{"id":"25d96c8a-406e-464a-a1c0-96adffdc75f2","arxiv_id":"2607.17782","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A single 3D MAE-pretrained xLSTM-UNet encoder transfers to infarct classification, meningioma segmentation, and brain-age estimation, taking second place on the FOMO 2025 Method Track.","lead":"BrainNext is a self-supervised 3D brain MRI model that combines masked autoencoder pretraining with a bidirectional xLSTM-UNet architecture. It ranked second on the FOMO 2025 benchmark across classification, segmentation, and brain-age regression, suggesting one pretrained encoder can serve multiple tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing from-scratch and standard-architecture baselines: the reported FOMO rank cannot be attributed to MAE pretraining, so the general-purpose foundation-model claim is not yet supported.","rationale":"The paper's intended contribution is a single self-supervised encoder that generalizes across brain MRI tasks. For that claim to hold, the official FOMO result must be attributable to the pretrained representation and robust enough to support the word 'general-purpose.' The weakest point is the absence of any comparison that isolates pretraining: a random-init version of the same architecture, or a standard 3D U-Net, would separate architecture and tuning effects from self-supervised learning. The FOMO challenge summary quoted in Section 2.3 cautions that MAE objectives favor segmentation and that no single objective wins all tasks; combined with the low absolute metrics, the reported 2nd/1st ranks are compatible with a task-specific outcome rather than a general representation. The paper deserves credit for participating in an official challenge and reporting the leaderboard, and the Section 6 limitation statement is honest about single-benchmark evaluation. However, an official rank without margins or baselines does not establish the causal role of self-supervised pretraining. The dataset-count and channel-handling inconsistencies further reinforce the need to release exact preprocessing and weights. A random-init control with confidence intervals would settle the attribution; if the results survive that check, the conditional acceptance can stand. This is not an accusation of misconduct, only a request for the minimal control that the central claim requires.","tokens_in":9527,"tokens_out":10190,"duration_ms":97224,"concrete_test":"Run the exact fine-tuning protocol of Section 3.6 (same five-fold splits, augmentation schedule, optimizer, epochs, and test-time ensembling) with the BrainNext encoder weights randomly initialized instead of MAE-pretrained, and evaluate all three FOMO tasks on the official test set. Compute bootstrap confidence intervals for each Table 3 metric from the test predictions. If the random-init model matches or exceeds BrainNext on any task, or if the Table 3 values fall within the bootstrap noise, then the self-supervised pretraining is not load-bearing for the reported ranking; the paper should then claim only that the architecture transfers under this protocol, not that MAE pretraining produces a general-purpose foundation model.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 5 reports only BrainNext's absolute FOMO test metrics (AUROC 0.684, DSC/NSD 0.261/0.232, MAE 12.67) and its official rank. There is no within-paper control condition: no randomly initialized encoder, no standard 3D U-Net, no MAE with a convolutional backbone, and no per-task leaderboard margins or confidence intervals. For the central claim to hold, the observed 'second place overall / first in meningioma segmentation' must be caused by the self-supervised volumetric representations and must be robust across classification, segmentation, and regression. The paper's own Section 2.3 states that no single SSL objective wins all FOMO tasks and that MAE-style objectives favor segmentation; a second-place overall driven by the segmentation rank would not establish general-purpose transfer. The Section 6 limitation paragraph concedes that evaluation is exclusively within FOMO, but the more immediate gap is internal: without a from-scratch control, the official rank is not evidence about MAE pretraining or about the xLSTM-UNet design. The low DSC value further means 'first place' may reflect a hard task where all methods cluster near chance, not a strong representation. Secondary reproducibility issues, including 60,551 vs 60,529 pretraining volumes (Section 3.2 vs Section 2.3) and 'multiple sequences as separate input channels' (Section 3.4) vs 'each MRI sequence treated independently' (Section 2.4), should also be resolved if the model is claimed to be the exact FOMO60K model.