{"id":"5d5d0ab2-ccef-40be-8b9f-d092fe851e82","arxiv_id":"2509.07042","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A dual-branch U-Mamba MRI model predicts gestational age at birth and preterm birth with ~3 weeks MAE, comparable to manual cervical length regression.","lead":"A deep learning model that reads placental and whole-uterus MRI predicts gestational age at birth with about three weeks of error, matching the accuracy of manual cervical length measurements. This suggests automated MRI could flag preterm risk when cervical length is unavailable or hard to measure.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No control for gestational age at scan: PUUMA's predictive signal may be driven by scan age rather than T2* placental/uterus features.","rationale":"The reader's weakest_assumption correctly identified both the small number of preterm test cases and the GA-at-scan confound. I focus on the latter because it is more directly load-bearing: the architecture explicitly uses GA at scan as an input, and the paper never reports a comparison against scan-age-only prediction. Such a comparison is essential for any claim that MRI-derived T2* features are informative. The small-sample issue affects the precision of the sensitivity estimate but does not by itself invalidate the architectural contribution; the scan-age confound, if real, would invalidate the core attribution to functional imaging. Since this is an addressable methodological gap rather than an internal contradiction, the existing CONDITIONAL verdict remains appropriate: the paper should be accepted only with an added scan-age baseline or ablation demonstrating added value, and ideally with confidence intervals for sensitivity. My agreement is 'partial' because the reader's primary emphasis was on sample size, whereas I see the scan-age control as the more fundamental threat. A concrete test can settle this quickly and without new data collection.","tokens_in":8039,"tokens_out":3439,"duration_ms":43309,"concrete_test":"Evaluate a baseline predictor that outputs GA at birth = GA at scan (or a linear regression trained on scan age only) on the same 26-subject test set, and separately retrain PUUMA with the GA-at-scan concatenation removed (or with scan GA permuted across subjects). If the scan-age baseline matches or exceeds PUUMA's MAE/sensitivity, or if PUUMA without scan GA degrades to near-chance, the central claim that T2* imaging adds predictive value beyond known gestational age is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that T2* whole-uterus and placental imaging predicts GA at birth. But PUUMA's final prediction directly concatenates gestational age at scan as an input (Section 2.3, Figure 3), and no baseline using only GA at scan is reported. GA at scan is strongly correlated with GA at birth, especially in a cohort where scans are acquired before 37 weeks and many preterm births occur soon after the scan. A trivial model that outputs scan GA could plausibly achieve an MAE of 2–4 weeks and, for preterm classification, would label any scan before 37 weeks as preterm, potentially reproducing or exceeding the reported sensitivity of 0.67. Without an ablation that removes the GA-at-scan input or a scan-age-only baseline, the observed performance cannot be attributed to T2* tissue properties. The small test set (6 preterm cases) is a valid concern, but the scan-age confound is more load-bearing because it undermines the mechanistic claim that functional placental imaging contributes predictive information.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces PUUMA, a dual-branch U-Mamba-based architecture that predicts gestational age (GA) at birth and classifies preterm versus term birth from T2* whole-uterus fetal MRI and high-resolution placental patches. The model is trained on 295 pregnancies (after exclusions) and benchmarked against whole-uterus U-Mamba, a U-Net baseline, and linear regression on manual cervical length measurements. On a 26-subject test set containing 6 preterm births, PUUMA and the cervical-length regression achieve mean absolute errors of about 3 weeks and a preterm-detection sensitivity of 0.67, while whole-uterus U-Mamba alone reaches sensitivity 0.33. The authors frame the results as a proof of concept for automated functional MRI prediction of GA at birth and preterm risk, while acknowledging the small test set and class imbalance.","tokens_in":8291,"tokens_out":4015,"duration_ms":44369,"significance":"If the predictive contribution of placental and whole-uterus T2* signal were established, the work would be a useful step toward non-invasive MRI-based preterm-risk assessment, and the architectural design (global whole-uterus branch plus local placental patch branch) is reasonable and clearly described. The authors are also transparent about the imbalanced cohort and the small test set, and they benchmark against a clinically motivated cervical-length baseline. The main barrier to accepting the central claim is the unaddressed GA-at-scan confound: because GA at scan