{"id":"5305b0f9-c800-4466-b87f-fdde135cbf3e","arxiv_id":"2508.10950","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":2,"one_line_summary":"FastFOD-Net uses deep learning to turn routine clinical diffusion MRI into research-quality fiber orientation estimates, validated across healthy and six disease groups.","lead":"This paper presents FastFOD-Net, a deep learning model that enhances fiber orientation estimates from clinical brain scans, and validates it across healthy controls and six neurological diseases. It claims performance comparable to high-quality research scans, at 60 times the speed of its predecessor, which could make advanced brain wiring analysis practical in clinics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Comparability claim rests on train/test independence and reference FOD ground truth; abstract alone does not rule out circular evaluation or domain shift.","rationale":"The reader's verdict is UNVERDICTED due to the corrupted full text, and the weakest assumption is exactly that enhanced FODs faithfully represent true fiber architecture and that clinical populations are well represented. My concern matches: the abstract's strong claim of comparability with research acquisitions is unsupported without evidence of an unbiased gold standard and proper train/test separation. The provided full text is unreadable, so no internal details can be checked; thus the correct verdict remains unchanged at UNVERDICTED. The concrete test would settle the concern if the manuscript becomes readable or the code/data are released, but for now the concern is genuinely load-bearing and cannot be dismissed.","tokens_in":3332,"tokens_out":2988,"duration_ms":37655,"concrete_test":"Inspect the released code/data or, if unavailable, request from the authors: (1) confirm that training and evaluation were split by subject, with no imaging sessions from the same individual in both; (2) confirm that the evaluation metric compares enhanced FODs against an independently generated reference—e.g., high-quality multi-shell acquisitions fitted with a different algorithm, or a tractography/connectome metric—rather than against the same FODs used as training targets. If either condition is violated, the comparability claim is inflated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that FastFOD-Net enables analysis comparable to research-quality acquisitions across six disorders—requires that (a) the research FODs used as training targets are unbiased ground truth, and (b) the model generalizes to clinical populations distinct from its training distribution. The abstract provides no information about the subject-wise train/test split, whether the evaluation metric is identical to the training loss (e.g., voxel-wise MSE against the same research FODs used as targets), or whether the six disorder cohorts were held out. If the same subjects or acquisition sessions appear in both training and evaluation, or if the reference FODs are the very same regression targets, then 'comparable' becomes circular. Moreover, if the model was trained predominantly on healthy controls, pathological conditions—edema, atrophy, lesions—represent a distribution shift that can degrade FOD fidelity. The full text is corrupted/unreadable, so these methodological details cannot be audited. This is the most load-bearing concern because the entire clinical utility rests on it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces FastFOD-Net, a deep-learning framework for enhancing fiber orientation distributions (FODs) from clinical diffusion MRI. The abstract claims that the method delivers superior FOD enhancement, is 60x faster in training/inference than its predecessor, and is validated across healthy controls and six neurological disorders, enabling analysis comparable to that of high-quality research acquisitions. The abstract also claims reduced sample size requirements and increased clinical applicability. However, the full text of the submitted PDF is corrupted and unreadable; no methods, cohort descriptions, experimental results, tables, figures, or statistical analyses are accessible. The abstract itself contains no quantitative supporting data.","tokens_in":3619,"tokens_out":4311,"duration_ms":50914,"significance":"If the claims are substantiated, the work would be significant: a clinically validated, fast deep-learning FOD enhancement method applicable to multiple neurological disorders would address a well-recognized barrier to clinical adoption. However, the submitted manuscript provides no verifiable evidence. There are no reproducible code artifacts, no parameter-free derivations, no machine-checked proofs, and no numerical results that could be independently assessed. The only readable portion is an abstract with strong but unsupported claims. As such, the potential significance is high, but the current manuscript cannot support it.","major_comments":[{"comment":"The central claims—'superior performance', 'comparable to high-quality research acquisitions', and 'reducing measurement errors to lower sample size requirements'—are asserted without a single numerical result, confidence interval, effect size, or p-value. The full text of the PDF is unreadable (mojibake), so no methods, cohort sizes, acquisition parameters, or outcome metrics are available for verification. This is load-bearing because the paper's contribution is an empirical validation claim.","section":"Abstract / Full text"},{"comment":"Even granting the abstract's intent, I cannot determine whether evaluation was done on held-out subjects or held-out disorder groups, whether the evaluation metric is identical to the training loss, or whether the research-quality FODs used as training targets are independent of the reference used for evaluation. If the same subjects/acquisitions appear in both training and testing, or if the 'research-quality' benchmark is the same regression target the model was trained on, the comparability claim becomes circular. No legible text rules this out.","section":"Methods / Evaluation (unreadable)"}],"minor_comments":[{"comment":"The PDF character encoding is corrupted; the full text is unreadable. The authors must resubmit a properly rendered PDF with embedded fonts and a readable character map.","section":"PDF formatting"},{"comment":"Even in the abstract, key performance metrics (e.g., mean/standard deviation, subject counts, comparison baselines) should be reported numerically to support the stated claims.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The submitted PDF is corrupted to the point that I cannot review any part beyond the abstract. This appears to be a submission/pipeline issue rather than a scientific flaw, but as it stands the manuscript is not reviewable. I recommend the editor return the manuscript to the authors with a request for a readable full text containing complete methods, dataset-split details, and numerical results. If the full text cannot be produced, the manuscript should be rejected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing in this paper is the scope of validation: FastFOD-Net across healthy controls plus six neurological disorders, with a 60x speedup over its predecessor. If that is real, it matters. Most prior deep learning enhancement work stays on healthy subjects, and the authors are right that clinical adoption depends on showing performance where pathology alters tissue. The framing of the clinical problem is honest and the motivation is sound.\n\nBut the abstract alone cannot support what it claims. There are no numbers, no sample sizes, no error bars, no statistical comparisons. We are asked to trust that the method has “superior performance” and produces FODs “comparable to research acquisitions” without a single quantitative result. The stress-test concern about circularity is not a nitpick: if the evaluation metric is voxel-wise similarity to the same research-quality FODs used as training targets, and if the patient cohorts overlap with the training data, then “comparable” is close to guaranteed. The abstract says nothing about held-out subjects, acquisition sites, or whether the six disorder groups were seen during training. The domain-shift worry is also legitimate—edema, atrophy, lesions are not healthy white matter, and a model trained mostly on controls can degrade silently.\n\nPractically, the provided PDF full text is corrupted to the point of being unreadable, so I cannot audit the methods or results. That is the biggest obstacle. I would not cite this work yet, and I would not bring it to reading group until the actual numbers are available. But the potential significance is real, and the study design—if it truly validates across six disorders—is the kind of evidence the field needs. So I would send it to peer review rather than desk reject, with strong instructions to the authors: report the evaluation protocol explicitly (train/test split, subject-level separation, metric independence from training loss), give per-disorder results, and make the full text legible. If the posted PDF is the version the authors intend others to read, they need to fix that first.","headline":"Clinical breadth is the real contribution here, but the abstract gives no numbers to back the claims—worth refereeing if the full text is actually readable.","tokens_in":4073,"tokens_out":1745,"would_cite":false,"duration_ms":25094,"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":"FastFOD-Net brings research-grade fiber orientation analysis to clinical diffusion MRI, validated across healthy controls and six neurological disorders.","keywords":["diffusion MRI","fiber orientation distribution","deep learning","FOD enhancement","clinical translation","tractography","connectome","neurological disorders"],"falsifier":"Acquire both clinical and research-quality diffusion MRI in the same subjects, including disorders beyond the six studied, and compare FastFOD-Net-enhanced FODs from the clinical scan against FODs estimated from the research scan voxel by voxel. If the enhanced FODs systematically diverge—for instance, missing known tracts on tractography or disagreeing with expert manual white-matter parcellation—the claim of comparability to research acquisitions would be falsified.","tokens_in":3312,"feed_emoji":"🧠","tokens_out":4338,"duration_ms":42865,"temperature":0.7,"pith_summary":"This paper claims that a deep learning network can close the quality gap between routine clinical diffusion MRI and high-end research scans, specifically for mapping the brain's white-matter fiber architecture. It presents FastFOD-Net, an accelerated end-to-end framework that enhances fiber orientation distributions (FODs) estimated from single-shell, low-angular-resolution clinical acquisitions so they match FODs from research-quality data. The framework is validated across healthy controls and six