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BrainNext, a volumetric brain MRI foundation model combining masked autoencoder (MAE) pretraining with a native 3D Bi-Directional xLSTM-UNet architecture. The model is pretrained on the FOMO60K dataset (reported as 60,551 or 60,529 examinations in different sections) and then fine-tuned for three downstream tasks in the FOMO 2025 Method Track: infarct classification, meningioma segmentation, and brain-age estimation. The authors report official leaderboard results of second place overall and first place in meningioma segmentation, with AUROC 0.684, DSC/NSD 0.261/0.232, and MAE/CORR 12.67/0.448. The paper claims that a single pretrained encoder can transfer across classification, segmentation, and regression, establishing BrainNext as a general-purpose foundation model.","tokens_in":9877,"tokens_out":5097,"duration_ms":46176,"significance":"If the claims are well supported, the paper makes a useful contribution by demonstrating a native 3D self-supervised foundation model for brain MRI and by providing results under the standardized FOMO 2025 evaluation protocol. The use of an official blind test set and a common pretraining dataset is a clear strength, and the plan to release model weights is commendable. However, the current evidence is insufficient to attribute the reported performance to MAE pretraining or to the proposed xLSTM-UNet architecture, because the paper presents no within-manuscript controls, no per-task comparison with other leaderboard entries, and no statistical uncertainty. The modest absolute metric values and the challenge's own finding that MAE favors segmentation further temper the 'general-purpose' claim. The significance is therefore conditional on an additional round of analysis.","major_comments":[{"comment":"The central claim that BrainNext's self-supervised pretraining and architecture yield generalizable representations is not supported by any within-paper control. The paper reports only its own absolute test-set metrics (AUROC 0.684, DSC 0.261, NSD 0.232, MAE 12.67, CORR 0.448) and an aggregate rank. There is no randomly initialized encoder, no standard 3D U-Net or UNETR baseline, no MAE without the xLSTM component, and no per-task scores of other participants. Consequently, the observed leaderboard position cannot be attributed to MAE pretraining or to the Bi-Directional xLSTM-UNet design; the official rank alone is not evidence for the proposed method's components. Please add at least a from-scratch control and direct per-task leaderboard comparisons, including margins to the best and median methods.","section":"§5.1, Table 3"},{"comment":"The size of the pretraining dataset is inconsistent across the manuscript. Section 2.3 states that FOMO60K comprises 60,529 MRI scans, while the Abstract, Section 3.2, Section 4.1, and the description of Figure 1 state 60,551. If the discrepancy reflects the difference between scans and examinations, or between raw and preprocessed volumes, this should be stated explicitly. As written, the two numbers are presented as the same dataset without clarification, which is a reproducibility issue.","section":"§2.3 vs §3.2"},{"comment":"The handling of multiple MRI sequences is described contradictorily. Section 2.4 says 'each MRI sequence treated independently during pretraining,' while Section 3.4 says 'Multiple MRI sequences are treated as separate input channels, allowing the network to jointly learn complementary information across imaging modalities.' These two statements define different input tensors (single-channel volume processed separately per sequence versus multi-channel volume processed jointly). Please clarify the actual input representation and unify the two descriptions, as this affects the architecture and the pretraining objective.","section":"§2.4 vs §3.4"},{"comment":"The paper reports no error bars or confidence intervals for any metric, despite the five-fold cross-validation and ensembling described in Section 3.6. Without uncertainty estimates, the reader cannot assess whether DSC 0.261 or MAE 12.67 is stable. In addition, Section 2.3 itself notes that no single SSL objective wins all FOMO tasks and that MAE-style objectives favor segmentation; this paper reports second overall with a first-place segmentation score but a modest AUROC (0.684) and MAE (12.67). To substantiate the 'general-purpose' claim, please report the official per-task rankings and per-task scores of competing methods, and discuss explicitly whether the overall rank is driven predominantly by the segmentation result.","section":"§5, Table 