{"id":"341ea3b3-34e7-48e5-be39-341bfe543be4","arxiv_id":"2501.16400","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A multimodal spatiotemporal network that fuses two timepoint CT scans and clinical data reaches 0.897 accuracy and 0.939 AUC for lung nodule malignancy on a new NLST-based dataset.","lead":"This paper presents CSF-Net, a deep-learning model that combines two CT scans of a lung nodule taken at different times with patient clinical data to predict whether the nodule will become cancerous. Its authors report higher accuracy on a newly built set of 443 cases from the NLST screening study than six existing methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The clinical variables fed to the text encoder are never enumerated; if Section 3.1's 'screening results as defined by the NLST' is one of them, the Table 2 gains could reflect a strong proxy for the malignancy label rather than cross-modal fusion.","rationale":"The reader's flagged weakest assumption is exactly the most load-bearing concern in this paper. The paper's central empirical claim is that cross-modal spatiotemporal fusion with follow-up CT and clinical data outperforms existing methods. That claim depends on the clinical features being legitimate, non-leaking inputs. Section 3.1 lists screening results as part of the clinical information, but never enumerates the features. In the NLST, the screening result is a direct radiological assessment of the same CT scan, strongly predictive of eventual cancer diagnosis. If this field is included in the text encoder, the cross-modal attention module can read a near-label proxy, inflating all metrics. The ablation in Table 2 is supposed to show the contribution of clinical data, but it only compares with and without all clinical features; it does not isolate the screening-result field. Since none of the six comparison methods uses this field, the comparison in Table 1 is also potentially unfair. This concern is concrete, testable, and directly threatens the central claim. It does not prove the result is wrong, because the feature list and code could reveal that screening results are excluded; hence REJECT is too strong. The appropriate response is a conditional acceptance requiring the feature list to be disclosed and the screening-result feature tested. The reader already chose CONDITIONAL, so no verdict adjustment is needed. In addition to this main concern, the paper lacks error bars and does not describe the train/validation/test split, but those are secondary to the feature-leakage issue because even a well-split experiment would be uninterpretable if a screening-result feature is present.","tokens_in":6946,"tokens_out":6126,"duration_ms":60552,"concrete_test":"Inspect the released code's NLST-cmst data loader to list the exact clinical columns passed to the text encoder, and check whether any field encodes the NLST screening result (e.g., positive or suspicious for cancer). Then retrain CSF-Net on the same split with that field removed, keeping age, sex, smoking status, and all imaging inputs unchanged, and compare Table 1 and Table 2 metrics. If AUC and accuracy drop materially, or the clinical-data ablation gain largely disappears, the central claim is not supported; if performance is unchanged, the concern is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that CSF-Net 'surpasses existing methods' rests on the experiments in Tables 1-2. Section 3.1 says the clinical information includes 'age, gender, smoking status, and screening results as defined by the NLST,' but the exact fields passed to the text encoder are never listed. In NLST, the screening result is a radiologist's categorization of the CT screen (e.g., negative vs. positive or suspicious), generated from the same images that supply the nodule ROI. It is strongly associated with eventual cancer diagnosis, so including it as a clinical feature can let the cross-modal attention module attend to a near-label signal. The ablation in Table 2 attributes large gains to clinical data, yet no version removes only the screening-result field. Because none of the six baselines uses this field, the comparison is also potentially unfair. If screening result is in the input, the headline superiority and the clinical-data contribution are not evidence for the proposed spatiotemporal fusion; they may just show that the label proxy is informative. This is a load-bearing, addressable concern rather than a proven error.