Geng Report: Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training

Academic fraud report · Geng Detector

Summary

A peer-style integrity review of the manuscript 'Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training' (Qiao et al., submitted to ACM MM '25; DOI: 10.1145/3746027.3755336). The report flags a highly suspicious rating. The principal issue is an internal numerical contradiction: the main table (Table 1) reports a final-model AUC of 81.4 on ChestX-ray14 with 1% training data, while the ablation tables (Tables 4 and 5) for the same full model in the same setting report 80.8, a 0.6-point unexplained gap. A second concern is a verbatim duplication: MRM is reported as achieving an AUC of 88.5 on CheXpert under both 1% and 10% training data, a coincidence that strains plausibility. Third, citation mislabeling is documented (Med-UniC attributed to ref [24] which is actually SAT; 'ReCO' description attached to ref [12] which is MAE). Fourth, a tentative flag is raised over the feasibility of pretraining on MIMIC-CXR (~370k images, 200 epochs) on only two RTX 4090 GPUs, though this is noted as insufficient on its own. The review is based on text only; no image-duplication analysis was performed.

Verdict

Highly suspicious. Multiple internal inconsistencies and citation errors undermine the credibility of the reported results. The central finding is a direct numerical contradiction between the main results table and the ablation tables for the same model under the same setting.

Key findings

Internal data contradiction (main table vs. ablation). Table 1 reports the full model at 81.4 AUC on ChestX-ray14 (1% data), while Tables 4 and 5 report the same full model at 80.8 AUC under the same 1% CXR14 setting, an unexplained 0.6-point gap.
Implausibly identical baseline scores. MRM is reported with an AUC of 88.5 on CheXpert under both 1% and 10% training data, an exact match to one decimal place despite a 10× change in training data.
Mismatched citations. Med-UniC is attributed to reference [24] (which is SAT), and the description of "ReCO" is attached to reference [12] (which is MAE).
Questionable compute budget. A MIMIC-CXR pretraining run (~370k images, ~220k reports) with 200 reconstruction epochs is described as performed on only 2× NVIDIA RTX 4090 GPUs, which is atypical for this scale of multimodal pretraining.

Evidence highlights

Table 1 (Ours, 1% CXR14) AUC = 81.4
Table 4 (PAR + DKBA, 1% CXR14) AUC = 80.8
Table 5 (MGCA prior, CXR14) AUC = 80.8
Table 1, MRM [46] on CheXpert: 88.5 (1%), 88.5 (10%), 88.7 (100%)
Reference [24] cited as Med-UniC; reference [24] is SAT
Reference [12] cited as ReCO; reference [12] is MAE
Section 4.3 states hardware: 2× NVIDIA RTX 4090; reconstruction stage 200 epochs; alignment stage 15 epochs; dataset: MIMIC-CXR
DOI: 10.1145/3746027.3755336

Notes

Image-based duplication and splicing checks (Geng detection modalities 1 and 3) were not performed because the source PDFs/figures were not provided.
The compute concern is flagged as anomalous but not, on its own, evidence of misconduct; long wall-clock training on consumer hardware is not strictly impossible.
All numerical values above are reproduced exactly as cited; no additional figures have been inferred.
Authors and venue should be given the opportunity to provide raw logs, code, and training records to resolve the main-vs-ablation discrepancy.