{"id":"b3118d3e-c25d-4476-9781-cb868a05f5f9","arxiv_id":"2502.01661","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Using pseudo-labeled censored data, semi-supervised regression improved lung cancer overall survival time prediction by 26.5% in mean absolute error for PET radiomics features versus supervised learning in a 199-patient study.","lead":"This paper tests whether a semi-supervised learning approach, which adds machine-generated labels for censored lung cancer patients, improves survival time prediction from PET/CT images and clinical data. In their analysis of 199 patients, the method reduced mean absolute error by 26.5% for PET handcrafted radiomics features compared to supervised learning alone.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The TCIA 'external' test set is part of the 199-patient cohort used to train and pseudo-label, so the headline 26.5% SSL gain may be a leakage artifact; the claim is unverified until a clean split is run.","rationale":"I agree with the reader: the weakest assumption is the independence of the TCIA external set. This is not a matter of consensus but of internal consistency: the paper simultaneously calls TCIA external and includes it in the study cohort. Because the headline number and the 'external' c-index come from this contaminated split, they cannot support the conclusion as written. I would move the verdict from CONDITIONAL to UNVERDICTED rather than REJECT: the proposed SSL strategy is plausible, the code is public, and a clean split could settle the question; but the current manuscript does not provide enough uncontaminated evidence to be accepted even conditionally. If the clean-split reanalysis fails, REJECT would be warranted. The pseudo-labeling-before-CV issue is a second reason the current SSL comparison is uninterpretable and should be fixed in the same reanalysis.","tokens_in":13250,"tokens_out":6396,"duration_ms":64648,"concrete_test":"Exclude the 33 TCIA patients from every fitting step: PCA fitting, KNNR pseudo-label training, grid-search model selection, and SSL/SL training. Train only on the 166 BC Cancer patients, or use nested cross-validation inside BC Cancer, and evaluate once on the TCIA patients. Report SSL vs SL MAE on the TCIA uncensored cases (only 6 events, so provide bootstrap confidence intervals) and c-index on all 33 TCIA patients. In addition, move pseudo-labeling inside each CV fold so KNNR sees only training-fold labels before labeling validation-fold censored patients. If SSL still shows roughly the 26.5% MAE reduction with non-overlapping intervals on this clean split, the concern is resolved; if the gain shrinks or reverses, the reported external results were leakage artifacts.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central empirical claim—SSL with pseudo-labeled censored patients reduces MAE by 26.5% for PET_HRF (Table 2) and yields an external c-index of 0.656 (Table 4)—depends on a genuinely independent test set. The manuscript's own Methods contradict that requirement. Section 2.I states that the study cohort of 199 patients comprises BC Cancer (n=166) plus TCIA (n=33), and Section 2.II (xii) says external validation is performed on 'the external dataset, comprising 33 patients from TCIA.' Thus the 33 TCIA patients are part of the same 199-patient cohort used for PCA, KNNR pseudo-labeling, regressor training, and hyperparameter selection. Any 'external' result is a within-cohort subsample, not an independent test. A second, independent leakage path is in Section 2.II (x): KNNR pseudo-labels the 100 censored patients before five-fold cross-validation, so the pseudo-label generator can be fit on labels of samples that later appear in validation folds. This makes the SSL-vs-SL MAE differences and paired t-test p-values uninterpretable and could easily manufacture the reported 26.5% gain. If either leakage path is active, the central claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a censor-aware semi-supervised learning (SSL) framework for overall survival time prediction in lung cancer, combining clinical features with handcrafted and deep radiomic features from PET and CT images. Using 199 patients, it compares seven regression algorithms under supervised learning (SL) and SSL, where SSL pseudo-labels the 100 censored patients, and also evaluates three hazard-ratio survival analysis methods. The headline claims are that SSL reduces the MAE for PET handcrafted features by 26.5% (from 2.11 to 1.55 years) and that a CT handcrafted-feature survival model achieves an external c-index of 0.656. The central claim is that censor-aware SSL improves survival time prediction over SL.","tokens_in":13518,"tokens_out":4917,"duration_ms":45491,"significance":"If the central result held, censor-aware SSL would offer a practical way to exploit censored patients in outcome prediction, with clinical relevance in lung cancer prognosis. The manuscript has clear strengths: the code is publicly shared, the radiomics pipeline via ViSERA is reproducible, and the authors report negative results for some CT handcrafted-feature