{"id":"ff55deb9-e494-471a-96e1-196d673748ef","arxiv_id":"2506.16784","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A new text-image brain MRI dataset and a text-guided segmentation model report improved accuracy, but the text is derived from the ground truth masks, creating a direct target leak.","lead":"This paper introduces TextBraTS, a brain MRI dataset paired with text reports, and a text-guided segmentation model that reports higher Dice scores than image-only baselines on BraTS2020. The main result is compromised because the text reports were generated from the ground truth tumor masks, giving the model target information at test time.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-time text reports are generated from ground-truth masks, so the reported segmentation gain is evidence of target leakage, not of cross-modal fusion.","rationale":"The reader's weakest assumption identifies exactly the load-bearing flaw: text reports used at test time are generated from the ground-truth tumor masks. I read the manuscript in full and confirm that Section 2 explicitly describes slicing ground truth labels into videos, feeding them to GPT-4o, and having experts correct the reports with access to the labels. Section 4 then uses these reports as test-time inputs. This is not an artifact of the review pipeline; it is the stated dataset creation procedure. Because the text contains the target locations, the Table 1 gain over image-only methods cannot be interpreted as evidence that text-guided fusion improves segmentation. The paper does have independent strengths: a publicly released volume-level text-image dataset, reproducible code, and an ablation study over fusion modules. I also note honestly that the bidirectional cross-attention module is a reasonable architectural contribution and the ablation results are internally consistent. However, the central empirical claim, stated as outperforming SOTA across all regions, depends on a contaminated evaluation. A revision that generates reports from images alone, or that excludes mask-derived location content from test reports, could support a weaker claim about text conditioning, but the current submission does not support its headline result. The verb should remain REJECT because the reported evidence for the central claim is invalid as presented, even though the dataset may be reusable after corrective disclosure and re-evaluation.","tokens_in":7227,"tokens_out":1998,"duration_ms":23461,"concrete_test":"Using the released code and dataset, retrain or re-evaluate the proposed model with test-time text reports stripped of all ground-truth-derived anatomical location words, replacing every lobe, hemisphere, and ventricle phrase with a generic token while keeping image inputs and training unchanged. If Dice on the 94-case test set falls from 85.3 toward the 84.1 image-only baseline or lower, the reported improvement is attributable to location leakage. A complementary check: assign each test image a report from a different patient and measure Dice; if it stays near 85.3, the text channel is not carrying case-specific mask information, and if it collapses, the text is being used as a label surrogate.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that text guidance improves volumetric brain tumor segmentation. This claim rests on test-time use of text reports whose content was derived from the ground-truth segmentation masks. Section 2 states: 'the corresponding tumor ground truth labels were sliced and converted into videos. These were then input into the GPT-4o model' and experts later corrected the reports 'based on the images and segmentation labels.' Section 4 then feeds these same reports into the model at test time. The text therefore encodes the answer (tumor location and tissue characteristics) before the model sees the image. The comparison against image-only baselines in Table 1 thus does not measure fusion of independent information; it measures how well the model can decode a textual description of the target mask. The location-only ablation in Table 2 strengthens this concern, since location phrases alone nearly reproduce the full gain of 85.3 Dice, consistent with the text acting as a compressed mask annotation rather than as a radiological report that would be available at inference. This is target leakage, not a modeling failure in the fusion module itself. The dataset contribution may still be valuable for other purposes, but the paper's headline segmentation improvement is not supported by the current experimental design.