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CSF-Net, a deep learning model for predicting pulmonary nodule malignancy from two timepoint CT scans and clinical information. The architecture consists of a 3D ResNet with CBAM for spatial feature extraction, a Temporal Residual Fusion module that combines features from the two timepoints (Eqs. 1-2), and a Cross-Modal Attention Fusion module (Eqs. 3-4) that integrates image features with clinical text features. The authors construct a new dataset, NLST-cmst, of 443 NLST subjects with physician-annotated nodule ROIs and pathologically confirmed malignancy labels. On this dataset, CSF-Net reports accuracy 0.8974, precision 0.8235, F1 0.8750, AUC 0.9389, and recall 0.9333, outperforming six comparison methods across most metrics. The code is publicly available.","tokens_in":7144,"tokens_out":4082,"duration_ms":34465,"significance":"The potential contribution is twofold: a publicly documented dataset for longitudinal nodule malignancy prediction, and a modular architecture that explicitly models spatiotemporal and cross-modal fusion. If the reported gains are robust, the paper would support the clinical practice of combining follow-up imaging with structured clinical data. The strengths include public code, an external pathological gold standard for labels, and an ablation study that isolates several architectural choices. The main weakness is empirical: no confidence intervals or significance tests are reported, and the clinical feature set is not fully specified. In particular, the possible inclusion of NLST screening results - a variable derived from the same CT images and strongly associated with the outcome - creates a risk of label leakage that could invalidate the headline comparison. This concern is addressable and should be resolved before publication.","major_comments":[{"comment":"The clinical information fed to the text encoder is never enumerated. Section 3.1 lists 'age, gender, smoking status, and screening results as defined by the NLST' as available clinical information, but the manuscript does not state which of these fields are actually used as input features. The NLST screening result is a radiologist-generated categorization of the CT screen and is strongly predictive of an eventual cancer diagnosis; if this field is among the text-encoder inputs, the cross-modal attention module can effectively read the label from the input. Because none of the six baselines uses this field, the performance advantage in Table 1 would not demonstrate the value of spatiotemporal or cross-modal fusion. Please enumerate the exact clinical variables used and provide an ablation that removes the screening-result field, or explicitly confirm that this field was never included.","section":"Section 3.1"},{"comment":"The empirical comparison lacks confidence intervals, significance tests, and a description of the data split. The abstract claims 'significant performance improvements,' but no statistical measure supports this wording. The manuscript should state how the 443 subjects were split into train/validation/test sets, for example the split ratio, whether the split was at the patient level, and the random seed, and it should report bootstrap confidence intervals or a paired permutation test over the test set. It should also state explicitly whether all baselines were evaluated on exactly the same partitions.","section":"Section 3.3, Table 1"},{"comment":"Several comparison methods, including SCANs, NAS-Lung, T-LSTM, RadFusion, DeepCAD, and MFCN, were reimplemented by the authors because public code was not available. The reported numbers may not reflect the official implementations, and the sentence 'The results presented are the actual results from our experiments' does not address how hyperparameters were tuned or whether the reimplementations were validated against known results. Please provide the exact training settings for each baseline and, where possible, report results from official code or from prior publications on the same data.","section":"Section 3.3, comparative experiments"},{"comment":"The ablation study does not isolate the contribution of clinical data from the contribution of the screening-result field. The row 'without CMAF+clinical' removes both the cross-modal attention module and all clinical data together, so any drop in performance cannot be attributed separately to the fusion mechanism or to the clinical variables. Given the concerns in the first major comment, an ablation that removes only the screening-result feature while keeping the remaining clinical variables is essential to support the claim that clinical information improves prediction.","section":"Section 3.4, Table 2"}],"minor_comments":[{"comment":"The PDF header