is provided as a direct model input and no scan-age-only baseline or ablation is reported, the observed performance cannot yet be attributed to T2* tissue properties. The manuscript would be strengthened by targeted experiments that isolate the imaging contribution.","major_comments":[{"comment":"The final prediction layer concatenates gestational age at scan with the two branch outputs. No baseline using only GA at scan, and no ablation that removes the GA-at-scan input, is reported. In this cohort, scans are acquired before 37 weeks (Section 2.2) and GA at scan is necessarily strongly correlated with GA at birth, especially for preterm births that occur soon after the scan. A trivial model that outputs or thresholds scan GA could plausibly produce MAE in the 2–4 week range and, for preterm classification, could label most pre-37-week scans as preterm. Therefore the reported PUUMA performance (Table 2) does not currently support the mechanistic claim that functional placental/uterine T2* features contribute predictive information. Please add (a) a scan-age-only baseline, (b) an ablation of PUUMA without GA-at-scan input, and (c) a report of the scan-to-birth time distribution.","section":"Section 2.3, Figure 3"},{"comment":"The central comparison rests on a test set of 26 subjects with only 6 preterm births. A sensitivity of 0.67 corresponds to 4 of 6 positives, with a wide 95% confidence interval (roughly 0.22–0.96). Differences in MAE (3.05 vs. 2.94 weeks) and sensitivity (0.67 vs. 0.33) between models are not assessed for statistical significance, and no confidence intervals or calibration curves are provided. Please report exact confusion matrices, bootstrap or exact binomial confidence intervals for sensitivity/specificity, and, where appropriate, paired tests for the primary comparisons. Without this, the claims of 'comparable' performance and 'good sensitivity' are not quantitatively supported.","section":"Section 3, Table 2"},{"comment":"Please clarify the patient-selection and baseline-fitting procedures. The sentence 'All fetal scans acquired after 37 weeks GA (term patients) were removed' is ambiguous: does it remove all term deliveries, or only scans acquired after 37 weeks while retaining term deliveries scanned earlier? This directly affects the correlation between scan GA and birth GA and the generalizability of the results. In addition, the cervical-length regression was fitted on 170 training cases; state whether the remaining 73 training cases lacked cervical-length measurements and whether the test set composition with respect to this baseline is comparable.","section":"Section 2.2, Section 2.4"}],"minor_comments":[{"comment":"Typographical errors: 'cortiscosteroid' should be 'corticosteroid' and 'pre-eclampisa' should be 'pre-eclampsia'.","section":"Section 1"},{"comment":"The patch-inclusion threshold of 33% placental tissue is a free parameter and is used both in training and at inference. Please report how sensitive the results are to this threshold and how many valid patch pairs are typically used per subject during sliding-window inference.","section":"Section 2.3"},{"comment":"The statement that PUUMA was the only model to predict both very preterm subjects as born before 34 weeks GA is based on visual inspection. Please report the point predictions for these two cases for all models.","section":"Section 3, Figure 4"},{"comment":"Please specify how the regression output is converted to the binary preterm/term classification used for accuracy, sensitivity, and specificity (e.g., predicted GA < 37 weeks). This is important for interpreting the reported metrics.","section":"Section 2.4, Table 2"}],"recommendation":"major_revision","confidential_remarks":"The GA-at-scan confound is the most serious issue. I would insist on the scan-age-only baseline and ablation before publication, as the current abstract and conclusions overstate the evidence for T2*-based prediction. The paper is otherwise within the scope of the journal and the methodology is presented clearly, but the quantitative evidence needs strengthening."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"PUUMA is a legitimate new application: a dual-branch U-Mamba that regresses gestational age at birth from whole-uterus T2* MRI, with a patch branch for placental detail. The work is also honest about its proof-of-concept status, and the comparison against manual cervical length measurements is a sensible, low-cost benchmark that many deep-learning papers skip. The authors state the class imbalance and the small preterm count plainly, and the per-case scatter in Figure 4 tells you more than the summary metrics alone.