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces TextBraTS, a volume-level dataset pairing BraTS2020 MRI volumes with expert-refined textual reports, and proposes a text-guided 3D brain tumor segmentation model that fuses text and image features with bidirectional cross-attention. The authors report that their method achieves an average Dice of 85.3 on their test split, outperforming several image-only baselines (e.g., 84.1 for nnU-Net and NestedFormer), and they conclude that text guidance significantly improves segmentation accuracy. The central claim is that integrating textual reports with imaging data improves volumetric brain tumor segmentation.","tokens_in":7550,"tokens_out":3731,"duration_ms":40054,"significance":"If the claimed result were valid, the public TextBraTS dataset, code, and pretrained models would be a useful contribution to multimodal medical image segmentation, and the proposed fusion baseline could inform future work on text-guided volumetric segmentation. The authors deserve credit for releasing these resources. However, the experimental design makes the headline claim circular: the text reports were constructed from the ground-truth segmentation masks, so at test time the text input encodes the answer. The reported gains over image-only baselines are therefore evidence of target leakage rather than of effective cross-modal fusion. The dataset may still be valuable for other purposes, such as radiological report generation or segmentation under a protocol where reports are produced without access to the masks, but the central segmentation improvement claimed in the paper is not established by the current evidence.","major_comments":[{"comment":"The text reports are generated directly from the ground-truth masks and then used as test-time inputs. Section 2 states that 'the corresponding tumor ground truth labels were sliced and converted into videos. These were then input into the GPT-4o model,' and that experts refined the reports 'based on the images and segmentation labels.' Section 4 then feeds these same reports into the model at test time. This is target leakage: the text is a compressed description of the target mask, so the model's task reduces to decoding the answer from the text rather than learning to segment from images. The Table 1 improvement over image-only baselines is therefore not evidence of multimodal fusion. The location-only row in Table 2 (Dice 84.6) nearly matching the full-text row (Dice 85.3) is consistent with this interpretation, since the location phrases explicitly encode tumor position from the ground truth.","section":"Section 4, Table 1"},{"comment":"The comparison with SOTA methods is not adequately controlled. The paper does not state whether the baselines (3D-UNet, nnU-Net, SegResNet, SwinUNETR, NestedFormer) were retrained on the same 220/55/94 split with the same fixed 128x128x128 resampling, the same checkpoint selection, and the same post-processing protocol. If the baseline numbers are taken from other publications or trained under different conditions, the comparison is not valid. The reported t-test (p=0.0077) is against NestedFormer only, and no confidence intervals are given for any method. Most importantly, there is no ablation of the proposed architecture without the text branch, which would be the necessary image-only control for measuring the contribution of text.","section":"Section 4, Table 1"},{"comment":"The fusion-module ablations are confounded by the same leakage. Since every condition in Table 3 receives text derived from the ground-truth mask, the comparison among dot-sum, one-step cross-attention, and bidirectional cross-attention measures how well each module can decode the answer-bearing text, not how well it fuses independent radiological information. The conclusion that bidirectional cross-attention is 'effective' is thus not supported as a general claim about multimodal fusion. A valid evaluation would require text that is generated without access to the masks, or an explicit comparison where the text is a genuine radiological report available at inference time.","section":"Section 3 and Table 3"}],"minor_comments":[{"comment":"The abstract refers to a 'sequential cross-attention method,' while Section 3 presents a 'bidirectional cross-attention' mechanism; please align the terminology throughout the paper.","section":"Abstract and Section 3"},{"comment":"Table 2 has the typo 'Ablaion study' and Table 1 contains an extraneous space in 'T able'; these should be corrected.","section":"Tables 1 and 2"},{"comment":"The sentence 'in the field of brain tumor segmentation, no similar text-image dataset and text-image multimodal segmentation methods' is incomplete; please revise to a full sentence.","section":"Section 2, last paragraph"},{"comment":"The caption states that template words, location descriptions, and features are colored red, green, and blue, but the report text in the figure is not visibly color-coded