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: Cleared (✅ 清白). This review examined the ACM MM '25 paper by Lihong Qiao et al. for image reuse/splicing, data fabrication, hardware/timeline inconsistencies, and methodological or citation anomalies. Cross-tabular consistency checks across Tables 1–6 (including the full-model ablation row in Table 4) showed no contradictions with detailed 1% data splits in Tables 1–3; reported gains (AUC +1–2%, mAP +1–2%) fall within typical ranges for medical vision-language pre-training. Hardware claims (two NVIDIA RTX 4090 GPUs, 200 reconstruction epochs, 15 alignment epochs on MIMIC-CXR with ViT-B/16) are physically plausible. Figures (architecture diagrams, training curves, t-SNE, zero-shot visualisations) could not be rigorously forensically analysed from text alone, but no descriptive or logical inconsistencies were detected. No self-contradictions, dimensional errors, or fabricated citations were found. Limitations: image-level pixel forensics not performed; final determination requires institutional review.

Verdict

Cleared (✅ 清白). No substantive evidence of academic fraud was identified across image reuse, data fabrication, hardware/timeline consistency, or methodological/citation checks. Image-level pixel forensics were not feasible from text-only artefacts; all other checks passed.

Key findings

Cross-tabular consistency (Tables 1–6): The full-model ablation entries in Table 4 (1% setting: ChestX-ray14 AUC 80.8, COVIDx ACC 82.3, SIIM Dice 66.9, RSNA mAP 21.7) are internally consistent with the "Ours" rows in Tables 1–3 at the corresponding 1% data split. No sign of tampering or round-number fabrication.
Plausible improvement margins: Gains reported are in the order of AUC +1–2% and mAP +1–2%, consistent with typical progress in medical vision-language pre-training; no implausibly large leaps observed.
Hardware/timeline realism: Implementation Details (Section 4.3) specify two NVIDIA RTX 4090 GPUs, 200 reconstruction epochs and 15 alignment epochs on MIMIC-CXR (~377k images) with a ViT-B/16 backbone — physically plausible.
Methodology and citations: References to MGCA [35], MRM [46], MedKLIP [40] and others correspond to real, topically appropriate prior work. Mathematical formulations of PAR and DKBA (contrastive loss, cross-entropy, GAT computations) appear consistent, with no matrix-dimension or self-contradictory errors observed in the text.
Figures (1–6): Limited to textual verification due to absence of raw pixel data. Figures consist of architecture diagrams (Fig. 1–3), training curves (Fig. 4), and t-SNE / zero-shot visualisations (Fig. 5–6). No descriptive inconsistencies or logical conflicts were detected in captions or surrounding text.

Evidence highlights

DOI: 10.1145/3746027.3755336
Venue: Proceedings of the 33rd ACM International Conference on Multimedia (MM '25), October 27–31, 2025
Authors: Lihong Qiao, Shiyi Gao, Yucheng Shu, Bin Xiao, Weisheng Li, Xinbo Gao
Table 4 (1%) full-model values: ChestX-ray14 AUC 80.8, COVIDx ACC 82.3, SIIM Dice 66.9, RSNA mAP 21.7 — matches Tables 1–3 "Ours" entries.
Section 4.3 hardware: two NVIDIA RTX 4090 GPUs; 200 epochs (reconstruction) + 15 epochs (alignment) on MIMIC-CXR (~377k images), ViT-B/16.

Notes

Image-level splicing / re-use forensics (lane-background or medical-imaging noise analysis) were not performed; only textual and logical consistency was assessed.
No recommendation for contacting the authors, posting on PubPeer, or reporting to the conference organisers.
Follow-up action flagged: peer readers may reproduce the work using the publicly referenced GitHub code.
Disclaimer: This report is AI-assisted, intended for academic discussion only. Final determinations of academic misconduct require formal institutional investigation. False positives and false negatives are possible.