Integrity Review Report: Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training

Academic fraud report · Geng Detector

Summary

This report examines the paper 'Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training' by Lihong Qiao et al., published at ACM MM '25 (DOI: 10.1145/3746027.3755336). Overall verdict: the paper appears CLEAN / no clear fraud indicators detected. Statistical analysis of Tables 1-3 shows natural digit distributions and methodologically consistent improvement patterns; the proposed method outperforms baselines strongly on classification but more modestly on detection/segmentation, which matches the algorithmic design. Crucially, the Table 4 ablation on ChestX-ray14 reports a slight AUC drop from 79.1 to 78.9 when adding the PAR module alone, with an honest textual explanation, which is a strong authenticity marker against systematic data fabrication. Hardware/citation timelines (RTX 4090, NeurIPS 2024 Med-UniC, CVPR 2024 CARZero) are internally consistent for a 2025 publication. Pixel-level image reuse/duplication analysis (Figures 1-6) was NOT performed due to absence of raw image data, so a residual uncertainty remains on this axis only. Confidence in the textual/statistical verdict is high; image-level confidence is limited.

Verdict

Clean (✅). No evidence of academic fraud was identified. Pixel-level image analysis could not be performed due to missing source images, leaving a minor residual uncertainty on that dimension only.

Key findings

Statistical tables (Table 1, 2, 3): Trailing-digit distribution is natural (0–9 appear without artificial rounding), and the magnitude of gains follows methodological logic — large gains on classification, modest gains on detection/segmentation vs. MRM. No "uniformly dominant" pattern typical of fabrication.
Ablation honesty (Table 4, Page 7): Adding the PAR module alone causes AUC on ChestX-ray14 to drop from 79.1 to 78.9. The paper reports this honestly and explains it in text (PAR focuses on fine-grained tasks and may slightly hurt global classification). Reporting a small regression is a strong authenticity marker that argues against systematic data invention.
Timeline consistency: Hardware (2× NVIDIA RTX 4090, released 2022) is appropriate for a 2025 paper; cited references Med-UniC (NeurIPS 2024) and CARZero (CVPR 2024) are plausibly available for a 2025 submission.
Mathematical self-consistency: Equations 3 (weighted MSE), 5 (contrastive loss), and 6–8 (GAT-style attention) follow standard formulations with no internal contradictions.
Image-level checks (Figure 1–6): Not performed — only text was available, so Western blot, microscopy, or feature-map duplication/tampering cannot be evaluated.

Evidence highlights

DOI: 10.1145/3746027.3755336
Venue: Proceedings of the 33rd ACM International Conference on Multimedia (MM '25), 2025
Table 4 ablation: PAR-only variant → 78.9 vs. baseline 79.1 on ChestX-ray14 AUC (honestly reported decline).
Table 1 shows strong gains over baselines on classification; Tables 2–3 show narrower gains on fine-grained detection/segmentation vs. MRM, consistent with design.
Hardware: 2× NVIDIA RTX 4090; cited baselines include Med-UniC (NeurIPS 2024) and CARZero (CVPR 2024).
Minor cosmetic note: title writes "Che-Xray" (likely intended "Chest-Xray"); not an integrity issue.

Notes

The image duplication / splicing check (Figures 1–6, including the feature-visualization panels in Figures 5–6) was limited by lack of original image data; an independent image-forensics pass is recommended if the figures become critical to a specific claim.
Authors may wish to correct the "Che-Xray" → "Chest-Xray" typographical error in the title.
An optional code-replication check using the authors' public repository would further corroborate the reported numbers.