in the manuscript; please colorize the example or adjust the caption.","section":"Figure 3"},{"comment":"Equation (4) has a mismatched parenthesis in the Softmax expression, and the shape of the intermediate feature f'_i (token_num x 768) versus the claimed output f_joint (H/32 x W/32 x D/32 x 768) is not explicitly derived; please clarify the tensor dimensions.","section":"Equations (4) and (6)"}],"recommendation":"reject","confidential_remarks":"The target leakage in the dataset construction is fundamental to the paper's central claim. The segmentation improvements in Tables 1 and 2 are better interpreted as the model decoding ground-truth-derived text, not as successful multimodal fusion. A redesigned study would need to generate reports without mask access, retrain baselines under identical protocols, and include a text-free control of the proposed architecture. These changes go beyond a standard revision, so I recommend rejection of the current manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this one: the dataset is genuinely new, and the headline result is not to be believed. The text reports used at test time are generated from the ground truth tumor masks, so the model is essentially reading a description of the answer.\n\nWhat is actually new and good: TextBraTS is the first volume-level text-image brain tumor dataset, 369 cases from BraTS2020 with expert-refined radiological reports. The creation pipeline — GPT-4o pre-annotation plus expert correction — is a reasonable way to build such a resource, and they release code and models. The ablation work on template formats and fusion modules is thorough, and reporting HD95 alongside Dice is a plus.\n\nThe soft spot is not soft at all. Section 2 explicitly says ground truth labels were sliced into videos and fed to GPT-4o, and experts then refined the reports based on the images and segmentation labels. Section 4 feeds those same reports into the model at test time. That is textbook target leakage. The comparison against image-only baselines in Table 1 therefore does not measure fusion of independent information; it measures how well the model decodes a textual summary of the mask. The location-only ablation in Table 2 confirms this: location phrases alone almost reproduce the full gain (84.6 vs 85.3 Dice), consistent with the text acting as a compressed mask annotation.\n\nMinor issues are easy to list: no confidence intervals in Table 1, unclear whether baselines were retrained on the identical split and protocol, and the t-test (p=0.0077) is irrelevant once the inputs leak the target. But those are secondary to the leakage, which invalidates the central claim.\n\nIs anything salvageable? The fusion module itself is an incremental but sensible adaptation of cross-attention; the comparison across fusion strategies could be informative if the text input were not generated from the labels. The dataset might be useful for other purposes, e.g., as a testbed for annotation-conditioned segmentation, but not for the claim of text-guided improvement without a redesigned evaluation.\n\nWho is this for? People working on multimodal medical segmentation, and anyone looking for a clean example of dataset leakage in a supposed state-of-the-art comparison. It deserves a serious referee only because the dataset is a plausible community resource; the current manuscript, however, should not be accepted. I would ask the authors to either generate test-time text without access to masks (e.g., from images only), or reframe the task as mask-conditioned annotation and stop claiming external guidance. Then report baselines on the identical split with confidence intervals.\n\nRecommendation: send to peer review only with the expectation of major revision; desk rejection would also be defensible given the load-bearing flaw.","headline":"The TextBraTS dataset is a real contribution, but the reported segmentation gains come from test-time leakage: the text reports are written from the ground-truth masks, so the paper overstates what text-guided fusion achieves.","tokens_in":7949,"tokens_out":2210,"would_cite":false,"duration_ms":22137,"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":"Adding volume-level radiological text reports to MRI input improves brain tumor segmentation beyond image-only baselines, reaching an average Dice of 85.3.","keywords":["text-guided segmentation","multimodal fusion","brain tumor MRI","volumetric segmentation","cross-attention","dataset construction","radiology report templates"],"falsifier":"Re-run the exact test procedure after regenerating the reports from the MRI images alone, without seeing the segmentation masks, either by prompting the multimodal model directly on the 3D Flair volumes or by having radiologists write