Integrity Review Report: Dual-Energy CT Deep Learning Radiomics to Predict Macrotrabecular-Massive Hepatocellular Carcinoma

Academic fraud report · Geng Detector

Summary

Verdict: Cleared (textual and arithmetic audit only). This Radiology paper (DOI: 10.1148/radiol.230255) by Mengsi Li, Yaheng Fan, Bingsheng Huang, Jin Wang et al. was assessed via cross-verification of reported statistics, timeline consistency, and disclosure of marginally significant variables. All checked arithmetic is internally consistent: sensitivity/specificity fractions reconcile exactly (e.g., 42/44=95.45%, 79/102=77.45%, 24/30=80%, 46/51=90%); likelihood ratios derived from these values match Table 4 (4.23 and 0.06); and the AFP coefficient (β=1.86, e^β≈6.42) matches the reported OR (6.43). The follow-up window (enrollment June 2019–June 2022, follow-up cutoff March 31, 2023, submission February 2023) is logically tight but feasible. P-values for Sex (P=.05) and Platelet count (P=.08) are honestly reported rather than rounded to significance. Limits: this audit is text-only; CT images, radiomics features, and deep-learning saliency maps could not be inspected at the pixel level.

Verdict

Cleared based on a textual, arithmetic, and timeline audit. No substantive integrity concerns were identified within the scope of this analysis.

Key findings

All reported sensitivity/specificity fractions in Table 4 reconcile precisely with the underlying counts (42/44, 79/102, 24/30, 46/51).
Likelihood ratios in Table 4 footnoted formula reproduce the paper's values when computed from the reported sensitivity and specificity (4.23 and 0.06).
Logistic regression coefficient for AFP level (β=1.86) yields an odds ratio consistent with the reported OR (6.43 vs 6.42 expected).
The authors transparently report variables with P-values at or above 0.05 (Sex P=.05; Platelet count P=.08), rather than manipulating them below the threshold.
AUC values (0.85–0.91 range) and confidence interval widths (e.g., internal test set 0.72–1.00) reflect realistic small-sample variability rather than suspiciously perfect performance.
Software versions (ITK-SNAP v3.8.0, UCTransNet AAAI 2022 reference, PyRadiomics v3.0.1) are temporally consistent with the 2019–2023 experimental window.
The follow-up timeline (enrollment June 2019–June 2022; follow-up cutoff March 31, 2023; submission February 2023; revision June 2023; publication August 2023) is internally coherent.

Evidence highlights

Table 4 training set: sensitivity 95% [42/44] → 42÷44 = 95.45%; specificity 77% [79/102] → 79÷102 = 77.45%; positive LR = 0.9545/(1−0.7745) = 4.23 (matches); negative LR = 0.0455/0.7745 ≈ 0.058 ≈ 0.06 (matches).
Table 4 external test: sensitivity 80% [24/30] and specificity 90% [46/51] — both verified.
Table 3: AFP β=1.86 → e^1.86 ≈ 6.42; reported OR = 6.43 (within rounding tolerance).
DOI: 10.1148/radiol.230255
Journal: Radiology, Volume 308, Number 2, August 2023.

Notes

This review is text-only; CT images, radiomics feature matrices, and deep-learning saliency/heatmap outputs were not subjected to pixel-level forensic analysis.
No image reuse, splicing, duplication, or rotation artifacts could be evaluated; absence of findings here reflects scope limits, not a confirmed clean bill of health for the imaging figures.
The arithmetic and disclosure patterns observed are consistent with genuine statistical software output rather than fabricated data.