line breaks the word as 'Spa tiotemporal'; it should read 'Spatiotemporal'.","section":"Title/header"},{"comment":"The Adam optimizer is reported with beta1=0.5, which is unusual (the standard value is 0.9); please clarify the choice or correct the value, and specify what is updated every 20 epochs.","section":"Section 3.2"},{"comment":"The method name is inconsistent: it appears as both 'MCFN [10]' in the introduction and 'MFCN [10]' in the comparative experiments and Table 1; please make it consistent.","section":"Section 3.3"},{"comment":"The row for CSF-Net contains formatting artifacts in the delimiters; the numbers should be displayed cleanly so that the improvements over the best baseline are immediately readable.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The label-leakage concern via the NLST screening result is the key risk. If the authors confirm that the screening-result field was never used as an input, or rerun the experiments without it, the paper may be acceptable. I would also urge the editor to require a full description of the data split and statistical significance testing before the paper is considered further. The code release and dataset creation are valuable, but the current manuscript is not publishable as is."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's headline claim—that CSF-Net outperforms six baselines on the new NLST-cmst dataset—may be an artifact of label leakage. Section 3.1 lists \"screening results as defined by the NLST\" among the clinical variables, but the text encoder inputs are never enumerated. In NLST, the screening result is the radiologist's overall interpretation of the CT, strongly correlated with the cancer outcome. If that field is fed into the model, the cross-modal attention is reading a near-label proxy, and the Table 2 gains attributed to clinical data are not evidence for the fusion mechanism. This is the load-bearing soft spot.\n\nWhat is genuinely new is the dataset construction. The authors screened NLST, had physicians annotate nodule locations, and built a 443-subject longitudinal CT dataset with pathological gold-standard labels. That is real work and a useful resource if released. The architecture itself is incremental: 3D ResNet with CBAM, a temporal residual fusion block, and cross-modal attention borrowed from CMAFGAN. Nothing is wrong with assembling known parts, but it is not a conceptual advance.\n\nOther soft spots: no confidence intervals or significance tests; the train/validation/test split is not described; the baselines were reimplemented by the authors, and the code link does not include the baseline implementations or the dataset. The reported numbers are not independently checkable. The ablation removes whole modules but never removes just the screening-result field, which would be the obvious control. The paper is also honest about relying on pre-identified nodule locations, and the code link is a plus, but for a paper whose whole claim is empirical superiority, the missing details and potential leakage are serious.\n\nThis deserves a serious referee because the dataset construction is a genuine contribution and the leakage concern is addressable. But as written, I would not believe the superiority claim. The authors need to enumerate the clinical fields, re-run without the screening-result variable, report error bars, and release the baseline implementations.\n\nRecommendation: send to peer review with a major-revision expectation, and get a reviewer who is familiar with NLST.","headline":"The dataset is real, but the clinical feature list may include the NLST screening result, which could turn the reported gains into label leakage.","tokens_in":7737,"tokens_out":2320,"would_cite":false,"duration_ms":21996,"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":"Combining follow-up CT scans with clinical text via learned cross-modal attention predicts nodule malignancy more accurately than single-image or naive-fusion baselines, with 0.8974 accuracy and 0.9389 AUC on a curated NLST cohort.","keywords":["pulmonary nodule","malignancy prediction","multimodal fusion","spatiotemporal network","follow-up CT","cross-modal attention","NLST dataset","deep learning"],"falsifier":"Open the released code and list the clinical variables passed to the text encoder; if the NLST screening-result field (or any variable derived from the follow-up cancer diagnosis) is present, retrain CSF-Net without it and check whether the accuracy and AUC remain near 0.8974 and 0.9389.","tokens_in":6716,"feed_emoji":"🫁","tokens_out":10321,"duration_ms":81497,"temperature":0.7,"pith_summary":"Most