3; §2.3"}],"minor_comments":[{"comment":"The reference to 'the authors’ previously published BrainFound [15]' is incorrect: BrainFound is introduced as [14] in Section 2.2, while [15] is cited for BrainIAC. Please fix the citation.","section":"§2.4"},{"comment":"Several implementation details are missing or imprecise: the masking ratio for MAE, the patch size used during fine-tuning, the fine-tuning learning rate schedule, and the per-task head architecture are not specified. The sentence 'He . initialization' contains a typo and should read 'He initialization.'","section":"§3.6"},{"comment":"If Figure 2 is a screenshot of the official leaderboard, the resolution or font size in the manuscript may make team names and scores unreadable. Please provide a table with the top entries' per-task and overall scores so that the comparison is legible and machine-readable.","section":"§5.1, Figure 2"},{"comment":"Several references contain formatting errors, such as a double period at the end of entries [10] and [11], and reference [29] should include the date of access consistently with the URL. Please proofread the reference list.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a competition report based on the official FOMO 2025 leaderboard. The core scientific claim—that MAE pretraining with a novel 3D architecture produces a general-purpose foundation model—is plausible but currently under-subsidized by evidence within the manuscript. The absence of any ablation or baseline, together with the dataset-size and input-representation inconsistencies, requires a substantive revision rather than a minor one. If the authors can add within-manuscript controls or, at minimum, detailed per-task leaderboard comparisons and uncertainty estimates, the paper could become an acceptable challenge-style contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take on arXiv:2607.17782. BrainNext is a real benchmark entry: MAE pretraining on 60k brain MRIs with a native 3D Bi-Directional xLSTM-UNet, evaluated on the official FOMO 2025 Method Track, second overall and first on meningioma segmentation. That combination is new, and the evaluation is on a blind test set, so the circularity burden is genuinely low. Credit where due: this is a serious engineering effort and the authors are transparent about the FOMO-only evaluation in the limitations.\n\nBut the core claim—that this is a general-purpose foundation model with strong transferability—is not backed by the evidence presented. The paper reports only the absolute FOMO test metrics (AUROC 0.684, DSC 0.261, MAE 12.67 years) and the leaderboard rank. There are no error bars, no ablations, no comparison to a randomly initialized encoder or a standard 3D U-Net, and no per-task leaderboard margins. Without a from-scratch control, the second-place finish cannot be attributed to MAE pretraining or to the xLSTM-UNet design. The DSC of 0.261 is low enough that first place in meningioma segmentation may reflect a hard task where all methods cluster near chance, not a strong representation. The paper's own Section 2.3 notes that no single SSL objective wins all FOMO tasks and that MAE favors segmentation; a rank driven by segmentation would not establish general-purpose transfer. The stress-test note is on target: the missing internal baselines are the load-bearing gap.\n\nThere are also small but telling inconsistencies: Section 3.2 says 60,551 pretraining volumes and Section 2.3 says 60,529, and Section 3.4 says multiple MRI sequences are separate input channels while Section 2.4 says each sequence is treated independently. These should be fixed if the model is claimed to be exactly the FOMO60K model. The paper says weights and code will be released, but they are not yet available, which limits reproducibility.\n\nWho is this for? Anyone working on brain MRI foundation models will want to know about this architecture and its FOMO results. It is a useful data point, but it is a system paper with a benchmark result, not a demonstration of a general-purpose foundation model.