Investigative report on 'Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training' (Qiao et al., ACM MM '25)

Academic fraud report · Geng Detector

Summary

Verdict: highly suspicious. The report raises four concerns regarding the paper, which introduces a vision-language pretraining method for chest X-rays. (1) Compute plausibility: the authors claim to have pretrained a 12-layer ViT-B/16 with a 4-layer decoder on 377,110 chest radiographs resized to 448x448 for 200 epochs, including GAT-based knowledge graph extraction and multimodal alignment, using only two NVIDIA RTX 4090 GPUs (24 GB each). The reviewer notes this is implausible for that scale relative to typical practice (8+ A100/V100 GPUs over weeks). (2) Suspiciously uniform SOTA gains over Med-UniC on RSNA of 0.1-0.2 AUC across 1%/10%/100% data regimes. (3) Identical baseline AUC values for MRM on CheXpert (88.5 at both 1% and 10% data). (4) The proposed model reaches 89.5 AUC on CheXpert with only 1% data versus 89.7 at 100%, an implausibly flat learning curve. Confidence is moderate; findings rely on authors' own numbers and standard hardware norms, and no independent reproduction has been performed.

Verdict

🟠 Highly suspicious. Multiple independent irregularities in compute claims, benchmark numbers, and performance curves warrant investigation. None is individually conclusive, but their combination is consistent with fabricated or massaged experimental results.

Key findings

Implausible compute claim: 200 epochs of reconstruction pretraining on 377,110 images at 448x448 with a 12-layer ViT-B/16 + 4-layer decoder, plus GAT and multimodal alignment, reportedly performed on only two NVIDIA RTX 4090 GPUs. Standard practice for comparable scale uses 8+ A100 80GB / V100 32GB GPUs over weeks.
Suspiciously uniform SOTA margins on RSNA: The proposed model outperforms Med-UniC by exactly 0.1-0.2 AUC at every data fraction (1%/10%/100%), yielding 92.1/93.2/93.8 vs. 91.9/93.1/93.7. Such regularity across regimes is atypical.
Identical baseline AUC for MRM on CheXpert: MRM is reported at 88.5 AUC for both 1% and 10% training data, an implausible zero-delta across a 10x data increase.
Deformed learning curve on CheXpert: The proposed model jumps from 89.5 (1%) to 89.7 (100%) AUC, suggesting that 1% of data already saturates 99.7%+ of full-data performance, which is atypical for multi-label chest X-ray classification.

Evidence highlights

Dataset size: 377,110 chest radiographs; image resolution: 448x448; pretraining epochs: 200 (reconstruction); hardware: two NVIDIA RTX 4090 GPUs (Section 4.1, Section 4.3).
Table 1, RSNA, ViT-based linear classification (Ours vs. Med-UniC): 1% 92.1 vs. 91.9; 10% 93.2 vs. 93.1; 100% 93.8 vs. 93.7.
Table 1, CheXpert, ViT-based linear classification, MRM: 88.5 AUC at 1% and 88.5 AUC at 10%.
Table 1, CheXpert, Ours: 89.5 (1%) / 89.5 (10%) / 89.7 (100%) AUC.
DOI: 10.1145/3746027.3755336

Notes

All cited numbers are transcribed directly from the paper as reported by the originating Chinese-language review; no independent verification of Table 1 or compute logs has been performed.
The hardware-plausibility finding is based on common-practice estimates rather than a formal FLOPs/throughput calculation; alternative configurations (mixed precision, gradient checkpointing, sharding across datasets) could partly mitigate the concern, though the reviewer judges full reproduction on two RTX 4090s within a reasonable timeframe to be unlikely.
The suspect MRM values may reflect a copy-paste error rather than deliberate fabrication; however, the reviewer flags this as a low-effort data integrity issue regardless of intent.
The recommendation to request pre-training logs, wall-clock times on the two RTX 4090s, and the MRM CheXpert test scripts is reasonable but should be directed through appropriate channels (authors, program chairs, or institutional research integrity offices).
No evidence of image manipulation or plagiarism is alleged in this report; concerns are limited to hardware feasibility and benchmark-result plausibility.
This report is AI-assisted and intended for academic discussion only; final determination of misconduct requires investigation by qualified institutions.