms existing methods for missing-modality cases; the evidence, however, is limited to single runs with no statistical testing, and the comparison protocol for baselines is not described in sufficient detail. Source code is provided via GitHub.","tokens_in":7243,"tokens_out":5676,"duration_ms":55066,"significance":"If substantiated, the method would be a useful contribution to missing-modality brain tumor segmentation, combining reconstruction and segmentation with recent xLSTM-based attention. The paper's strengths include evaluation on the large BraTS 2024 dataset, ablation of each proposed module, and public source code. However, as presented, the central claim of significant improvement is not established: no repeated runs, confidence intervals, or significance tests are reported, and the baseline protocol is ambiguous. The work is not circular in the sense of fitting parameters to the target outcome; it is an empirical comparison. Its significance is therefore conditional on additional controlled experiments.","major_comments":[{"comment":"The central claim in the abstract that the model 'significantly outperforms existing advanced methods' is not supported by the evidence in Table 1. The table reports only point estimates from what appears to be a single run; no standard deviations, number of seeds, or significance tests are given. Several reported advantages are small (average WT Dice 74.53 vs. mmformer's 74.16; average ET Dice 53.83 vs. 52.99), and in some single-modality configurations baselines are substantially better (T2-only TC Dice: mmformer 47.24 vs. 36.60; T1-only TC Dice: mmformer 44.39 vs. 23.23). The phrase 'significantly' therefore needs repeated experiments and paired statistical tests, or the claim must be withdrawn.","section":"§3.2, Table 1"},{"comment":"The baseline comparison is not demonstrably fair. The paper says 'All the baselines and our model were trained and tested using the same backbone network' and then states 'We selected U-HVED [3] as the backbone network,' but mmformer is a transformer architecture and RMBTS uses feature disentanglement and gated fusion; neither can literally be built on U-HVED. The text also does not specify how any baseline was trained for missing modalities: no modality-dropout schedule, loss functions, data augmentation, hyperparameter tuning, training epochs, or convergence criteria are reported. Training all models for the same 72-hour wall clock is mentioned in §3.1, but this does not guarantee convergence. Without this information, the reported improvements could reflect an unfavorable baseline setup.","section":"§3.1, §3.2"},{"comment":"The title advertises an 'MRI Reconstruction Method,' and §2.1 says the model 'can reconstruct existing MRI data' and 'supplement missing data,' yet no reconstruction-quality results are reported anywhere. Table 1 contains only Dice and HD95 segmentation scores; there are no PSNR/SSIM values, no qualitative reconstruction examples, and no comparison of reconstructed modalities. Either add reconstruction evaluation or revise the title and contribution claims to focus on segmentation.","section":"§2, §3.2, Table 1"},{"comment":"The ablation study in Table 2 is also based on single runs without repeated seeds or significance tests. The conclusion in §4 that 'removing any module significantly diminishes the model's overall performance' is not statistically supported; for example, the WT HD95 for w/o ViLA is 22.69 vs. 17.86 for the full model, and no error bars are provided. The ablation claims should be backed by multiple runs and appropriate tests, or stated descriptively.","section":"§3.3, Table 2"}],"minor_comments":[{"comment":"The caption contains a typo ('A valiable Modalities') and does not define the symbols ◦ and • used to indicate available modalities; please add a legend.","section":"Table 1"},{"comment":"The 80/20 train/test split should specify the number of patients in each set, the random seed, and whether the split was performed at the patient level to avoid data leakage.","section":"§3.1"},{"comment":"References [4] and [13] are the same paper and should be consolidated into a single citation; reference [10] is listed as 'Anonymous' and under review and should be updated to its published version if available.","section":"References"},{"comment":"The module name is written inconsistently as 'SA VE' and 'SAVE'; please use one spelling throughout.","section":"§2.2"}],"recommendation":"major_revision","confidential_remarks":"The main risk is baseline fairness: if the authors cannot provide a controlled comparison with official implementations and matched training protocols, the central claim should not be accepted. I would also urge the editor to require statistical rigor (multiple seeds and significance tests) before considering the paper further."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this is an incremental but sensible engineering integration for missing-modality brain tumor segmentation. It combines the HVED framework with the Vision-LSTM backbone and two attention modules (SAVE, SFECA), and reports experiments on BraTS 2024 with code released. If the comparisons held up, it would be a useful incremental data point for clinical workflows missing MRI sequences. But the central claim that it \"significantly outperforms\" existing methods is not supported by the evidence as presented.\n\nWhat's actually new: the specific combination of Vision-LSTM with HVED, and the two attention modules. The paper describes the architecture clearly, includes an ablation study showing each module contributes, and uses a public benchmark. That is real work, and the direction is relevant.