Investigation Report on "Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training" (ACM MM '25, DOI: 10.1145/3746027.3755336)

Academic fraud report · Geng Detector

Summary

This report evaluates a chest X-ray vision-language pre-training paper published at ACM Multimedia 2025, with a verdict of highly suspicious due to internal numerical inconsistencies and logically impossible ablation results. Key issues: (1) the main results table (Table 1) reports AUC of 81.4 on ChestX-ray14 with 1% training data for the proposed model, while Tables 4 and 5 under identical configurations report 80.8, a 0.6-point internal discrepancy; (2) Table 4 ablation shows that adding the headline module PAR on top of DKBA reduces AUC from 81.0 to 80.8, directly contradicting the paper's claim that PAR contributes complementary gains. Confidence in the textual/numerical findings is high because the inconsistencies are verifiable from reported tables alone, but image-level manipulation could not be assessed. Limitations include lack of access to code, raw logs, and high-resolution figures.

Verdict

🟠 Highly suspicious. The paper contains internally contradictory AUC values for the same experimental configuration across its own tables, and an ablation in which the headline module (PAR) is shown to degrade performance relative to its predecessor (DKBA alone), contrary to the claims in the text.

Key findings

Internal AUC conflict on ChestX-ray14 (1% training data):
Table 1 (main linear classification): "Ours" = 81.4
Table 4 (ablation, full PAR + DKBA model): 80.8
Table 5 (ablation, MGCA prior setting): 80.8
The same configuration (1% ChestX-ray14 + full proposed model) reports 81.4 in one table and 80.8 in two others, a gap of 0.6 AUC points.
Ablation violates stated module contribution: Table 4 reports Baseline = 79.1, +DKBA = 81.0, +DKBA+PAR = 80.8 on ChestX-ray14. Adding the title-promoted module PAR *decreases* AUC by 0.2, yet the text asserts that PAR provides complementary benefits (citing ~2% gains on ChestX-ray14).
Image-level analysis not verifiable: Figures 1–6 (architecture, heatmaps, t-SNE) could not be pixel-checked from the text-only source; potential cherry-picked t-SNE visualizations cannot be excluded.

Evidence highlights

Quantitative conflict (exact values): 81.4 (Table 1, "Ours", ChestX-ray14 1%) vs. 80.8 (Table 4 full model, same setting) vs. 80.8 (Table 5 MGCA prior, same setting).
Logical/arithetic inconsistency in ablation order on ChestX-ray14: 79.1 (Baseline) → 81.0 (+DKBA) → 80.8 (+DKBA+PAR). The +PAR step is negative, contradicting the paper's narrative and its titled contribution "Pathology-Aware Reconstruction."
Paper text (paraphrased) claims DKBA yields ~2% gains on ChestX-ray14, but Table 4 shows +DKBA improves over Baseline by ~1.9 points (79.1→81.0), while further adding PAR *erases* 0.2 of that gain.
DOI verified: 10.1145/3746027.3755336 (ACM MM '25).

Notes

Scope limitations: only text and reported tables were available; no code, training logs, or high-resolution figure files were accessible. Pixel-level manipulation checks (duplication, splicing, selective t-SNE cropping) were therefore not performed.
The internal numerical inconsistencies are, however, directly verifiable from the paper's own tables and do not depend on image analysis.
Recommendations from the source report remain appropriate: request raw training logs, raise a PubPeer comment, and notify the ACM MM '25 program committee.
This is not a final determination of misconduct; institutional investigation is required for any formal finding.