configurations, which is informative. However, the current validation contains two leakage paths that could manufacture the reported gains, and the 'external' test set is shown by the manuscript's own Methods to be part of the training cohort. The significance of the 26.5% improvement and the external c-index is therefore not established as presented.","major_comments":[{"comment":"The stated 'external validation' is not external. Section 2.I defines the 199-patient cohort as BC Cancer (n=166) plus TCIA (n=33), and Section 2.II(xii) validates on 'the external dataset, comprising 33 patients from TCIA.' Because those 33 TCIA patients are also part of the cohort used for PCA, pseudo-label generation, and regressor training, the external c-index of 0.656±0.02 in Table 4 and the bootstrap estimates are within-cohort results. The authors must either obtain a genuinely independent validation cohort or remove all 'external' claims from the manuscript.","section":"§2.I, §2.II(xii), Table 4"},{"comment":"Pseudo-labels are generated before five-fold cross-validation. If the KNNR pseudo-labeler is fit on the full labeled set and the pseudo-labeled censored samples are included in every training fold, then labeled samples in the validation fold contribute to the pseudo-labels used for training, and the validation MAE reflects information leakage. This makes the SSL-versus-SL comparisons and paired t-tests reported in Tables 2 and 3 uninterpretable. The pseudo-labeler must be refit inside each training fold, or a proper nested CV scheme must be used, before the 26.5% gain can be assessed.","section":"§2.II(x), Tables 2 and 3"},{"comment":"The abstract states that the SL baseline for PET_HRF is PCA+KNNR with an MAE of 2.11 years, but Table 2 and the main text identify the SL model as PCA+DTR (abbreviated DTC in the narrative). This is an internal inconsistency in the headline comparison, and the authors need to specify which regressor actually produced the 2.11 baseline.","section":"Abstract, §3.1, Table 2"},{"comment":"The text describes p=0.053 as 'slightly significant,' but 0.053 exceeds the 0.05 threshold and is not statistically significant. More generally, the study selects the best among seven regressors, 13 feature sets, and two learning strategies without any multiple-comparison control, so isolated p<0.05 values in Tables 2 and 3 are likely inflated by selection. The authors should report corrected p-values or clearly frame the results as exploratory.","section":"Table 3, §3.2"}],"minor_comments":[{"comment":"The first row of Table 3 (MLPR, CF) duplicates the second row of Table 2 (MLPR, CF); consider removing the duplication or referencing the earlier table.","section":"Table 3"},{"comment":"The text says the SL strategy used '104 labeled cases,' while Section 2.I reports 99 patients with OS events; this numeric inconsistency should be resolved.","section":"§2.II"},{"comment":"References [67] and [68] are identical (both are the Kirienko et al. paper); one reference should be replaced with the intended distinct source.","section":"References"},{"comment":"The regressor is abbreviated DTR in Table 2 but DTC in the Section 3.1 narrative; a single abbreviation for decision tree regression should be used throughout.","section":"Table 2, §3.1"}],"recommendation":"major_revision","confidential_remarks":"The two leakage paths identified in the major comments are load-bearing: the TCIA 'external' set is part of the training cohort, and pseudo-labels are generated before cross-validation. Both can in principle be fixed by re-analysis with a clean split and nested CV, but the authors must also provide a rigorous answer on whether any truly independent data are available. The manuscript also contains several small self-citations and conference abstracts, which is not disqualifying but warrants a check for novelty overlap with prior work by the same group."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: the headline result—26.5% MAE improvement from pseudo-labeling censored patients—is not actually demonstrated by this paper's experiments. The 'external' test set of 33 TCIA patients is a subset of the 199-patient cohort described in the Methods. Calling it external is misleading, and any bootstrap results on that subset are within-cohort, not independent. That alone would make me skeptical of the reported gains.\n\nTo give credit where it's due: the paper lays out a clear pipeline, compares seven regressors and three survival models, and shares code on GitHub. The idea of treating censored patients as unlabeled examples is plausible and has prior support (Chai et al. 2017; Hermoza et al. 2022; the authors' own [22]). The manuscript also flags its small sample size and the risk of noisy pseudo-labels, which shows some honesty.\n\nThe bigger problem is a second leakage path. KNNR generates pseudo-labels for all censored patients before five-fold cross-validation. If the KNNR is fit on the full labeled set, then validation-fold patients contribute to pseudo-label generation, making the SSL-vs-SL comparison circular. The paper does not describe a nested or train-only pseudo-labeling scheme. There's also an internal inconsistency: the abstract says SL's 2.11 came from PCA+KNNR, but Table 2 shows DTR for that row. Not fatal by itself, but it undermines confidence in how carefully the results were assembled.