\n\nThe soft spots are substantial. First, the baseline comparison is not demonstrably fair. The paper says all models used \"the same backbone network\" and then says U-HVED was the backbone—but mmformer is a transformer and RMBTS is a feature-disentanglement method, so neither can literally use U-HVED. No missing-modality training protocol is given for the baselines: no modality dropout schedule, no loss functions, no seeds, no confidence intervals. Without that, the small average Dice advantages (e.g., +3.14 TC over mmformer) could be an artifact. The table itself shows mmformer beating XLSTM-HVED by large margins in several single-modality rows (TC Dice T2-only 47.24 vs 36.60; T1-only 44.39 vs 23.23). Second, the abstract's blanket claim is undermined by the paper's own note that ET Dice is lower in some configurations. Third, the title promises MRI reconstruction, but no reconstruction quality metrics are reported—only segmentation. Fourth, there are no error bars or significance tests anywhere.\n\nThese are addressable, not fatal. The core architecture is plausible, and the ablations suggest the modules do something. But the paper needs a rigorous evaluation: controlled baseline training under the same missing-modality protocol, multiple seeds with standard deviations, significance tests, and reconstruction metrics. As is, it's a conditional accept at best.\n\nWho it's for: readers working on missing-modality segmentation or multimodal fusion in medical imaging. It deserves a serious referee—the problem is relevant and the approach is coherent—but only after major revisions.\n\nRecommendation: send it to peer review with the expectation of heavy revision; the current evidence does not support \"significantly outperforms.\"","headline":"A plausible engineering integration for missing-modality brain tumor segmentation, but the headline superiority claim isn't supported by the reported evidence.","tokens_in":7809,"tokens_out":2622,"would_cite":false,"duration_ms":24781,"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":"The paper claims one network can reconstruct missing MRI scans while segmenting brain tumors, reporting better average Dice and HD95 than three baselines on BraTS 2024.","keywords":["brain tumor segmentation","missing MRI modalities","multimodal MRI","Vision XLSTM","variational encoder-decoder","MRI reconstruction","multi-task learning","BraTS 2024"],"falsifier":"Run all compared methods on the same BraTS 2024 split with an identical missing-modality training protocol, including the same random modality dropout schedule, augmentation, and training budget. If the average Dice and HD95 gaps reported in the paper's Table 1 drop below one point or reverse direction, the central claim that the architecture itself is better at missing-modality segmentation would not be supported.","tokens_in":6728,"feed_emoji":"🧠","tokens_out":8695,"duration_ms":83105,"temperature":0.7,"pith_summary":"The paper sets out to show that one neural network can handle both halves of a common clinical problem: some MRI sequences are missing, and the tumor still has to be segmented. The proposed XLSTM-HVED couples a heteromodal variational encoder-decoder with a Vision XLSTM module so that whatever subset of the four standard MRI modalities is present is used to reconstruct the absent sequences and to segment whole tumor, tumor core, and enhancing tumor. On the BraTS 2024 dataset the authors report higher average Dice and lower average HD95 than RA-HVED, RMBTS, and mmformer across nearly all 15 combinations of available modalities. The reason to care is practical: if the claim holds, segmentation no longer waits on a complete imaging protocol, and the reconstructed scans could be inspected by clinicians.","feed_headline":"Missing MRI? One model reconstructs and segments brain tumors","feed_subtitle":"When one of four MRI sequences is unavailable, the network fills it in and keeps tumor boundaries accurate.","key_machinery":"The central object is the three-module architecture itself. SAVE is the fusion engine: a multimodal variational autoencoder whose Product-of-Gaussians latent variable makes the model invariant to which modalities are present. ViLA is the sequence-processing engine: volumetric tokens are passed through mLSTM blocks, which maintain a matrix memory updated by covariance rules, and the module applies attention back to the original input, so spatial layout is not discarded. SFECA is the cross-task engine: channel-wise and spatial squeeze-fusion-excitation blocks recalibrate features flowing to the segmentation and reconstruction decoders, which is how the two tasks influence each other. Together these modules carry the claim that missing-modality robustness comes from better fusion and better task interaction rather than from simply more parameters.","core_discovery":"The paper's central claim is that joint reconstruction and segmentation under missing modalities is improved by combining three mechanisms: a Self-Attention Variational Encoder (SAVE), a Vision XLSTM Attention (ViLA) module, and a Squeeze-Fusion-Excitation Cross Awareness (SFECA) module. SAVE encodes each available modality, fuses their means and covariances through a Product of Gaussians, and samples a shared latent variable, so missing modalities are excluded rather than imputed with zeros. ViLA reshapes the fused volume into tokens and processes them with Vision-LSTM mLSTM blocks plus softmax attention, giving the model a long-range dependence mechanism across the 3D volume. SFECA lets the segmentation and reconstruction decoders exchange recalibrated channel and spatial features during upsampling. In the BraTS 2024 experiments, the full model obtains the best average Dice and HD95 among the compared methods on whole tumor and tumor core, and the ablation study indicates all three modules contribute.","pith_inferences":["If the reported advantage holds under matched baseline training, the approach could directly support clinical workflows in which contrast agent cannot be administered, since the missing-modality comparisons often involve T1ce.","The ViLA module's tokenization of the whole volume suggests the same shared-latent missing-modality design could transfer