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

Academic fraud report · Geng Detector

Summary

Verdict: Cleared / no fraud indicators detected. This ACM Multimedia (MM '25) paper by Lihong Qiao et al. was reviewed across image reuse, data consistency, statistical honesty, timeline plausibility, and writing provenance. Image-level (pixel) analysis could not be performed because the submitted text lacked usable figure data; this is a review-environment limitation, not a paper issue. Statistically, the main results (Table 1) and ablation tables (Tables 4, 5, 6) are internally consistent across configurations sharing the same setting, with ChestX-ray14 AUC = 80.8, COVIDx ACC = 82.3, SIIM Dice = 66.9, and RSNA mAP = 21.7 reproducing exactly where appropriate — suggesting careful record-keeping rather than fabrication. The authors also honestly report underperformance on CheXpert (89.5 AUC under both 1% and 10% settings) and attribute it to Med-UniC's use of PadChest, an unusual but credibility-enhancing disclosure. Timeline (RTX 4090, 2024 references including CARZero and Med-UniC) is internally coherent. A minor typo ("abesent") is interpreted as a natural human artifact. Overall confidence is moderate given the inability to inspect figures directly.

Verdict

Cleared. No substantive indicators of academic fraud were identified. The paper passes cross-table consistency checks, timeline plausibility, and methodological transparency tests. A minor typo is noted but is non-incriminating.

Key findings

Data self-consistency (positive indicator): Across Tables 1, 4, 5, and 6, the configuration Ours (PAR + DKBA) at 1% training data yields identical values: ChestX-ray14 AUC = 80.8, COVIDx ACC = 82.3, SIIM Dice = 66.9, RSNA mAP = 21.7. This consistency across independent ablation dimensions is consistent with careful experimental record-keeping rather than fabricated data.
Transparent negative results (positive indicator): In Table 1, CheXpert results (89.5 at both 1% and 10% settings) are reported, and Section 5.1 explicitly acknowledges underperformance vs. Med-UniC, attributing it to Med-UniC's use of the additional PadChest dataset. This open acknowledgment of a competitor's advantage is atypical of fraudulent work.
Timeline plausibility: Publication venue (ACM MM 2025), hardware (NVIDIA RTX 4090), and cited 2024 works (CARZero, CVPR 2024; Med-UniC, NeurIPS 2024) are all temporally coherent as of the review date.
Image-level analysis inconclusive: Figure 1–Figure 6 could not be inspected at the pixel level because figure images were not available in the submitted text — a review-environment limitation, not a paper defect.
Minor typo detected: The word "abesent" appears in Section 3.1 (intended: "absent"). Interpreted as a human writing artifact; recommended for camera-ready proofreading.

Evidence highlights

DOI: 10.1145/3746027.3755336
Authors: Lihong Qiao, Shiyi Gao, Yucheng Shu, Bin Xiao, Weisheng Li, Xinbo Gao
Venue: Proceedings of the 33rd ACM International Conference on Multimedia (MM '25), 2025
Exact reproducing figures: 80.8 / 82.3 / 66.9 / 21.7 across Tables 1, 4, 5, 6 (same configuration)
CheXpert self-reported result: 89.5 (1% and 10% settings, Table 1)
Competitor disclosure quote: underperformance on CheXpert attributed to Med-UniC's use of PadChest
Typo location: Section 3.1, "abesent"

Notes

Image manipulation and duplicate-image checks were not feasible; pixel-level reuse or splicing could not be ruled in or out.
"Identical numbers across ablation tables" is presented here as a positive consistency signal; it is not, on its own, statistical proof of authenticity, but it removes a common fabrication artifact (mismatched self-contradictory numbers).
Confidence in the clearance verdict is moderate. A definitive judgment would require access to the original figures, training logs, and raw experimental outputs.
No action items beyond a routine spell-check (e.g., "abesent" → "absent") are warranted.