Data Anomaly Investigation: Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for CheX-ray Vision-Language Pre-training

Academic fraud report · Geng Detector

Summary

Verdict: Questionable (suspected data-entry error, insufficient grounds for fabrication). The investigation centers on Table 1 of the paper (DOI: 10.1145/3746027.3755336) by Qiao et al., published at ACM MM '25. The principal finding is that the authors' ViT-based model reports identical AUC scores of 89.5 on the CheXpert dataset under both the 1% and 10% training-data regimes, with only the 100% setting showing a marginal change to 89.7. Such exact equality across a tenfold increase in labeled training data is statistically implausible for linear-probe evaluation, and is most plausibly explained by a copy-paste transcription error rather than deliberate fabrication, though intent cannot be ruled out without raw experiment logs. A secondary observation is that multiple table cells appear visually or parser-wise merged (e.g., MRM's values rendered as '82.784.4', Med-UniC* as '90.891.993.193.7'), reflecting either tight typesetting or PDF extraction artifacts. No image-based findings were assessed because pixel-level data was unavailable. Confidence in the copy-paste hypothesis is moderate; confidence in scientific misconduct is low without author clarification.

Verdict

Questionable. The reported equality of 89.5 AUC at both 1% and 10% CheXpert training splits for the authors' ViT-based model is highly atypical and most consistent with a data-entry/transcription error. There is insufficient evidence to support a finding of deliberate falsification. Authors should be requested to provide raw evaluation logs.

Key findings

Identical metric values across training-set scales: In Table 1, the proposed method's CheXpert AUC is 89.5 at 1% training data and 89.5 at 10% training data, differing only at 100% (89.7).
Comparison with baselines: Competing methods such as Med-UniC show monotonically increasing or fluctuating trends across splits (e.g., 89.4 → 89.7 → 90.8), highlighting the anomaly.
Formatting/parsing artifacts: Several cells in Table 1 are concatenated without separators (e.g., 82.784.4, 90.891.993.193.7), indicating either compressed typesetting or PDF-to-text conversion errors.
Image-based analysis not performed: Pixel-level inspection of t-SNE and attention maps (Figures 5–6) was not feasible from the available text-only extraction.

Evidence highlights

Source table cell values: "Ours — CXP (AUC): 1% → 89.5, 10% → 89.5, 100% → 89.7."
Baseline reference for contrast: Med-UniC CXP (AUC): 1% → 89.4, 10% → 89.7, 100% → 90.8.
Conjoined numerical strings extracted from Table 1: 82.784.4 (MRM, ChestX-ray14) and 90.891.993.193.7 (Med-UniC*, RSNA).
DOI of investigated work: 10.1145/3746027.3755336.
Severity: orange (moderate) for the data duplication; yellow (low) for the formatting issue; insufficient evidence for image-related claims.

Notes

The report's recommended follow-up actions include contacting the authors for raw experiment logs and inspecting the public repository at https://github.com/Felix1118/PADKB for potential data-leakage or metric-computation bugs.
Limitations: No image forensics were conducted; conclusions rely solely on textual/numerical evidence from the manuscript.
All numeric values above are reproduced exactly as cited in the source report. No additional figures or findings have been inferred.