Geng Integrity Report — Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training (DOI: 10.1145/3746027.3755336)

Academic fraud report · Geng Detector

Summary

Verdict: Suspect (中等存疑). This report evaluates the ACM MM '25 paper by Lihong Qiao et al. and flags two confirmed concerns. First, Table 1 (ViT-based row, 'Ours') reports AUC = 89.5 on CheXpert at both 1% and 10% training data, an implausible exact match across a 10x data scale change that suggests either fabricated numbers or copy-paste error. Second, Section 3.1 states that an original chest X-ray is 'down-sample[d]' to create both high-resolution and low-resolution versions, a clear methodological/logical contradiction since downsampling can only reduce resolution. Image reuse analysis was not possible due to lack of high-resolution figures. Overall the findings are limited to textual and tabular evidence; no institutional investigation has been conducted, and authorship rebuttal is pending. Confidence in Finding 1 and 2 is high based on direct quotations, but a definitive misconduct determination requires raw experimental logs and author clarification.

Verdict

🟡 Suspect — Two confirmed textual/numerical inconsistencies identified; image-level analysis pending; not yet a confirmed misconduct ruling.

Key findings

Finding 1 — Suspiciously identical AUC values across data scales (Severity: 🟡, Confirmed): In Table 1, the ViT-based 'Ours' row reports an AUC of 89.5 on CheXpert (CXP) for both 1% and 10% training data. A 10× increase in training data yielding an identical AUC to one decimal place is statistically implausible and suggestive of fabricated numbers or copy-paste error.
Finding 2 — Methodological/logical contradiction in Section 3.1 (Severity: 🟡, Confirmed): The paper states, *"We down-sample an original chest X-ray image to create both high-resolution and low-resolution versions."* Downsampling by definition reduces spatial resolution and cannot produce a high-resolution version of the same image.
Finding 3 — Image reuse analysis not possible (Severity: N/A, Insufficient evidence): Figures 1–6 were not available at sufficient resolution for pixel-level or splicing checks.

Evidence highlights

DOI: 10.1145/3746027.3755336
Table 1, ViT-based, 'Ours' row, CheXpert (AUC) column: 1% = 89.5, 10% = 89.5 (exact match to one decimal).
Section 3.1 (Page 4), direct quote: *"We down-sample an original chest X-ray image to create both high-resolution and low-resolution versions."*
Authors: Lihong Qiao, Shiyi Gao, Yucheng Shu, Bin Xiao, Weisheng Li, Xinbo Gao.
Venue: Proceedings of the 33rd ACM International Conference on Multimedia (MM '25), 2025.

Notes

Both confirmed findings are textual/tabular only; no raw training logs, code, or checkpoints were inspected.
Image manipulation (PSNR/ELA, splicing artifacts in architecture diagrams, t-SNE plots, heatmaps) could not be assessed without source files.
Recommended next steps: (1) request from the authors the original CXP experiment logs and random seeds for the 1% and 10% splits; (2) raise a PubPeer comment seeking clarification on the identical AUC and the downsampling wording; (3) defer institutional and editorial escalation pending author response.
This report is AI-assisted and intended for academic discussion only; final misconduct determinations require formal investigation.