neurological disorders, and it trains and runs 60 times faster than its predecessor. This matters because FODs are the foundation for tractography and connectome analysis, tools that could support clinical neuroscience if they worked on widely available clinical scans. If the claims hold, hospitals could perform research-grade diffusion MRI analysis without research-grade scanning hardware.","feed_headline":"60x faster network makes clinical diffusion MRI research-grade","feed_subtitle":"Validated across healthy brains and six disorders, FastFOD-Net brings tractography and connectome analysis to routine scans.","key_machinery":"The fiber orientation distribution (FOD) is the central object: it models the complex arrangement of white-matter fibers at each brain voxel and serves as the basis for tractography and connectome analysis. FastFOD-Net is the accelerated end-to-end deep learning network that takes low-quality clinical diffusion MRI signals and regresses high-quality FODs, with a 60x speedup in training and inference relative to its predecessor. The network is the mechanism that transforms routine clinical acquisitions into research-grade FODs, and the clinical evaluation across healthy controls and six disorders is the evidence that this transformation works in practice.","core_discovery":"The central claim is that FastFOD-Net—an optimized deep learning framework for enhancing fiber orientation distributions—produces FODs from clinical diffusion MRI acquisitions that are comparable to those estimated from high-quality research acquisitions, and does so across healthy controls and six neurological disorders. The authors support this with what they describe as the most comprehensive clinical evaluation of deep learning FOD enhancement to date, reporting a 60x improvement in training and inference efficiency over the predecessor network. The contribution is to demonstrate that a clinically practical deep learning pipeline can support FOD-based tractography and connectome analysis","pith_inferences":["If the same enhancement approach transfers to other diffusion-derived quantities, such as tensor metrics or fixel-based measures, it might unify clinical diffusion MRI analysis around a single trained model.","The comparability claim is only as strong as the training distribution; independent validation on unseen scanner vendors and patient cohorts would be the natural next test.","A direct head-to-head comparison of FastFOD-Net-enhanced FODs against multi-shell research acquisitions on the same subjects would quantify the residual gap and reveal which fiber populations are most affected.","The 60x speedup suggests the method could be embedded in real-time acquisition pipelines, potentially allowing radiologists to see enhanced FODs while the patient is still in the scanner."],"forward_implications":["Clinical diffusion MRI datasets that were previously too low-quality for tractography could now be analyzed for white-matter fiber architecture and connectomics.","Deep-learning FOD enhancement may reduce measurement error, lowering the sample sizes needed to detect group differences in clinical neuroscience studies.","FastFOD-Net's 60x efficiency gain makes training and inference practical in hospital workflows, potentially enabling near-real-time tractography in a clinical setting.","The validation across six neurological disorders supports generalization beyond healthy controls, building a case for clinical adoption of deep-learning diffusion MRI enhancement."],"supporting_citations":[],"fun_headline_variants":["FastFOD-Net: 60x faster clinical diffusion MRI, research-grade","Deep learning makes routine MRI match research scans","Faster FODs: clinical diffusion MRI now research-grade","60x speedup brings deep FOD enhancement to clinics","Tractography on routine scans: FastFOD-Net validated"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The enhanced FODs are assumed to faithfully represent true white-matter fiber architecture, implying the research-quality FODs used as training targets are an unbiased gold standard and the clinical populations studied are representative of all patients the network might encounter.","fun_headline_variants_meta":{"raw":{"variants":["FastFOD-Net: 60x faster clinical diffusion MRI, research-grade","Deep learning makes routine MRI match research scans","Faster FODs: clinical diffusion MRI now research-grade","60x speedup brings deep FOD enhancement to clinics","Tractography on routine scans: FastFOD-Net validated"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000561,"raw_usage":{"total_tokens":2517,"prompt_tokens":776,"completion_tokens":1741,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":1655}},"tokens_in":520,"tokens_out":1741,"duration_ms":13505,"temperature":1.0,"reasoning_tokens":1655,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:39:28.539212+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Acquire both clinical and research-quality diffusion MRI in the same subjects, including disorders beyond the six studied, and compare FastFOD-Net-enhanced FODs from the clinical scan against FODs estimated from the research scan voxel by voxel. If the enhanced FODs systematically diverge—for instance, missing known tracts on tractography or disagreeing with expert manual white-matter parcellation—the claim of comparability to research acquisitions would be falsified.","supporting_citations":[],"review_version":1}