\n\nI would send it to peer review, but with major revision: add from-scratch and standard-architecture baselines, report confidence intervals, clarify the data inconsistencies, and temper the claims accordingly. It deserves referee time, but the current version should not be accepted as is.","headline":"A legitimate FOMO challenge entry with a new MAE plus 3D xLSTM-UNet recipe, but the general-purpose foundation-model claim outruns the evidence because the paper lacks within-paper baselines, ablations, and error bars.","tokens_in":10378,"tokens_out":1470,"would_cite":false,"duration_ms":14592,"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":"BrainNext is a single self-supervised 3D encoder that transfers across brain-MRI classification, segmentation, and regression, ranking second overall and first on meningioma segmentation in the FOMO 2025 Method Track.","keywords":["Brain Foundation Models","Brain MRI","Self-Supervised Learning","Masked Autoencoders","Bi-Directional xLSTM-UNet","Neuroimaging","Volumetric Representation Learning"],"falsifier":"Run the same three downstream tasks with the identical fine-tuning pipeline but replace the pretrained encoder with a randomly initialized one of the same architecture. If the from-scratch model matches or beats BrainNext's AUROC of 0.684, Dice of 0.261, and brain-age MAE of 12.67 years, then the reported transfer is not caused by MAE pretraining. A second check is external validation: fine-tune the released weights on an independent brain MRI dataset and compare against a from-scratch baseline under matched label counts.","tokens_in":9379,"feed_emoji":"🧠","tokens_out":10941,"duration_ms":84369,"temperature":0.7,"pith_summary":"BrainNext aims to show that one self-supervised encoder, pretrained on unlabeled volumetric brain MRI, can transfer across the three main families of medical image analysis: classification, segmentation, and regression. The authors pretrain a 45-million-parameter 3D Bi-Directional xLSTM-UNet with a masked autoencoder objective on 60,551 brain MRI examinations, then fine-tune the same encoder for infarct detection, meningioma segmentation, and brain-age prediction. The evidence is the official FOMO 2025 Method Track: BrainNext placed second overall and first on meningioma segmentation. If the result holds, it matters because a single reusable encoder could reduce the need for task-specific models and large annotated datasets in neuroimaging. The paper's contribution is therefore a demonstration, under a standardized protocol, that native-3D masked pretraining at this scale generalizes across tasks.","feed_headline":"One pretrained brain-MRI model takes second on a three-task benchmark","feed_subtitle":"Self-supervised 3D pretraining on 60,551 scans wins meningioma segmentation and transfers to classification and age regression.","key_machinery":"The load-bearing object is the native 3D Bi-Directional xLSTM-UNet, a five-stage encoder-decoder in which each encoder stage combines 3D convolutional layers with bidirectional xLSTM blocks that read the volume forward and backward, fuse their outputs through a learnable projection with a residual connection ($H_l = F_l + \\phi([\\vec H_l, \\overleftarrow H_l])$), and feed the decoder through skip connections. During pretraining, the model is treated as an MAE: input volumes are split into $96 \\times 96 \\times 96$ voxel patches, a random subset $\\mathcal{M}$ is masked, and the loss is the mean squared reconstruction error over masked patches only. This objective forces the encoder to represent global anatomical structure from visible context, while the xLSTM blocks supply long-range volumetric dependencies that plain convolutions would miss. After pretraining, the decoder is discarded and task-specific heads are attached to the encoder.","core_discovery":"The paper's central claim is that large-scale masked autoencoder pretraining in native 3D produces a single encoder whose lightweight fine-tuning is competitive across classification, segmentation, and regression. On the official FOMO 2025 Method Track test set, BrainNext reports an AUROC of 0.684 for infarct classification, a Dice of 0.261 and Normalized Surface Dice of 0.232 for meningioma segmentation, and a mean absolute error of 12.67 years with correlation 0.448 for brain-age estimation, ranking second overall and first on meningioma segmentation. The same pretrained encoder supports all three tasks via task-specific heads; the reconstruction decoder is discarded after pretraining. These numbers are the paper's evidence that transferable anatomical representations emerge from unlabeled volumetric data at this scale.","pith_inferences":["Inference, not the paper's claim: the 12.67-year brain-age MAE is likely inflated by the out-of-domain, limited-label protocol; a matched external evaluation would reveal whether the same encoder approaches errors typical of dedicated brain-age models.","Inference: because no ablation separates MAE pretraining from the xLSTM-UNet backbone, the relative contributions of the objective and the architecture remain entangled; a two-factor ablation would isolate them.","Inference: a natural extension is label-efficiency testing—freeze the encoder, train