\n\nThese issues are fixable. A clean train/test split and pseudo-labeling nested inside CV folds would make the comparison meaningful. As written, though, the central claim is unverified.\n\nWho is this for? People working on survival prediction with small censored datasets may find the setup useful, and it is a good teaching example of external-validation pitfalls. I would send it back for major revision, not desk-reject it, because the code is available and the authors can rerun the analysis properly. But I would not trust the 26.5% number until that happens.","headline":"The pseudo-labeling idea is plausible, but the evaluation leaks: the TCIA 'external' set is inside the training cohort and pseudo-labels are generated before CV, so the 26.5% SSL gain is unverified.","tokens_in":14095,"tokens_out":3194,"would_cite":false,"duration_ms":30338,"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":"Censor-aware semi-supervised learning, which pseudo-labels censored lung cancer patients, improves overall survival time prediction over supervised learning, cutting mean absolute error by 26.5% on PET radiomics.","keywords":["lung cancer","overall survival","semi-supervised learning","pseudo-labeling","radiomics","PET imaging","CT imaging","machine learning"],"falsifier":"Recompute the external-test metrics after verifying whether any of the 33 public-archive patients contributed to pseudo-label generation or feature scaling; if overlap exists, remove them and recompute the 26.5% mean-absolute-error gain and the external concordance index. If the gain disappears, the central claim is not supported.","tokens_in":13056,"feed_emoji":"🫁","tokens_out":5367,"duration_ms":46385,"temperature":0.7,"pith_summary":"This paper argues that patients who do not die during follow-up—censored observations—should not be thrown away or unevenly weighted when training survival-time predictors. Instead, the authors propose to pseudo-label these censored patients using a model trained on patients with known survival times, then retrain on the combined set. Applied to 199 lung cancer patients with clinical, PET, and CT features, this censor-aware semi-supervised strategy reduces the mean absolute error of overall survival time prediction by 26.5% for handcrafted PET radiomics features, from 2.11 to 1.55 years (p<0.05). If valid, the approach offers a practical way to use routinely collected but incomplete follow-up data to build better prognostic models.","feed_headline":"Pseudo-labeling censored patients cuts survival-prediction error 26.5%","feed_subtitle":"Including censored patients as pseudo-labeled data cuts mean absolute error from 2.11 to 1.55 years in PET-based models.","key_machinery":"Pseudo-labeling: a K-nearest-neighbor regressor trained on uncensored cases assigns an approximate survival time to each censored patient (someone without an observed death), and the resulting pseudo-labeled censored patients are added to the training set for the seven regressors compared in the study. Principal component analysis reduces each feature set to 10 components before training. The survival-analysis arm uses three hazard-ratio survival algorithms, with risk groups split at the mean predicted survival time and evaluated by concordance index and Kaplan-Meier log-rank tests.","core_discovery":"The central discovery claimed is that censor-aware pseudo-labeling meaningfully improves survival-time regression, and the largest gain appears with handcrafted radiomics features from PET scans. The best supervised baseline (PCA plus decision-tree regression) gives a mean absolute error of 2.11 years; replacing the supervised training with pseudo-labeling of the censored patients lowers the mean absolute error to 1.55 years. The paper also reports that the semi-supervised approach beats supervised learning for clinical features alone, with a 19.3% gain, and for combined PET deep-radiomics features, with a 17.9% gain. For CT-based features the gains are smaller or negative; only the combination of deep and handcrafted CT radiomics shows a statistically significant improvement. In the survival-analysis component, the best model, a component-wise gradient-boosted survival model on CT handcrafted features, reaches an external concordance index of 0.65 with a significant log-rank split.","pith_inferences":["The headline 26.5% gain may be optimistic because the claimed external test cohort is also listed as part of the 199-patient study cohort, so genuine independence of that test is questionable.","The benefit of pseudo-labeling likely depends on the censoring pattern; with roughly half the patients censored and a short follow-up window, the gains may not transfer to datasets with longer follow-up or informative censoring.","Because the quality of pseudo-labels is set by a K-nearest-neighbor