automated malignancy predictors use a single CT scan, but radiologists follow nodules over time and weigh clinical context. This paper attempts to show that a network explicitly modeling change between two CT timepoints and fusing that with clinical text can beat both single-image models and simpler multimodal baselines. On a new curated dataset (NLST-cmst, 443 participants from the public NLST trial with physician-annotated nodule locations and pathologically confirmed malignancy), the proposed CSF-Net reports accuracy 0.8974, AUC 0.9389, recall 0.9333, F1 0.8750, and precision 0.8235, outperforming all six compared methods on every metric except a precision tie with NAS-Lung. If the result holds, it would support the clinical intuition that longitudinal imaging plus basic patient data carries more predictive signal than a single scan.","feed_headline":"Follow-up CT plus clinical text outperforms single-scan malignancy models","feed_subtitle":"The cross-modal network reports 0.897 accuracy and 0.939 AUC on a curated NLST cohort of 443 subjects.","key_machinery":"The architecture has three load-bearing components. The spatial feature extractor, built from 3D ResNet with CBAM, produces a feature map for each of the two CT timepoints ($t_0$ and $t_1$). The temporal residual fusion module concatenates these maps, upsamples and globally average-pools them, applies a 3D convolution, and then fuses the result with the raw $t_1$ features using learnable sigmoid-gated weights $\\lambda_0$ and $\\lambda_1$ ($\\mathit{STF}_{\\mathrm{feat}} = \\sigma(\\lambda_0) F_{\\mathrm{feat}} + \\sigma(\\lambda_1) t_1$). The cross-modal attention fusion module takes the spatiotemporal image features $x_i$ and clinical text features $y_i$, computes matching scores $\\beta_{j,i}$ and $\\rho_{j,i}$ via query-key products ($s_{ij} = q_1(x_i)^T k_2(y_j)$ and $t_{ij} = q_2(y_i)^T k_1(x_j)$), and uses these scores to reweight the image features before classification. Together these modules let the model attend to spatial regions that changed between scans and that align with clinical attributes.","core_discovery":"The central discovery is that a three-module architecture, consisting of a 3D ResNet+CBAM spatial extractor, a temporal residual fusion (TRF) module, and a cross-modal attention fusion (CMAF) module, yields a clear predictive gain when applied to two-timepoint CT data plus clinical text. The model's reported metrics on NLST-cmst exceed every baseline, and the ablation study shows each design choice contributes: the follow-up timepoint $t_1$ alone reaches 0.8462 accuracy versus 0.7079 for the baseline $t_0$, adding clinical features to $t_1$ raises accuracy to 0.8590, and removing either the TRF or the CMAF module drops the full model's accuracy from 0.8974 to 0.8718. The paper presents these results as evidence that spatiotemporal and cross-modal fusion together simulate the clinician's practice of reading follow-up scans in light of patient history.","pith_inferences":["The paper does not specify which clinical fields were fed to the text encoder; if the NLST screening-result variable (derived from the same follow-up cancer diagnoses that set the malignancy label) is among them, the cross-modal attention could be reading the label, and a retraining without that variable is the decisive test.","The dominance of $t_1$ over $t_0$ may reflect recency as much as temporal dynamics; testing the same architecture with more than two timepoints or on a time-to-event outcome would separate 'learning growth' from 'learning recency'.","The cross-modal attention maps could be visualized to ask whether the model grounds its predictions in clinically meaningful image-text correspondences (e.g., nodule regions matching smoking history), which the paper does not report.","The NLST-cmst cohort is drawn from a U.S. screening trial; generalizability to other screening programs, to non-screen-detected nodules, or to differing CT acquisition protocols is untested."],"forward_implications":["The reported performance implies that two-timepoint follow-up imaging plus routinely available clinical attributes is a higher-signal input for malignancy risk stratification than a single screening CT.","The ablation results imply that the follow-up timepoint $t_1$ dominates the image signal, so screening archives missing the follow-up scan would forgo most of the predictive benefit.","The model's gains over DeepCAD, the strongest multimodal temporal baseline, indicate that learned temporal residual fusion and cross-modal attention add value beyond simple multimodal temporal concatenation.","The