Integrity Review Report: Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training (DOI: 10.1145/3746027.3755336)

Academic fraud report · Geng Detector

Summary

Verdict: Questionable but no evidence of substantive academic fraud. This 2025 ACM Multimedia conference paper by Lihong Qiao et al. presents a vision-language pre-training framework combining reconstruction (PAR) and contrastive alignment (DKBA) modules for chest X-ray tasks. Cross-validation of Tables 1, 4, and 6 shows ablation results that align with established theory: DKBA mainly improves classification (e.g., COVIDx AUC rising from 77.0 to 81.9) while PAR primarily aids fine-grained tasks (Dice/mAP), avoiding the 'everything improves uniformly' red flag of fabricated numbers. The principal concerns are minor: a typographical error ('abesent' instead of 'absent') in Section 3.1, and missing spaces in Table 1 (MRM, Med-UniC) and Table 5 (KAD). Image pixel-level review was not possible from the supplied text-only extract, but figure cross-references are internally consistent. Code is publicly released at https://github.com/Felix1118/PADKB, supporting reproducibility. Confidence is moderate; limitations include inability to verify figures or run code.

Verdict

🟡 Questionable / Minor concerns only — No evidence of substantive academic fraud detected. Issues are confined to proofreading and formatting negligence.

Key findings

Data logic consistency (Tables 1, 4, 6): Ablation outcomes follow expected Vision-Language Pre-training theory — reconstruction helps fine-grained segmentation/detection tasks; contrastive alignment boosts global classification. This avoids the suspicious 'all-metrics-improve-uniformly' pattern common in fabricated deep-learning results.
Textual/typographical errors (Section 3.1, Tables 1 & 5):
'abesent' instead of 'absent' in Section 3.1.
Missing spaces in Table 1 MRM row: '88.5 88.5'.
Missing spaces in Table 1 Med-UniC row, e.g., '90.891.993.7'.
Missing spaces in Table 5 KAD row: '66.020.9'.
Figure integrity (Figures 1–6): Not directly inspectable from text extract; in-text figure references are internally consistent with captions.
Reproducibility signal: Authors released code at https://github.com/Felix1118/PADKB — a strong credibility indicator for a CV paper.
No numerical paradoxes detected in Tables 1, 4, or 6.

Evidence highlights

DOI: 10.1145/3746027.3755336
Venue: ACM International Conference on Multimedia (MM '25)
Specific value audited: COVIDx classification metric rises from 77.0 → 81.9 when DKBA is added (Table 4), consistent with claimed role of discriminative knowledge boosting for classification.
Author-provided GitHub: https://github.com/Felix1118/PADKB
Authors: Lihong Qiao, Shiyi Gao, Yucheng Shu, Bin Xiao, Weisheng Li, Xinbo Gao.

Notes

Pixel-level image manipulation analysis (Form 1/3) is not feasible from the supplied PDF text extract; this is a limitation of the current review.
The typographical defects alone are insufficient to infer data fabrication and are more parsimoniously explained by carelessness.
Recommended follow-up: spot-check the GitHub repository for code–paper consistency; consider an erratum for the spelling and table spacing issues.
Final adjudication of academic misconduct requires institutional investigation; this automated review is advisory only.