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

Academic fraud report · Geng Detector

Summary

This integrity review examines the ACM MM '25 paper (DOI: 10.1145/3746027.3755336) by Lihong Qiao et al. The overall verdict is EXONERATED (no fraud indicators found). Key positive signals include: (1) the authors honestly report underperforming a baseline (Med-UniC at AUC 90.8% vs. their 89.7% on CheXpert 100%) and provide a plausible methodological explanation (Med-UniC uses the additional PadChest pretraining corpus); (2) the data-saturation pattern across 1%/10%/100% training ratios (AUC 89.5/89.5/89.7) is consistent with linear probing on robust pretrained features rather than artificially generated monotonic improvements; (3) cited references (NeurIPS 2024 Med-UniC, RadGraph, UMLS, ViT-B/16, Sinkhorn-Knopp) are real and the publication timeline is coherent. Image-level forensics (pixel manipulation, copy-paste) were NOT performed because raw high-resolution figures were unavailable. Confidence is moderate for text/data findings and low for image findings. No fabricated reagents, datasets, or methods were detected.

Verdict

EXONERATED — No evidence of academic fraud. The report concludes the paper is likely authentic based on data-honesty, methodological coherence, and timeline checks. Image-level forensics were not performed due to lack of raw figure files.

Key findings

Honest reporting of sub-SOTA results: On CheXpert (100% data), the proposed model achieves AUC 89.7%, below Med-UniC's 90.8%. The authors explicitly acknowledge this in Section 5.1 and attribute it to Med-UniC's additional pretraining data (PadChest). Such transparency is atypical of fabricated studies, which usually claim across-the-board superiority.
Plausible data-saturation pattern: On CheXpert, AUC values at 1%/10%/100% training ratios are 89.5/89.5/89.7, mirroring the baseline MRM (88.5/88.5/88.7). This flat-then-slight-rise behavior is consistent with linear probing saturating quickly on robust pretrained features, not with fabricated monotonic gains.
Coherent citation and tool timeline: References include NeurIPS 2024 work (Med-UniC [34]), RadGraph [16], UMLS medical knowledge graph, ViT-B/16, and Sinkhorn-Knopp clustering — all real and widely used. Timeline fits a manuscript targeting ACM MM 2025.
Figure pixel analysis not performed: Figures 1–6 were unavailable for pixel-level inspection (no high-resolution images provided). Figure 5/6 descriptions (t-SNE, zero-shot attention heatmaps) appear logically arranged.

Evidence highlights

Table 1, Page 5: CheXpert AUC 89.7% (proposed) vs. 90.8% (Med-UniC), 89.2% (MRM), with an explicit justification in Section 5.1 about PadChest pretraining.
Table 1, Page 5 (ViT-based CXP column): 1%→10%→100% AUC = 89.5/89.5/89.7; MRM baseline 88.5/88.5/88.7 — consistent saturation behavior.
References [16] (RadGraph) and [34] (Med-UniC, NeurIPS 2024) — verifiable peer-reviewed sources.

Notes

DOI: 10.1145/3746027.3755336
Conference: 33rd ACM International Conference on Multimedia (MM '25), 27–31 October 2025.
Detection scope was limited to text and tabular data supplied in the source PDF; raw figures were not provided, so image reuse/PS-trace checks could not be executed. Recommend requesting high-resolution figures from the authors' GitHub if image verification is later required.
Confidence: HIGH for text/data analysis; LOW for image-related integrity.
Disclaimer: This is an AI-assisted preliminary assessment, not an institutional finding.