\n\nThe main soft spot is the gestational-age-at-scan confound. The model concatenates GA at scan into the final prediction head, and the paper reports no ablation without that input, nor a baseline that uses only scan age. So you cannot attribute the performance to placental T2* features. That said, the stress-test note overstates the problem: the dataset excludes scans after 37 weeks but includes term deliveries, so every scan is pre-term. A naive baseline that labels everyone preterm would get sensitivity 1 and specificity 0, which is not a meaningful competitor. A regression baseline using only scan age would likely produce a much larger MAE than 3 weeks, because term births are scanned weeks before delivery. So the omitted baseline is a real gap in attribution, but the mechanistic claim isn't obviously false. The paper still needs the ablation or baseline to be convincing.\n\nThe second soft spot is statistical: sensitivity of 0.67 is 4 of 6 preterm test cases, with no confidence intervals, and the MAE differences among models are within one standard deviation. No significance tests are attempted. That doesn't sink the paper, but it means the headline numbers shouldn't be quoted as stable. No code or data either, which limits reproduction.\n\nWho gets value: researchers working on fetal MRI, preterm prediction, or applying U-Mamba to regression. It deserves peer review because the problem is meaningful, the design is described in enough detail to replicate, and the missing analyses are addressable rather than fatal. I'd want the ablation and scan-age baseline before accepting; the small test set can be handled with honest uncertainty intervals.","headline":"New application, honest proof-of-concept, but missing scan-age-only baseline leaves attribution of predictive signal to MRI unresolved.","tokens_in":8793,"tokens_out":3201,"would_cite":false,"duration_ms":37747,"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":"A fully automated MRI pipeline predicts gestational age at birth from T2* fetal MRI with about 3 weeks mean error, matching manual cervical-length regression for preterm detection.","keywords":["preterm birth prediction","fetal MRI","T2* relaxometry","placental function","deep learning","U-Mamba","gestational age at birth","cervical length"],"falsifier":"Take a held-out cohort with at least 30 preterm births and compare PUUMA's preterm sensitivity against a control model trained only on gestational age at scan, using the same architecture and loss. If the control matches or beats PUUMA's sensitivity and mean absolute error, the T2* imaging signal is not carrying the predictive information attributed to it; alternatively, if PUUMA's sensitivity on the larger cohort falls well below 0.5 while the scan-age-only model remains near chance, the original sensitivity was not a stable estimate.","tokens_in":7968,"feed_emoji":"🤰","tokens_out":10265,"duration_ms":108076,"temperature":0.7,"pith_summary":"Preterm birth is hard to predict because its causes are many and poorly understood, and current clinical markers are operator-dependent or invasive. This paper tries to establish whether one safe, non-invasive measurement—T2* functional MRI of the placenta and uterus—can predict when a pregnancy will end. The authors build PUUMA, a dual-branch deep-learning model that reads the whole uterus at low resolution and the placenta at high resolution, and test it on 295 pregnancies with a deliberately imbalanced mix of term and preterm deliveries. On 26 held-out cases, PUUMA reached a mean absolute error of about 3 weeks in gestational age at birth and detected 67% of preterm births, the same sensitivity as a linear regression on cervical length measured manually from anatomical MRI. The paper presents this as proof of concept that automated MRI analysis can flag at-risk pregnancies and that MRI-derived cervical length is also informative, with implications for steroid timing and delivery planning.","feed_headline":"MRI model predicts preterm birth as well as cervical length","feed_subtitle":"Whole-uterus T2* images plus placental patches match manual cervical-length screening for preterm risk.","key_machinery":"The load-bearing mechanism is PUUMA's two-branch design built on U-Mamba, a U-Net whose encoder and decoder use Mamba state-space blocks to capture long-range dependencies. A global branch processes downsampled 128×128×64 whole-uterus T2* volumes and also predicts the placental segmentation mask; a local branch processes 16×16×16 high-resolution patches sampled from placental regions, requiring more than 33% placental tissue per patch. A fully connected layer concatenates both branch predictions with gestational age at scan and yields the final GA-at-birth regression and preterm classification. T2* maps come from mono-exponential fitting clipped at 300 ms, and placental masks are generated a","core_discovery":"On its own terms, the paper's central discovery is that the functional state of the placenta and uterus, as encoded by T2* MRI, carries enough information about birth timing that a fully automated model can match a clinician's cervical length measurement. PUUMA fuses a global U-Mamba branch on whole-uterus volumes with a local U-Mamba encoder on 16×16×16 placental patches and combines both with gestational age at scan. On 26 test cases it reaches 3.05-week mean absolute error, 0.65 accuracy, and 0.67 sensitivity; cervical-length regression reaches 2.94 