Integrity Review Report: Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Chest X-ray Vision-Language Pre-training

Academic fraud report · Geng Detector

Summary

Verdict: Suspicious (insufficient evidence for a definitive call). The review examined the ACM MM '25 paper (DOI: 10.1145/3746027.3755336) by Qiao et al. across three dimensions: image reuse/tampering, statistical anomalies, and timeline/equipment verification. No statistical red flags were found: Tables 1–6 show plausible monotonic gains (1%–3% margins) across 1%/10%/100% data ratios, with no terminal-digit anomalies or arithmetic inconsistencies. Timeline logic is internally consistent (RTX 4090 GPUs, 2024 NeurIPS reference, October 2025 conference date). The principal limitation is that the review was conducted on text alone; the paper consists of computer-vision figures (architectures, t-SNE, attention maps) whose forensic analysis (ELA, splicing detection) requires pixel data, which was unavailable. No fabricated references, impossible equipment, or anachronistic citations were detected. Recommendation: verify reproducibility via the linked GitHub repository; no institutional referral warranted at this time.

Verdict

🟡 Suspicious / Undetermined — No substantive misconduct indicators detected in textual and tabular content, but image-level forensic analysis was not possible due to the absence of pixel data.

Key findings

Statistical data (Tables 1–6): No red flags. Reported AUC/ACC and detection metrics across data ratios (1%, 10%, 100%) show plausible monotonic improvement margins (~1%–3%), consistent with current SOTA-chasing norms in CV conference papers.
Numeric hygiene: No terminal-digit bias, no obvious arithmetic artifacts (e.g., fabricated arithmetic progressions), and no impossible jumps.
Reporting conventions: No p-values, standard deviations, or confidence intervals reported — common in top-tier ML venues, though it limits independent reproducibility assessment.
Timeline & equipment: RTX 4090 GPUs and a 2024 NeurIPS reference (Ref [34] Med-UniC) are temporally consistent with the October 2025 ACM MM '25 publication date. No anachronistic equipment or citation inconsistencies found.
Figures (Fig. 1–6): Not analyzable. Image reuse, splicing, or ELA-based tampering checks require pixel data, which the reviewer did not have access to.

Evidence highlights

Paper title: *Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Chest X-ray Vision-Language Pre-training*
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
DOI: https://doi.org/10.1145/3746027.3755336
Claimed compute: 2× NVIDIA RTX 4090 GPUs (Section 4.3)
Code link cited: https://github.com/Felix1118/PADKB
Performance margins observed across data scales are within the 1%–3% range typical for incremental CVPR/ICCV/MM-class contributions.

Notes

Confidence: moderate on textual/tabular findings; low on image integrity (untestable without figures).
Scope limitation: only the manuscript text was reviewed; figures, supplementary materials, and the GitHub repository were not independently examined.
Recommended follow-up: (1) clone and attempt to reproduce Tables 1–3 from the linked repository; (2) scrutinize generalization claims and potential overfitting to public chest X-ray benchmarks; (3) consider PubPeer discussion on real-world robustness.
Current evidence does not support a referral to an institutional research-integrity office.