baseline, the reported results are a lower bound for what a stronger pseudo-labeler might achieve, but the negative CT results show the approach is not universally beneficial."],"forward_implications":["Censored patients, normally excluded from regression training, can instead contribute pseudo-labeled data and improve survival-time prediction when label data are scarce.","Handcrafted PET radiomics benefit most from the semi-supervised treatment, with the reported 26.5% reduction in mean absolute error.","Combining deep and handcrafted CT radiomics yields a smaller but statistically significant improvement, suggesting modality and feature-type dependence.","The survival-analysis results support using CT handcrafted features to separate high- and low-risk lung cancer patients with an external concordance index of 0.65."],"supporting_citations":[{"why":"supplies the 33-patient public-archive cohort used for external testing.","marker":"[40]"},{"why":"prior demonstration of semi-supervised pseudo-labeling on lung cancer PET/CT datasets, which this study extends.","marker":"[22]"},{"why":"introduced censor-aware semi-supervised learning for survival time from medical images, the direct methodological predecessor.","marker":"[59]"},{"why":"combined Cox and accelerated failure time models with semi-supervised learning for cancer survival, giving background for SSL in survival analysis.","marker":"[58]"},{"why":"provides the ViSERA software used for handcrafted and deep radiomic feature extraction.","marker":"[32]"},{"why":"censor-aware deep ordinal regression for survival prediction, supporting the general censor-aware approach.","marker":"[17]"}],"fun_headline_variants":["Pseudo-labeling censored patients cuts survival error 26.5%","Censor-aware pseudo-labeling reduces survival MAE by 26.5%","Semi-supervised pseudo-labels on censored data cut error 26.5%","Pseudo-labels on censored PET data cut survival error 26.5%","Censor-aware SSL on PET cuts survival error 26.5%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed external validation assumes the 33 patients drawn from the public imaging archive are truly separate from model training, yet the Methods section lists those same 33 patients as part of the 199-patient study cohort used for feature extraction and model development.","fun_headline_variants_meta":{"raw":{"variants":["Pseudo-labeling censored patients cuts survival error 26.5%","Censor-aware pseudo-labeling reduces survival MAE by 26.5%","Semi-supervised pseudo-labels on censored data cut error 26.5%","Pseudo-labels on censored PET data cut survival error 26.5%","Censor-aware SSL on PET cuts survival error 26.5%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001422,"raw_usage":{"total_tokens":5829,"prompt_tokens":1124,"completion_tokens":4705,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":740,"completion_tokens_details":{"reasoning_tokens":4599}},"tokens_in":740,"tokens_out":4705,"duration_ms":29580,"temperature":1.0,"reasoning_tokens":4599,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T19:50:38.524156+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the external-test metrics after verifying whether any of the 33 public-archive patients contributed to pseudo-label generation or feature scaling; if overlap exists, remove them and recompute the 26.5% mean-absolute-error gain and the external concordance index. If the gain disappears, the central claim is not supported.","supporting_citations":[{"cited_title":"A radiogenomic dataset of non -small cell lung cancer,","cited_arxiv_id":null,"evidence_quote":"supplies the 33-patient public-archive cohort used for external testing."},{"cited_title":"Enhanced Lung Cancer Survival Prediction Using Semi -Supervised Pseudo-Labeling and Learning from Diverse PET/CT Datasets,","cited_arxiv_id":null,"evidence_quote":"prior demonstration of semi-supervised pseudo-labeling on lung cancer PET/CT datasets, which this study extends."},{"cited_title":"Censor -aware Semi-supervised Learning for Survival Time Prediction from Medical Images,","cited_arxiv_id":null,"evidence_quote":"introduced censor-aware semi-supervised learning for survival time from medical images, the direct methodological predecessor."},{"cited_title":"A new semi-supervised learning model combined with Cox and SP-AFT models in cancer survival analysis,","cited_arxiv_id":null,"evidence_quote":"combined Cox and accelerated failure time models with semi-supervised learning for cancer survival, giving background for SSL in survival analysis."},{"cited_title":"ViSERA: Visualized & Standardized Environment for Radiomics Analysis - A Shareable, Executable, and Reproducible Workflow Generator,","cited_arxiv_id":null,"evidence_quote":"provides the ViSERA software used for handcrafted and deep radiomic feature extraction."},{"cited_title":"Censor-Aware Semi-supervised Learning for Survival Time Prediction from Medical Images,","cited_arxiv_id":null,"evidence_quote":"censor-aware deep ordinal regression for survival prediction, supporting the general censor-aware approach."}],"review_version":1}