reports blinded to the labels; if the average Dice advantage over the image-only baseline disappears or shrinks below the t-test threshold, the gain is attributable to mask-derived text rather than to text-guided fusion.","tokens_in":7031,"feed_emoji":"🧠","tokens_out":6775,"duration_ms":60224,"temperature":0.7,"pith_summary":"This paper argues that combining volume-level radiological text reports with MRI input can improve brain tumor segmentation beyond image-only state of the art. To test this, the authors build TextBraTS, a multimodal dataset pairing 369 BraTS2020 MRI volumes with expert-refined textual reports, and propose a text-guided segmentation model whose bidirectional cross-attention block aligns text and image features. On the test split the model reaches an average Dice of 85.3%, ahead of the best image-only baseline at 84.1%, with a statistically significant t-test against NestedFormer. The paper also shows that fully templated reports combining location and feature descriptions, and the bidirectional fusion design, are what deliver the gain. A careful reader should note that the text reports used at test time were generated with access to the ground-truth segmentation masks, which bears on what the comparison establishes.","feed_headline":"Adding radiology text lifts MRI tumor segmentation to 85.3 Dice","feed_subtitle":"TextBraTS pairs 369 MRI volumes with expert reports; a bidirectional attention block turns them into a segmentation gain.","key_machinery":"The load-bearing component is the bidirectional cross-attention fusion block: a text-to-image attention pass (text as query, image as key and value) followed by an image-to-text pass (image as query, refined text-guided features as key and value), with normalization layers in between. This two-step mechanism lets the textual report reshape the image feature map before decoding. Equally central is the templated report formulation, where location phrases and feature phrases are presented in a fixed structure; the ablations show these templated inputs, rather than raw text, are what make the fusion effective.","core_discovery":"The paper's central claim is that a text-guided segmentation network can outperform vision-only methods on volumetric brain tumor segmentation when trained and tested with paired radiological reports. The authors introduce TextBraTS, which they describe as the first publicly available volume-level text-image brain MRI dataset, with 369 cases; each report is produced by a large multimodal model and then corrected by expert radiologists against the images and segmentation labels. The proposed architecture combines a transformer-based image encoder with a biomedical language model, and a bidirectional cross-attention fusion block first uses text to refine image features, then uses those refined features to re-attend to the image features, producing a fused representation fed to the decoder. Experiments report a mean Dice of 85.3% (ET 83.3, WT 89.9, TC 82.8) and a 95th-percentile Hausdorff distance of 5.13 mm, improving over all compared baselines, and a t-test (p=0.0077) against NestedFormer over ten runs. Ablations show that templated text containing both location and feature phrases outperforms raw or partial text, and that bidirectional cross-attention outperforms single-direction cross-attention and dot-sum fusion.","pith_inferences":["The evaluation protocol uses reports derived from ground-truth masks, so the reported gain likely reflects how well the model can read location information that is already implicit in the segmentation labels; a test with reports written from images alone would separate the value of genuine radiological description from label leakage.","The bidirectional fusion design could transfer to other volumetric medical imaging tasks that have structured reports, such as chest CT or ultrasound, where location-feature templates are available.","A direct extension would be to train the same image-only backbone with the identical 200-epoch schedule and report Dice; the paper compares against published baselines rather than an equally trained SwinUNETR without text, so part of the gain could be training configuration rather than text guidance."],"forward_implications":["If text-guided fusion holds, adding volume-level radiology reports to brain tumor segmentation yields a Dice gain of about 1.2 points on average and a substantially lower edge error (HD95 of 5.13 mm vs. 7.07 mm for the image-only SwinUNETR baseline), which matters for boundary-sensitive downstream tasks like surgical planning.","TextBraTS provides a public benchmark with paired MRI volumes and expert-refined textual reports, enabling direct comparisons of text-image fusion strategies on volumetric brain tumor data.","Fully templated