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

Academic fraud report · Geng Detector

Summary

Verdict: Highly suspicious. This review identifies four concerns in the paper 'Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training' (DOI: 10.1145/3746027.3755336), accepted at ACM Multimedia 2025. First, Table 1 reports an identical CheXpert ViT-based AUC of 89.5 for both the 1% and 10% training subsets, an implausible coincidence suggesting fabrication or copy-paste error. Second, citation labels in Section 2.1 and Section 5.2 are misattributed: reference [24] (SAT) is described as Med-UniC, while [34] is the actual Med-UniC; reference [12] (MAE) is called MGCA, which corresponds to [35]. Third, the reported compute footprint (two NVIDIA RTX 4090 GPUs running 200 epochs of multimodal reconstruction pretraining on 377,110 MIMIC-CXR image-report pairs) appears implausibly tight for the stated scale. Fourth, the reported gains on COVIDx (82.3 / 92.3 / 96.0 for 1/10/100% data) far exceed prior SOTA, raising concerns about data tuning. Confidence is moderate-to-high for the tabular anomaly and citation errors, which are directly verifiable; the hardware and COVIDx concerns are suggestive but require original logs and code for confirmation. No image manipulation analysis was possible because figures were not provided.

Verdict

Highly suspicious. The combination of an identical AUC across two training-set fractions, multiple citation misattributions, and implausible compute claims warrants formal investigation by the conference editors.

Key findings

Identical CheXpert AUC (89.5) reported for both 1% and 10% training subsets in Table 1, contrary to normal learning-curve behaviour.
Systematic citation misattribution: Med-UniC incorrectly cited as [24] (which is SAT) instead of [34]; MGCA incorrectly cited as [12] (which is MAE) instead of [35].
Implausible compute footprint for the scale of the experiments (two RTX 4090 GPUs, 200 reconstruction pretraining epochs on 377,110 MIMIC-CXR image-report pairs).
Suspiciously large, monotonic gains on COVIDx (82.3 / 92.3 / 96.0 for 1/10/100% data), substantially exceeding prior SOTA without architectural or data-scale justification.

Evidence highlights

Table 1 (p. 5), CheXpert CXP column: 1% = 89.5, 10% = 89.5 for the authors' ViT-based method; comparable methods such as Med-UniC move from 89.4 to 89.7.
Section 2.1 (p. 2): sentence 'MedUnic [24] integrates multimodal medical data...' misidentifies [24]; reference [34] corresponds to Med-UniC.
Section 5.2 (p. 7): claim that 'the reconstruction-based MRM [46] outperforms MGCA [12]' mislabels MGCA; [12] corresponds to MAE, and MGCA is [35].
Section 4.3 (p. 6), Implementation Details: 377,110 image-report pairs from MIMIC-CXR, ViT-B/16 + BERT, 200 epochs reconstruction pretraining at 448x448 / 224x224, 15 epochs alignment pretraining, on 2x NVIDIA RTX 4090.
Table 1 (p. 5), COVIDx column: 82.3 (1%), 92.3 (10%), 96.0 (100%); versus MRM at 75.6 / 88.3 / 92.2 and Med-UniC at 80.3 / 89.5 / 94.5.

Notes

DOI: 10.1145/3746027.3755336
Venue: ACM Multimedia (MM '25), conference dates October 27–31, 2025; source PDF: acmmm-submission.pdf
No image-manipulation analysis was conducted because figures were not provided in the source material.
The Table 1 anomaly and citation errors are directly verifiable from the supplied PDF; the compute and COVIDx findings are suggestive and would be best confirmed by requesting raw training logs, random seeds, and the official code release from the authors.