Integrity Review Report: Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training

Academic fraud report · Geng Detector

Summary

This review examines the ACM MM '25 paper (DOI: 10.1145/3746027.3755336) for potential academic integrity concerns. The verdict is 'Questionable' (yellow). A previous allegation of contradictory mAP values between Table 4 and Table 5 was retracted upon re-examination; the MGCA rows match correctly. However, Table 1 (linear classification) contains suspiciously regular numeric patterns. On COVIDx (ACC), the reported method rises from 82.3 to 92.3 to 96.0 across 1%/10%/100% data regimes, including an exact +10.0 point jump from 1% to 10%. On CheXpert (AUC), the method reports identical 89.5 values at both 1% and 10% data scales. Genuine deep-learning training rarely produces such integer-perfect, monotonically clean improvements due to stochasticity. Pixel-level analysis of figures was not possible because only text was supplied. Confidence in the regularity concern is moderate; no definitive fabrication proof exists without author logs or raw runs.

Verdict

Questionable (yellow). One substantiated concern (implausibly regular Table 1 numbers) and one previous allegation retracted after re-examination. No conclusive proof of fabrication.

Key findings

Table 1 contains suspiciously regular numeric patterns that are atypical for deep-learning benchmark results.
COVIDx ACC progression: Ours (1%) = 82.3 → Ours (10%) = 92.3 → Ours (100%) = 96.0, including an exact +10.0 percentage-point gain from 1% to 10% data.
CheXpert AUC duplication: Ours reports identical 89.5 at both 1% and 10% data regimes, a value repetition that is unlikely under realistic stochastic training.
Retracted allegation: The earlier claim of conflicting mAP values between Table 4 and Table 5 for MGCA was found to be a misreading; the values align correctly and are not evidence of misconduct.
Figure analysis limited: Pixel-level inspection of architecture diagrams and t-SNE plots (Figures 1–6) was not performed because original high-resolution images were not available.

Evidence highlights

Location: Table 1, Linear classification results.
COVIDx ACC column: 82.3, 92.3, 96.0 across 1%, 10%, 100% (exact 10.0-point jump between the first two rows).
CheXpert AUC column: 89.5 repeated identically for the 1% and 10% rows of the proposed method.
DOI preserved: 10.1145/3746027.3755336.

Notes

The regularity in Table 1 is statistically unusual but not, by itself, proof of data fabrication; common benign causes include aggressive rounding, post-hoc cherry-picking of a single seed, or template entry errors.
The reviewer recommends caution when using the COVIDx and CheXpert numbers reported in Table 1 pending verification.
A definitive ruling would require the authors' raw training logs, random seeds, and multiple-seed replicates.
Confidence in the regularity-based suspicion: moderate. Confidence in the retraction of the mAP-conflict claim: high.
This report is AI-assisted and intended for academic discussion only; final determination of misconduct requires an institutional investigation.