Geng Integrity Report: Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for CheXray Vision-Language Pre-training

Academic fraud report · Geng Detector

Summary

This report assesses the ACM Multimedia 2025 paper (DOI: 10.1145/3746027.3755336) by Lihong Qiao et al. for possible data fabrication and citation irregularities. The central finding concerns Table 1, where CheXpert (CXP) ViT-based AUC values show an implausibly constant +1.0 improvement over the MRM baseline across all three data scales (1%, 10%, 100%), while the MRM values themselves (88.5, 88.5, 88.7) are nearly invariant, inconsistent with stochastic deep-learning experiments. This pattern is not mirrored in other datasets in the same table. A secondary finding flags a citation inconsistency: reference [34] (Med-UniC) is attributed to NeurIPS vol. 36 (2024), but vol. 36 corresponds to NeurIPS 2023, raising timeline concerns. Image-level forensics were inapplicable as the paper contains no biomedical images. Confidence is moderate: the arithmetic anomaly is concrete but could in principle stem from selective reporting or rounding rather than fabrication. Recommended actions include demanding raw logs, posting on PubPeer, auditing public code, and notifying the conference ethics chair.

Verdict

🟠 Highly Suspicious. The paper presents numerical evidence in Table 1 that is statistically implausible for independent deep-learning runs, alongside a citation error concerning the volume–year mapping of NeurIPS. Image-level forensics are not applicable. Final determination requires institutional investigation.

Key findings

Suspiciously constant baseline gap (Table 1, CheXpert / ViT). Across 1%, 10%, and 100% training data, the reported AUC gap over MRM is exactly +1.0 (89.5 vs 88.5; 89.5 vs 88.5; 89.7 vs 88.7), while the MRM baseline itself is nearly frozen (88.5, 88.5, 88.7). Other datasets in the same table show irregular, dataset-specific deltas, making the CXP pattern an outlier.
Citation/year mismatch in reference [34]. The paper cites Med-UniC as *Advances in Neural Information Processing Systems 36 (2024)*. NeurIPS vol. 36 corresponds to the 2023 conference (held December 2023); NeurIPS 2024 would be vol. 38. The error affects a core baseline cited throughout the comparison tables.
No biomedical images to analyze. Figures 1, 2, 5, and 6 are algorithm-generated visualizations (heatmaps, t-SNE), which cannot be evaluated with Western-blot-style pixel forensics.

Evidence highlights

Table 1, CheXpert (CXP), ViT-based AUC (paper vs. MRM):
1% data: paper 89.5 vs. MRM 88.5 (Δ +1.0)
10% data: paper 89.5 vs. MRM 88.5 (Δ +1.0)
100% data: paper 89.7 vs. MRM 88.7 (Δ +1.0)
Cross-dataset comparison (same table) shows non-constant gaps, e.g., ChestX-ray14 deltas of +2.6, +1.1, +0.6 and RSNA deltas of +1.2, +0.9, +1.1, contrasting sharply with the CXP pattern.
Reference [34] metadata: cited as *NeurIPS 36 (2024)*; correct mapping is NeurIPS 36 → 2023, NeurIPS 38 → 2024.
DOI of the paper under review: 10.1145/3746027.3755336 (ACM MM '25).
Source PDF: acmmm-submission.pdf.

Notes

The constant +1.0 gap is *suggestive* of hand-adjusted or templated reporting but is not, on its own, proof of fabrication. Plausible alternative explanations include selective reporting of best seeds, post-hoc rounding, or a data leakage that produces uniform lift. Distinguishing these requires the original training logs, seeds, and evaluation scripts.
The authors have a public repository referenced in the report (Felix1118/PADKB on GitHub) that could be audited for hard-coded offsets; this was not independently verified here.
The reference [34] error could be a typographical slip, but because Med-UniC is a key comparator used in the headline tables, the misattribution is material and should be corrected regardless of intent.
Recommended next steps: request 10 independent runs with seeds and standard deviations for the CXP row; raise a PubPeer comment focusing on the constant-delta anomaly; audit the public evaluation code for hard-coded biases; and notify the ACM MM 2025 program chairs / ethics committee of both the numerical irregularity and the citation discrepancy.
Limitations: this analysis is based on a single submission PDF; no raw logs, code, or reviewer comments were available. Image-based duplication checks were not applicable to the paper's content.