Integrity Review Report: 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: Suspected issues / inconclusive (yellow rating). This 2025 ACM MM '25 paper by Qiao, Gao, Shu, Xiao, Li, and Gao proposes a chest X-ray vision-language pretraining framework combining masked reconstruction with knowledge-boosted alignment. The review analyzed Figures 1–6, Tables 1–4 and 6, and the Implementation Details section (4.3). Key issues: (1) Image forensics could not be performed because only plain-text extraction was available, so potential reuse or splicing in architecture diagrams, t-SNE plots, and heatmaps remains unverified and is flagged for re-examination with pixel-level material. (2) Statistical logic was internally consistent: the authors honestly report underperformance versus Med-UniC on CheXpert (1% 89.4, 10% 89.7, 100% 90.8 vs their 89.5/89.5/89.7), and a 'reverse' ablation in Table 4 shows AUC dropping from 79.1 to 78.9 on ChestX-ray14 after adding PAR, indicating non-cooked results. (3) Computational setup (2× RTX 4090, 200 reconstruction epochs, 15 alignment epochs on MIMIC-CXR) is plausible. Limitations: confidence is moderate due to the absence of source images; conclusions should be treated as preliminary pending verification of the original PDF figures and open-source code reproducibility.

Verdict

🟡 Suspected / Inconclusive — No direct evidence of academic fraud was found in textual and tabular content, but image-level forensics could not be completed. Overall assessment leans toward integrity, with an explicit caveat that the rating is constrained by the lack of pixel-level evidence.

Key findings

Honest baseline comparison (Table 1): The authors transparently report that Med-UniC outperforms their model on CheXpert (1%: 89.4 vs 89.5; 10%: 89.7 vs 89.5; 100%: 90.8 vs 89.7), and acknowledge underperformance in Section 5.1 — atypical behavior for fabricated results.
'Reverse' ablation result (Table 4): Adding the PAR module decreases ChestX-ray14 AUC from 79.1 to 78.9, suggesting non-manipulated experimental runs.
No fabricated monotonic improvements: Tables 1–3 and Table 6 show realistic, modest gains (1–2%) consistent with current SOTA ranges in medical image pretraining.
Plausible computational footprint: Section 4.3 specifies 2× NVIDIA RTX 4090 GPUs, 200 epochs for reconstruction pretraining and 15 epochs for alignment pretraining on MIMIC-CXR, which matches realistic hardware–time expectations.
No methodological anomalies detected: Cited prior methods (MAE, MRM, MGCA) are genuine and recognized works in the field.
Image forensics pending: Figures 1–6 (architecture diagram, t-SNE visualizations, heatmaps) could not be examined at pixel level; ELA, noise-pattern, or splicing analysis requires the original PDF.

Evidence highlights

Table 1 — CheXpert comparison: Med-UniC 1% 89.4 / 10% 89.7 / 100% 90.8 vs proposed method 1% 89.5 / 10% 89.5 / 100% 89.7. Authors openly concede underperformance.
Table 4 — Ablation on ChestX-ray14: AUC 79.1 (baseline) → 78.9 (with PAR), a 0.2-point drop rather than an artificial gain.
Section 4.3 Implementation Details: "completed on two NVIDIA RTX 4090 GPUs, with 200 epochs for reconstruction pretraining and 15 epochs for alignment pretraining."
DOI: https://doi.org/10.1145/3746027.3755336
Code repository referenced for reproducibility: https://github.com/Felix1118/PADKB

Notes

Confidence is moderate, not high: the report is constrained by the absence of pixel-level image data, so the 'suspected' rating should be revisited once the original PDF figures or high-resolution images are available.
The authors' willingness to report negative results and to acknowledge a stronger baseline is a soft indicator of integrity, but it is not a guarantee; fabrication methods can be subtle.
Recommended follow-up: (i) re-run the open-source code (Felix1118/PADKB) to verify numerical claims, (ii) perform pixel-level forensic checks (ELA, copy-move, noise consistency) on Figures 1–6 from the original PDF, (iii) cross-check Figures and Tables against the camera-ready version on ACM DL.
No p-value or variance analysis was warranted, as the paper reports absolute metrics (AUC, mAP, Dice) typical of computer-vision evaluation.
This report is AI-assisted, intended for academic discussion only; final adjudication requires an institutional or editorial investigation.