heads on small labelled subsets, and measure how much annotation the pretrained representation actually saves."],"forward_implications":["A single pretrained 3D encoder can be adapted to classification, segmentation, and regression through lightweight task-specific heads, so new brain MRI tasks may not require training a full model from scratch.","MAE-style masked reconstruction at 60,551 scans is sufficient to make the encoder competitive on a common benchmark, supporting the use of unlabeled hospital-scale MRI collections for pretraining.","Because BrainNext is natively 3D, the learned representations carry volumetric context that slice-based models discard, which is directly relevant to anatomy-dependent tasks such as meningioma segmentation.","The planned public release of weights and inference code would make BrainNext a reusable starting point for downstream neuroimaging studies."],"supporting_citations":[{"why":"supplies the masked autoencoder objective that BrainNext's pretraining stage minimizes.","marker":"[11]"},{"why":"supplies the bidirectional xLSTM blocks that capture long-range volumetric dependencies in the encoder.","marker":"[23]"},{"why":"supplies the U-Net encoder-decoder topology with skip connections that BrainNext adapts to 3D.","marker":"[20]"},{"why":"defines the FOMO 2025 benchmark protocol and the FOMO60K pretraining pool used for evaluation.","marker":"[28]"},{"why":"is the official leaderboard source for the reported second-place overall and first-place meningioma segmentation results.","marker":"[29]"}],"fun_headline_variants":["Self-supervised brain MRI model takes first in meningioma segmentation","One 3D brain MRI model excels at three tasks after pretraining","BrainNext: 60k scans pretraining yields top meningioma segmentation","Volumetric brain MRI foundation model ranks second on FOMO 2025","General-purpose brain MRI model wins one task, places second overall"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the official FOMO 2025 leaderboard is a complete, unbiased measure of cross-task generalization for brain MRI foundation models; if that single benchmark is unrepresentative or unfair, the claim that BrainNext transfers across tasks weakens.","fun_headline_variants_meta":{"raw":{"variants":["Self-supervised brain MRI model takes first in meningioma segmentation","One 3D brain MRI model excels at three tasks after pretraining","BrainNext: 60k scans pretraining yields top meningioma segmentation","Volumetric brain MRI foundation model ranks second on FOMO 2025","General-purpose brain MRI model wins one task, places second overall"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000421,"raw_usage":{"total_tokens":2164,"prompt_tokens":942,"completion_tokens":1222,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":1127}},"tokens_in":558,"tokens_out":1222,"duration_ms":7932,"temperature":1.0,"reasoning_tokens":1127,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:33:41.466393+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same three downstream tasks with the identical fine-tuning pipeline but replace the pretrained encoder with a randomly initialized one of the same architecture. If the from-scratch model matches or beats BrainNext's AUROC of 0.684, Dice of 0.261, and brain-age MAE of 12.67 years, then the reported transfer is not caused by MAE pretraining. A second check is external validation: fine-tune the released weights on an independent brain MRI dataset and compare against a from-scratch baseline under matched label counts.","supporting_citations":[{"cited_title":"Looking into the functional architecture of the brain with diffusion MRI","cited_arxiv_id":null,"evidence_quote":"supplies the masked autoencoder objective that BrainNext's pretraining stage minimizes."},{"cited_title":"Self -supervised pre -training of swin transformers for 3d medical image analysis","cited_arxiv_id":null,"evidence_quote":"supplies the bidirectional xLSTM blocks that capture long-range volumetric dependencies in the encoder."},{"cited_title":"U -net: Convolutional networks for biomedical image segmentation","cited_arxiv_id":null,"evidence_quote":"supplies the U-Net encoder-decoder topology with skip connections that BrainNext adapts to 3D."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines the FOMO 2025 benchmark protocol and the FOMO60K pretraining pool used for evaluation."},{"cited_title":"An OpenMind for 3D medical vision self -supervised learning","cited_arxiv_id":null,"evidence_quote":"is the official leaderboard source for the reported second-place overall and first-place meningioma segmentation results."}],"review_version":1}