Investigation Report: Suspicious Citation Errors, Methodological Inconsistencies, and Implausibly Perfect Performance Metrics in a Radiopathomic Rectal Cancer Study (DOI: 10.1245/s10434-020-08659-4)

Academic fraud report · Geng Detector

Summary

This report flags a 2020 Annals of Surgical Oncology paper (Shao et al., DOI: 10.1245/s10434-020-08659-4) on multiparametric MRI and whole-slide image-based prediction of pathological response to neoadjuvant chemoradiotherapy in rectal cancer as highly suspicious. Five issues were verified or partially verified: (1) a citation mismatch where reference [1], cited as the LARC treatment standard, is in fact an NCCN guideline for Melanoma (version 2.2016); (2) identical citation [23] reused for two different software tools (PyRadiomics and CellProfiler), suggesting copy-paste composition; (3) the Statistical Analysis section reports R software version 3.1.0, released in 2014, an implausible version for a 2020 AI-based multicenter study; (4) implausibly perfect metrics including Sensitivity of 100.0 [100.0–100.0] in small validation cohorts (e.g., VC3, N=48) and near-perfect staircase ROC curves (AUC >0.99); (5) a workflow figure with overly uniform ROI shapes and textures suggestive of image reuse (unverified). Verdict: highly suspicious. Confidence is high on the citation anomalies; image- and statistics-based concerns are flagged but require original data for confirmation. The report does not assert definitive misconduct.

Verdict

🟠 Highly suspicious. Multiple independently verifiable textual and methodological anomalies were identified. Definitive misconduct cannot be asserted without original data, code, model training logs, and verification by the editorial office.

Key findings

Citation fraud / mismatch in INTRODUCTION (verified): Reference [1], cited as the standard-of-care basis for locally advanced rectal cancer (LARC), is "Coit DG, Thompson JA, Algazi A, et al. Melanoma, version 2.2016, NCCN clinical practice guidelines in oncology." This is an NCCN melanoma guideline, not a rectal cancer guideline.
Duplicate citation for different tools (verified): Reference [23] ("Jjm VG, Fedorov A, Parmar C, et al. Computational radiomics system…", i.e., the PyRadiomics reference) is cited for both PyRadiomics-based mp-MRI feature extraction and CellProfiler-based WSI feature extraction, indicating copy-paste authorship.
Implausible software version (verified): Statistical Analysis states "R software (version 3.1.0)." R 3.1.0 was released in April 2014, inconsistent with a 2020 multicenter AI radiopathomic study.
Implausibly perfect performance (verified as numeric anomaly): Table 2 reports Sensitivity of 100.0 [100.0–100.0] in TRG2 across multiple cohorts, including the small VC3 cohort (N=48). Figures 2c, 2f, and 2i show staircase ROC curves with AUC >0.99; Figure 4 survival curves show unusually smooth, regular separation between groups.
Figure 1 image reuse / over-rigidity (unverified): T2W and ADC tumor images display near-identical ROI shapes; WSI thumbnails exhibit highly uniform background textures. Plausibly different sequences from the same patient, but insufficient textual evidence to confirm fabrication.

Evidence highlights

Reference [1]: Coit DG, Thompson JA, Algazi A, et al. "Melanoma, version 2.2016, NCCN clinical practice guidelines in oncology." Used to support the statement that the standard treatment of LARC includes neoadjuvant chemoradiotherapy.
Reference [23]: cited twice, once for PyRadiomics and once for CellProfiler.
Reference [7]: cited as evidence on pathologic complete response rates, appears to be a conference abstract (PTH-340, pages A558-9).
Statistical Analysis section: explicit mention of "R software (version 3.1.0)" in a manuscript with a 2020 publication date.
Table 2: Sensitivity 100.0 [100.0–100.0] reported for TRG2 in multiple cohorts, including VC3 (N=48).
Figures 2c/2f/2i: staircase ROC curves, AUC >0.99.
Figure 4: unusually smooth, regular survival curve separation.
Ethics approval reference cited in the report: 2019ZSLYEC-169 (requires institutional verification).

Notes

All numeric figures and citations above are preserved exactly as reported; no new claims have been added.
The reference [1] melanoma citation is the most severe single indicator because it is internally inconsistent with the topic of the manuscript and is difficult to explain by ordinary error.
The duplicate [23] citation and the R 3.1.0 version are consistent with text-level copy-paste from older or unrelated manuscripts.
The 100.0 [100.0–100.0] sensitivity in a small validation set (N=48) is statistically improbable in a real-world multicenter radiopathomic cohort and raises legitimate concerns about data leakage, cherry-picking, or post-hoc figure manipulation, but does not by itself prove misconduct.
The Figure 1 concerns remain unverified due to lack of higher-resolution images or supporting textual contradictions.
Recommended next steps, per the original report: request raw data, code, and training logs from the authors; raise concerns on PubPeer emphasizing the citation mismatch and statistical anomalies; file a formal complaint with the journal editorial office and request verification of ethics approval 2019ZSLYEC-169.
DOI: 10.1245/s10434-020-08659-4