Academic Integrity Report: Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for CheX-ray Vision-Language Pre-training

Academic fraud report · Geng Detector

Summary

This report assesses the paper "Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training" by Lihong Qiao, Shiyi Gao, Yucheng Shu, Bin Xiao, Weisheng Li, Xinbo Gao, published in Proceedings of the 33rd ACM International Conference on Multimedia (MM '25), DOI: 10.1145/3746027.3755336. The overall verdict is that the paper appears CLEAN based on the available evidence. A systematic review of the data logic and consistency (Tables 1, 2, 3, 4, 6) found that comparative results with competitor methods align with accepted benchmarks in medical vision-language pre-training, and ablation results show realistic trade-offs rather than suspiciously uniform improvements. Methodology and timeline checks confirmed the plausibility of the reported training setup (two NVIDIA RTX 4090 GPUs, ViT-B/16 cascade pre-training with 200 reconstruction epochs plus 15 alignment epochs). Pixel-level image analysis could not be performed because only text-extracted content was available. Limitations include the inability to verify Figures 1-6 for visual manipulation, and reliance on text-based evidence only.

Verdict

✅ Clean (within the limits of text-only analysis). No evidence of data fabrication, methodological inconsistency, or timeline anomalies was detected.

Key findings

Quantitative results across Tables 1, 2, 3, 4, and 6 show logical consistency and align with accepted benchmarks in medical vision-language pre-training; no signs of cherry-picked or suspiciously inflated numbers.
Ablation study (Table 4) exhibits realistic trade-offs: using PAR alone slightly reduces classification accuracy on ChestX-ray14 (79.1 → 78.9) while improving detection and segmentation tasks; DKBA shows the opposite pattern, consistent with genuine model behavior rather than artificial inflation.
Methodology section (Section 3, Equations 1–13) is internally consistent, with uniform notation and no reference to non-existent architectures.
Reported training setup (two NVIDIA RTX 4090 GPUs, ViT-B/16, 200 reconstruction epochs + 15 alignment epochs) is plausible given the current hardware landscape.
Pixel-level image forensics could not be performed because only extracted text was provided for review.

Evidence highlights

Table 4 (Ablation): PAR-only classification 79.1 → 78.9 on ChestX-ray14, but gains in detection/segmentation; DKBA shows the inverse pattern — a realistic trade-off profile.
Implementation Details (Section 4): Two NVIDIA RTX 4090 GPUs; 200 reconstruction epochs plus 15 alignment epochs for the ViT-B/16-based cascade pre-training.
DOI: 10.1145/3746027.3755336
Figures 1–6: No pixel-level data available; only caption-level descriptions reviewed (e.g., Figure 5 t-SNE visualization).

Notes

The detector explicitly notes that pixel-level inspection of Figures 1–6 was not possible due to the absence of high-resolution source images. Future verification with image-analysis tools is recommended if the original figures become available.
This assessment is based solely on text-extracted content; final determination of academic misconduct requires investigation by qualified institutions.
The original report's commentary about the absence of "Western Blot" is a colloquial remark and should not be interpreted as a substantive finding.