reports containing both location and feature information are the recommended input format; raw text or partially templated reports measurably reduce the segmentation benefit.","Bidirectional cross-attention is a better fusion strategy than one-directional cross-attention or dot-sum fusion, implying that the text and image branches both need to be iteratively aligned.","The reported improvement is consistent across the three tumor sub-regions (whole tumor, enhancing tumor, tumor core), not just on one region."],"supporting_citations":[{"why":"Supplies the base BraTS2020 MRI volumes and segmentation labels from which TextBraTS is built.","marker":"[15]"},{"why":"The large multimodal model used to generate the initial pseudo-reports that experts later refine.","marker":"[17]"},{"why":"Provides the Transformer-based image encoder backbone (SwinUNETR) that the text-guided model extends.","marker":"[6]"},{"why":"Provides the biomedical language model (BioBERT) used to extract text features from the reports.","marker":"[11]"},{"why":"Serves as a strong image-only segmentation baseline (nnU-Net) that the method must beat.","marker":"[10]"},{"why":"Serves as the classic 3D-UNet baseline for the image-only comparison.","marker":"[21]"},{"why":"Serves as the SegResNet baseline in the comparison table.","marker":"[7]"},{"why":"Serves as the NestedFormer baseline whose average Dice the method surpasses, and the comparison used for the statistical test.","marker":"[24]"}],"fun_headline_variants":["Text-guided fusion boosts brain tumor segmentation to 85.3 Dice","New text-image dataset improves MRI tumor segmentation","Bi-directional attention with text enhances brain tumor Dice","First volume-level text-MRI dataset lifts segmentation to 85.3","Text prompts sharpen brain tumor segmentation to 85.3 Dice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the volume-level text reports used at test time, which were generated with access to the ground-truth tumor masks and then refined by experts against those labels (Section 2), are valid inputs for measuring segmentation accuracy; if they leak mask-derived location information, the reported gain over image-only models would be inflated.","fun_headline_variants_meta":{"raw":{"variants":["Text-guided fusion boosts brain tumor segmentation to 85.3 Dice","New text-image dataset improves MRI tumor segmentation","Bi-directional attention with text enhances brain tumor Dice","First volume-level text-MRI dataset lifts segmentation to 85.3","Text prompts sharpen brain tumor segmentation to 85.3 Dice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00029,"raw_usage":{"total_tokens":1725,"prompt_tokens":1004,"completion_tokens":721,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":638}},"tokens_in":620,"tokens_out":721,"duration_ms":6242,"temperature":1.0,"reasoning_tokens":638,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:18:05.992184+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the exact test procedure after regenerating the reports from the MRI images alone, without seeing the segmentation masks, either by prompting the multimodal model directly on the 3D Flair volumes or by having radiologists write reports blinded to the labels; if the average Dice advantage over the image-only baseline disappears or shrinks below the t-test threshold, the gain is attributable to mask-derived text rather than to text-guided fusion.","supporting_citations":[{"cited_title":"In: Crimi, A., Bakas, S","cited_arxiv_id":null,"evidence_quote":"Serves as the SegResNet baseline in the comparison table."},{"cited_title":"In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F","cited_arxiv_id":null,"evidence_quote":"Serves as the classic 3D-UNet baseline for the image-only comparison."},{"cited_title":"IEEE transactions on medical imaging34(10), 1993–2024 (2014)","cited_arxiv_id":null,"evidence_quote":"Supplies the base BraTS2020 MRI volumes and segmentation labels from which TextBraTS is built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The large multimodal model used to generate the initial pseudo-reports that experts later refine."},{"cited_title":"In: Crimi, A., Bakas, S","cited_arxiv_id":null,"evidence_quote":"Provides the Transformer-based image encoder backbone (SwinUNETR) that the text-guided model extends."},{"cited_title":"Bioinformatics36(4), 1234–1240 (2020)","cited_arxiv_id":null,"evidence_quote":"Provides the biomedical language model (BioBERT) used to extract text features from the reports."},{"cited_title":"In: Wang, L., Dou, Q., Fletcher, P.T., Speidel, S., Li, S","cited_arxiv_id":null,"evidence_quote":"Serves as the NestedFormer baseline whose average Dice the method surpasses, and the comparison used for the statistical test."}],"review_version":1}