weeks and 0.67 sensitivity. PUUMA is the only model to predict both very preterm subjects as born before 34 weeks, whereas global-only U-Mamb","pith_inferences":["A direct way to test whether T2* adds independent information is to compare PUUMA against a model that receives only gestational age at scan; if that baseline matches the reported MAE, the placental signal would not be carrying the prediction.","The paper does not fuse its two signal sources—automated T2* imaging and MRI cervical length—into a single predictor; combining them may yield higher sensitivity than either alone, and is a natural next experiment.","The same patch-plus-whole-organ design could transfer to other functional MRI contrasts, such as diffusion or T1, or to placental ultrasound, provided the local patch sampling survives resolution differences.","Because the cohort is enriched for high-risk pregnancies and excludes late scans, the reported error may be specific to this population; external cohorts with a more representative delivery mix would test the generalisability."],"forward_implications":["If the test-set performance holds, a clinician could obtain a preterm-risk estimate from the same T2* acquisition already used to assess placental function—no extra scan, no manual contouring.","Because PUUMA's sensitivity for preterm detection (0.67) is double the global-branch U-Mamba's (0.33), local placental patches appear to carry signal that whole-uterus shape and topology do not.","MRI-derived cervical length could become a predictive marker in pregnancies where transvaginal ultrasound is unavailable or not desired.","A mean error near 3 weeks is in the range where the estimate could support decisions about antenatal corticosteroid timing and delivery at a centre with appropriate neonatal care."],"supporting_citations":[{"why":"Supplies the U-Mamba backbone that PUUMA modifies from segmentation into regression and classification.","marker":"[14]"},{"why":"Automatically generates the placental segmentation masks used to sample the high-resolution local patches.","marker":"[12]"},{"why":"Introduces the Mamba selective state-space model that forms the core of both PUUMA branches.","marker":"[7]"},{"why":"Establishes the structured state-space sequence models underlying Mamba's linear-complexity long-range dependencies.","marker":"[8]"},{"why":"Bases the choice of T2* relaxometry as a functional probe of placental oxygenation and microstructure.","marker":"[11]"},{"why":"Prior deep-learning work on preterm-birth markers that frames the comparison with automated prediction.","marker":"[26]"},{"why":"Provides the clinical interpretation of cervical length's strong negative predictive value, which explains the regression baseline's term-case performance.","marker":"[23]"}],"fun_headline_variants":["MRI model ties cervical length for preterm prediction","Whole-uterus MRI matches cervical length for preterm risk","Automated MRI matches manual cervical length for preterm risk","Deep MRI predicts preterm birth on par with cervical length"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The results stand on two linked premises: that a sensitivity of 0.67 measured on only six preterm births in the test set is a dependable estimate, and that the T2* placental and uterine signal adds predictive information beyond gestational age at scan, which the model receives as a direct input.","fun_headline_variants_meta":{"raw":{"variants":["MRI model ties cervical length for preterm prediction","Whole-uterus MRI matches cervical length for preterm risk","Automated MRI matches manual cervical length for preterm risk","Deep MRI predicts preterm birth on par with cervical length"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001023,"raw_usage":{"total_tokens":4186,"prompt_tokens":812,"completion_tokens":3374,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":3312}},"tokens_in":556,"tokens_out":3374,"duration_ms":26511,"temperature":1.0,"reasoning_tokens":3312,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T23:28:20.994380+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out cohort with at least 30 preterm births and compare PUUMA's preterm sensitivity against a control model trained only on gestational age at scan, using the same architecture and loss. If the control matches or beats PUUMA's sensitivity and mean absolute error, the T2* imaging signal is not carrying the predictive information attributed to it; alternatively, if PUUMA's sensitivity on the larger cohort falls well below 0.5 while the scan-age-only model remains near chance, the original sensitivity was not a stable estimate.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior deep-learning work on preterm-birth markers that frames the comparison with automated prediction."},{"cited_title":"https://doi.org/10.1016/j.siny.2018.09.006","cited_arxiv_id":null,"evidence_quote":"Provides the clinical interpretation of cervical length's strong negative predictive value, which explains the regression baseline's term-case performance."}],"review_version":1}