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: Cleared (no substantive issues identified). This review examined the ACM MM '25 paper by Qiao et al. (DOI: 10.1145/3746027.3755336) for potential academic fraud, focusing on data fabrication, timeline plausibility, statistical anomalies, and image manipulation. Cross-checking Table 4 (ablation study) against Table 5 (prior model variants) revealed fully consistent best-case figures (ChestX-ray14 AUC 80.8, COVIDx ACC 82.3, SIIM Dice 66.9, RSNA mAP 21.7), indicating internal logical coherence rather than fabrication. Hardware specifications (two NVIDIA RTX 4090 GPUs, released 2022) are temporally appropriate for a 2025 publication. Performance gains over SOTA baselines (Med-UniC, MRM, MGCA) fall within the expected 0.5%–3.0% range, consistent with diminishing returns typical of VL pretraining research. Image-based forensics (Western blot bands, microscopy noise, splicing traces) could not be performed because only the text stream was provided. Overall confidence is moderate-to-high; limitations stem from absence of source figures and raw data.

Verdict

Cleared. No evidence of data fabrication, timeline anomalies, or statistical irregularities was found. Image-level analysis could not be completed due to lack of source files.

Key findings

Internal data consistency: Table 4 (full PAR+DKBA configuration) and Table 5 (MGCA-prior configuration) share identical best-case values across all four benchmarks, confirming a coherent experimental setup rather than ad-hoc fabrication.
Plausible performance gains: Improvements over SOTA baselines (Med-UniC, MRM, MGCA) range from ~0.5% to ~3.0% on AUC, ACC, Dice, and mAP, consistent with typical diminishing-returns behavior in medical vision-language pretraining.
Hardware/timeline consistency: Reported use of 2× NVIDIA RTX 4090 GPUs is temporally valid (released 2022) for a 2025 submission; no "future hardware" or anachronistic anomalies detected.
Image forensics unavailable: Figures 1–6 (including t-SNE and zero-shot visualizations) could not be pixel-inspected because only text content was provided.
No suspicious narrative patterns: No exaggerated claims such as "uniform 5% SOTA gains across all datasets."

Evidence highlights

Table 4 best row vs. Table 5 MGCA row — exact match: ChestX-ray14 AUC 80.8, COVIDx ACC 82.3, SIIM Dice 66.9, RSNA mAP 21.7.
Implementation Details (Section 4.3): 2× RTX 4090, MIMIC-CXR, 200 epochs Reconstruction + 15 epochs Alignment.
Comparative tables (1–3): incremental gains of 1%–3% relative to MGCA and Med-UniC.

Notes

Review is constrained to textual evidence; visual artifacts (Western blot duplication, splicing, noise inconsistency) in Figures 1–6 remain unverified.
Per-step gains around 1%–2% are within normal SOTA-improvement margins and should not be interpreted as suspicious on their own.
Final institutional determination of misconduct requires formal investigation; this automated review is advisory only.