to other volumetric medical tasks, such as cardiac or abdominal multi-sequence imaging.","A natural next experiment is a controlled rerun in which every baseline is trained under the same missing-modality protocol, same augmentation, and same tuning budget; that would tell how much of the paper's margin is architecture and how much is training procedure."],"forward_implications":["Under the paper's results, a single trained model can be queried with any nonempty subset of T1, T1ce, T2, and FLAIR, so one deployment covers all missing-modality patterns.","The reconstructed MRI sequences are produced by the same network that segments the tumor, so the output could serve both automated contouring and radiologist review.","The ablation results imply the three modules are complementary, not redundant, with each module contributing to a different part of the segmentation quality.","Because both tasks share one backbone rather than being separate pipeline stages, reconstruction behavior is visible in the same output and can be audited alongside segmentation."],"supporting_citations":[{"why":"Supplies the Heteromodal Variational Encoder-Decoder backbone and the joint completion-and-segmentation formulation the paper extends.","marker":"[3]"},{"why":"Provides the Multimodal Variational Autoencoder used inside SAVE to encode each available modality under a shared latent variable.","marker":"[6]"},{"why":"Provides the Product of Gaussians fusion rule used to combine modality means and covariances while excluding missing modalities.","marker":"[8]"},{"why":"The RA-HVED baseline the paper compares against and the source of the Dimension Reduction Block used in SAVE.","marker":"[9]"},{"why":"Supplies the Vision-LSTM architecture and mLSTM blocks that power the ViLA module.","marker":"[10]"},{"why":"Supplies the squeeze-fusion-excitation co-attention mechanism on which SFECA is based.","marker":"[11]"},{"why":"Provides the BraTS 2024 dataset, preprocessing, and expert tumor labels used in all experiments.","marker":"[12]"},{"why":"The RMBTS baseline, a feature-disentanglement and gated-fusion method the paper compares against.","marker":"[13]"},{"why":"The mmformer baseline, a transformer-based incomplete multimodal learning method the paper compares against.","marker":"[14]"}],"fun_headline_variants":["One model fills missing MRI and maps tumors precisely","Missing MRI sequences? This AI rebuilds and segments","XLSTM-HVED: joint MRI reconstruction and tumor segmentation","AI reconstructs missing MRI views for accurate tumor mapping","Handles missing MRI data with a single network"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The main assumption is that the baseline methods were trained for missing modalities just as carefully as the proposed model, with the same augmentation, loss weighting, and tuning; the paper does not report those details for the baselines.","fun_headline_variants_meta":{"raw":{"variants":["One model fills missing MRI and maps tumors precisely","Missing MRI sequences? This AI rebuilds and segments","XLSTM-HVED: joint MRI reconstruction and tumor segmentation","AI reconstructs missing MRI views for accurate tumor mapping","Handles missing MRI data with a single network"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000276,"raw_usage":{"total_tokens":1654,"prompt_tokens":958,"completion_tokens":696,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":621}},"tokens_in":574,"tokens_out":696,"duration_ms":7017,"temperature":1.0,"reasoning_tokens":621,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:47:08.217374+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run all compared methods on the same BraTS 2024 split with an identical missing-modality training protocol, including the same random modality dropout schedule, augmentation, and training budget. If the average Dice and HD95 gaps reported in the paper's Table 1 drop below one point or reverse direction, the central claim that the architecture itself is better at missing-modality segmentation would not be supported.","supporting_citations":[{"cited_title":"Dataset and Implementation Details Our study utilizes the multimodal Brain Tumor Segmentation Challenge (BraTS) 2024 dataset [12]","cited_arxiv_id":null,"evidence_quote":"Supplies the Heteromodal Variational Encoder-Decoder backbone and the joint completion-and-segmentation formulation the paper extends."},{"cited_title":"A2304), Guangdong Basic and Applied Basic Research Foundation (No","cited_arxiv_id":null,"evidence_quote":"Provides the Multimodal Variational Autoencoder used inside SAVE to encode each available modality under a shared latent variable."},{"cited_title":"Missing mri pulse sequence synthesis using multi-modal generative adversarial network,","cited_arxiv_id":null,"evidence_quote":"Provides the Product of Gaussians fusion rule used to combine modality means and covariances while excluding missing modalities."},{"cited_title":"Hetero-modal vari- ational encoder-decoder for joint modality completion and segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the Vision-LSTM architecture and mLSTM blocks that power the ViLA module."},{"cited_title":"M3ae: multimodal representation learning for brain tumor segmentation with missing modalities,","cited_arxiv_id":null,"evidence_quote":"Provides the BraTS 2024 dataset, preprocessing, and expert tumor labels used in all experiments."},{"cited_title":"All the baselines and our model were trained and tested using the same backbone network to ensure consistency in the evaluation phase","cited_arxiv_id":null,"evidence_quote":"The RMBTS baseline, a feature-disentanglement and gated-fusion method the paper compares against."},{"cited_title":"Multimodal generative models for scalable weakly-supervised learning,","cited_arxiv_id":null,"evidence_quote":"The mmformer baseline, a transformer-based incomplete multimodal learning method the paper compares against."}],"review_version":1}