new NLST-cmst dataset provides a publicly available, pathologically labeled two-timepoint benchmark on which future two-timepoint and multimodal nodule models can be compared."],"supporting_citations":[{"why":"Supplies the source cohort: the National Lung Screening Trial dataset from which the authors select 443 subjects and annotate nodule ROIs.","marker":"[2]"},{"why":"DeepCAD: the strongest multimodal and multi-temporal baseline, which CSF-Net is compared against and surpasses on all metrics except precision.","marker":"[13]"},{"why":"The cross-modal attention fusion (CMAF) mechanism is adapted from this method; it is the core of the cross-modal attention fusion module.","marker":"[14]"},{"why":"CBAM: the convolutional block attention module integrated into the spatial feature extractor to emphasize channel and spatial features.","marker":"[15]"},{"why":"RadFusion: a multimodal baseline combining imaging and EHR data, representing the clinical-information fusion approach the paper builds on.","marker":"[16]"},{"why":"SCANs: a longitudinal CT baseline using Siamese convolutional attention, one of the six comparison methods.","marker":"[6]"}],"fun_headline_variants":["Fusing follow-up CT and clinical text boosts nodule malignancy prediction","CSF-Net: Follow-up scans plus patient data lift malignancy AUC to 0.939","Two-point CT plus clinical info outperforms single-scan nodule models","Cross-modal fusion of CT and clinical data improves malignancy prediction","Spatiotemporal and cross-modal fusion nets 0.939 AUC on NLST cohort"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The clinical features fed to the text encoder must not include the NLST screening-result variable, which is defined by the same follow-up cancer diagnoses used to label malignancy; the paper does not specify which clinical fields are encoded.","fun_headline_variants_meta":{"raw":{"variants":["Fusing follow-up CT and clinical text boosts nodule malignancy prediction","CSF-Net: Follow-up scans plus patient data lift malignancy AUC to 0.939","Two-point CT plus clinical info outperforms single-scan nodule models","Cross-modal fusion of CT and clinical data improves malignancy prediction","Spatiotemporal and cross-modal fusion nets 0.939 AUC on NLST cohort"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000822,"raw_usage":{"total_tokens":3632,"prompt_tokens":1015,"completion_tokens":2617,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":2516}},"tokens_in":631,"tokens_out":2617,"duration_ms":16743,"temperature":1.0,"reasoning_tokens":2516,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:55:47.105198+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Open the released code and list the clinical variables passed to the text encoder; if the NLST screening-result field (or any variable derived from the follow-up cancer diagnosis) is present, retrain CSF-Net without it and check whether the accuracy and AUC remain near 0.8974 and 0.9389.","supporting_citations":[{"cited_title":"CSF-Net: Cross-Modal Spatiotemporal Fusion Network for Pulmonary Nodule Malignancy Predicting","cited_arxiv_id":"2501.16400","evidence_quote":"Supplies the source cohort: the National Lung Screening Trial dataset from which the authors select 443 subjects and annotate nodule ROIs."},{"cited_title":"The na- tional lung screening trial: overview and study design,","cited_arxiv_id":null,"evidence_quote":"DeepCAD: the strongest multimodal and multi-temporal baseline, which CSF-Net is compared against and surpasses on all metrics except precision."},{"cited_title":"Learning efficient, explainable and discrimina- tive representations for pulmonary nodules classifica- tion,","cited_arxiv_id":null,"evidence_quote":"The cross-modal attention fusion (CMAF) mechanism is adapted from this method; it is the core of the cross-modal attention fusion module."},{"cited_title":"Evaluate the malignancy of pulmonary nod- ules using the 3-d deep leaky noisy-or network,","cited_arxiv_id":null,"evidence_quote":"CBAM: the convolutional block attention module integrated into the spatial feature extractor to emphasize channel and spatial features."},{"cited_title":"Study on the prediction method of long-term benign and malignant pulmonary lesions based on lstm,","cited_arxiv_id":null,"evidence_quote":"RadFusion: a multimodal baseline combining imaging and EHR data, representing the clinical-information fusion approach the paper builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SCANs: a longitudinal CT baseline using Siamese convolutional attention, one of the six comparison methods."}],"review_version":1}