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

Academic fraud report · Geng Detector

Summary

Verdict: Cleared (no evidence of academic fraud identified within the scope of this review). The paper, published in ACM MM '25, was assessed across three dimensions: statistical consistency of reported metrics, logical coherence of ablation studies, and methodological/citation plausibility. The terminal-digit distribution of performance scores (AUC, ACC, Dice, mAP) across Tables 1–3 appears natural, and performance gains across 1%, 10%, and 100% data scales follow a plausible non-linear curve consistent with diminishing returns at higher data volumes. Ablation results in Tables 4 and 6 reveal realistic trade-offs, such as gains on fine-grained segmentation tasks offset by minor fluctuations on classification tasks, which is more consistent with genuine experimentation than fabricated uniform improvements. Cited baselines (e.g., NeurIPS 2024 Med-UniC) align with normal publication timelines. The only limitation is that no image data was available for pixel-level duplication or splicing analysis; this is noted but not indicative of fraud. Confidence is high for text-based checks; image integrity remains unverified.

Verdict

Cleared. Based on the text, tabular data, and methodological description provided, no concrete indicators of academic fraud were identified. One dimension (image integrity) could not be evaluated due to lack of access to the original figures.

Key findings

Statistical consistency (Tables 1–3): Terminal-digit distribution of metrics (AUC, ACC, Dice, mAP) is uniform and natural. Performance curves across 1%, 10%, and 100% data ratios show expected non-linear, diminishing-returns behavior rather than suspiciously uniform deltas.
Ablation logic (Tables 4, 6): Genuine trade-offs observed. Adding the PAR (Pathology-Aware Reconstruction) module improves fine-grained tasks (SIIM Dice, RSNA mAP) while causing minor fluctuations on classification tasks, consistent with the stated trade-off between local detail reconstruction and global discriminative power. Table 6 supports the paper's opening hypothesis regarding gradient conflict between joint and cascaded training.
Citation and methodology timeline: Referenced models (e.g., Med-UniC, NeurIPS 2024) align with a 2025 publication. The methodological stack (MGCA-generated pathology priors, PubMedBERT, ViT-B/16, U-Net, YOLOv3) is internally consistent, with no anachronistic or fabricated components.

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
Tables 1–3: Metric distributions and scaling behavior consistent with natural experimental noise.
Table 4: PAR module yields task-specific trade-offs (fine-grained up, classification slightly down).
Table 6: Cascaded training outperforms joint training, supporting the gradient-conflict hypothesis.
No future-dated models or implausible hardware claims detected.

Notes

Limitation — image analysis not performed: Figures 1–6 were not accessible in text form, so pixel-level checks (duplication, splicing, noise inconsistencies, erasing artifacts) could not be conducted. This is a domain (CV/multimedia) where such checks are valuable, particularly on segmentation/visualization outputs. A follow-up review using the original high-resolution figures and any released code is recommended to fully close this gap.
Recommendation: If a public code repository is available, comparing the implementation against the paper's formulas may further strengthen or revise this verdict.
Disclaimer: This report is AI-assisted and intended for academic discussion only. Final determinations of misconduct require institutional investigation.