Integrity Assessment: Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for CheXray Vision-Language Pre-training (ACM MM '25)

Academic fraud report · Geng Detector

Summary

This report assesses the paper "Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training" by Lihong Qiao et al., submitted to ACM MM '25 (DOI: 10.1145/3746027.3755336). The overall verdict is highly suspicious. Three main concerns are documented: (1) Implausibly identical AUC scores (89.5) for the proposed model on the CheXpert dataset at both 1% and 10% training-data ratios in Table 1, which is statistically unreasonable given the ~10x increase in training samples; (2) a mathematical inconsistency between the text claim of a 2.0% AUC improvement over MRM on ChestX-ray14 (1%) and the actual table-derived difference of 2.6% (81.4 vs 78.8); and (3) formatting anomalies in Med-UniC's Table 1 row, where digits appear concatenated without proper spacing (e.g., 90.891.9), suggesting hasty data entry. Image-based analyses (e.g., t-SNE plots, heatmaps in Figures 1–6) could not be evaluated. Confidence is moderate-to-high for the textual/numerical findings; conclusions should be confirmed by inspecting raw logs and code.

Verdict

🟠 Highly Suspicious. Three independent concerns regarding data plausibility, numerical consistency, and table formatting were verified. No image-level forensics were performed.

Key findings

Implausible identical AUC (89.5) at 1% and 10% training-data regimes on CheXpert, suggesting possible fabrication or copy-paste of results.
Mathematical mismatch: text claims a 2.0% AUC gain over MRM on ChestX-ray14 (1%), but Table 1 shows 81.4 − 78.8 = 2.6%.
Malformed Med-UniC row in Table 1, with digits concatenated (e.g., 89.4 89.7 90.891.993.193.780.389.594.5), indicating rushed manual data entry.
Image-based checks (duplication, splicing, t-SNE manipulation in Figures 1–6) could not be performed due to lack of pixel access.

Evidence highlights

Table 1 (CXP, ViT-based): CheXpert AUC values for the proposed model reported as 89.5 at both 1% and 10% training fractions; the 1% → 10% expansion (roughly an order of magnitude in samples) typically produces non-trivial metric variation, making identical values to one decimal place highly improbable.
Section 5.1 vs Table 1 (ChestX-ray14, 1%): Stated improvement "2.0%" over MRM does not match the computed gap of 81.4 − 78.8 = 2.6% (Δ = 0.6 percentage points).
Table 1 (Med-UniC row, ViT-based): Token stream reads 89.4 89.7 90.891.993.193.780.389.594.5; expected values such as 90.8 and 91.9 are visually merged as 90.891.9, breaking the columnar structure of the table.
DOI: 10.1145/3746027.3755336 (ACM MM '25).

Notes

The 89.5/89.5 coincidence is strong evidence of suspicious reporting but is not definitive proof of fabrication; confirmation requires raw training logs, random seeds, and run-to-run variance estimates.
The 2.0% vs 2.6% discrepancy could alternatively reflect a last-minute edit to one location that was not propagated; nonetheless, it raises concerns about data integrity practices.
The Med-UniC formatting issue is consistent with manual LaTeX edits rather than automated table generation; alone it is suggestive but not conclusive.
Image forensics, code reproducibility checks, and author response are recommended next steps before any formal allegation of misconduct.