Data Integrity Concerns in "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 evaluates the paper by Lihong Qiao et al. (DOI: 10.1145/3746027.3755336), published at ACM Multimedia 2025, and assigns a verdict of Highly Suspicious. Five issues are flagged, with three confirmed as substantive. First, in Table 1, the proposed method yields an identical AUC of 89.5 on the CheXpert dataset under both 1% and 10% training data regimes, an outcome that is statistically implausible for deep learning fine-tuning at a tenfold difference in data volume. Second, the ablation study (Table 4) exhibits internal mathematical inconsistency: adding the PAR module to the baseline on ChestX-ray14 reduces performance from 79.1 to 78.9, yet the combined PAR+DKBA model scores 80.8, lower than DKBA alone at 81.0, violating the claimed complementary design. Third, COVIDx accuracy rises from 82.3 (1%) to 92.3 (10%), a suspiciously exact integer delta of 10.0. Fourth, an uncited mention of "ReCO" in Related Work suggests possible paraphrased copying. Visual image-level forensics were not possible because only text was available. Confidence is moderate-to-high for the numerical anomalies; the uncited reference and cherry-picking concerns remain unverified without additional evidence.

Verdict

Highly Suspicious (🟠) — Multiple confirmed statistical and logical anomalies in the reported results, combined with an uncited reference, indicate a non-trivial probability of data fabrication or manipulation. Visual-level forensics could not be performed because only text was available; confirmation would require raw training logs and original figures.

Key findings

Table 1 (CheXpert / ViT linear probing): Identical AUC of 89.5 reported for both the 1% and 10% training-data settings. A tenfold increase in training data producing identical, single-decimal AUC is statistically implausible.
Table 4 (Ablation, ChestX-ray14): Baseline = 79.1; +PAR = 78.9 (decreases); +DKBA = 81.0; PAR+DKBA = 80.8 (lower than DKBA alone). The combined model underperforming one of its components contradicts the paper's narrative of complementary modules.
Table 4 (Ablation, SIIM): Adding DKBA to the baseline drops Dice from 64.9 to 64.5, inconsistent with the module's claimed benefit.
Table 1 (COVIDx): Accuracy jumps from 82.3 (1%) to 92.3 (10%), a suspiciously clean integer delta of exactly 10.0.
Section 2.1 (Related Work): The method name "ReCO" is mentioned without any citation marker, and no corresponding entry appears in the reference list, raising concerns of unreferenced paraphrasing.

Evidence highlights

CheXpert AUC reproducibility: 89.5 → 89.5 across a 10× change in training-data fraction (Table 1).
ChestX-ray14 ablation paradox: 79.1 → 78.9 (+PAR) → 81.0 (+DKBA) → 80.8 (PAR+DKBA) — combined score is lower than the strongest single-component variant (Table 4).
COVIDx ACC step: 82.3 → 92.3, Δ = 10.0 exactly (Table 1).
ReCO mentioned in §2.1 with no bracketed citation and no entry in the bibliography.
DOI: 10.1145/3746027.3755336 preserved for traceability.

Notes

Visual checks (Western-blot-style noise analysis, pixel-level duplication, splicing) were not applicable in this pass because only text extracted from the PDF was available. Figure 5's clustering claims could not be independently verified.
Findings 1–4 are marked as confirmed; Finding 5 (possible image cherry-picking) is flagged as insufficient evidence pending access to original high-resolution figures.
The uncited "ReCO" reference and the COVIDx integer delta (10.0) are treated as supporting indicators rather than conclusive proof on their own; their weight comes from co-occurrence with the CheXpert AUC duplication and the ablation paradox.
Recommended actions: request raw training logs and TensorBoard traces from the authors; raise the statistical anomalies on PubPeer; ask the corresponding author to clarify the "ReCO" citation and run a similarity check on §2.1; consider a broader audit of the team's prior publications for systematic irregularities.
This is an AI-assisted preliminary assessment; final determination of misconduct requires investigation by the publisher or institutional body.