Forensic Review Report: Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training (DOI: 10.1145/3746027.3755336)

Academic fraud report · Geng Detector

Summary

Verdict: Highly suspicious (orange rating). The report, titled 'Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training' (Lihong Qiao et al., submitted to ACM MM '25, DOI: 10.1145/3746027.3755336), flags three main concerns. First, the authors claim that pretraining on the MIMIC-CXR dataset (over 370,000 images) at 448x448 resolution using a ViT-B/16 backbone, transformer text encoder, 4-layer ViT decoder, GAT modules, and multimodal cross-attention was completed on only two NVIDIA RTX 4090 GPUs (24GB each). Reviewers note this compute configuration is physically implausible for large-batch contrastive learning at this scale, with high probability of OOM. Second, confirmed citation mislabeling was found: Med-UniC is cited as reference [24] (which is actually SAT, IEEE TMI 2023), and MGCA is cited as [12] (which is actually MAE, CVPR 2022), indicating copy-paste artifacts. Third, textual truncation ('t. we compute the cosine similarities...') and numerical concatenation ('89.4 89.7 90.891.993.193.780.389.594.5') in Table 1 suggest rushed, uncorrected submission. The hardware claim is marked insufficiently substantiated; citation and formatting issues are confirmed. Final determination requires institutional investigation.

Verdict

🟠 Highly suspicious. Multiple methodological, editorial, and reproducibility concerns were identified. Citation and formatting issues are directly verified from the manuscript; the hardware feasibility claim is plausible but flagged as insufficiently substantiated pending the authors' release of training logs and batch-size configuration.

Key findings

Compute configuration anomaly (Section 4.3): Authors state the entire pretraining on MIMIC-CXR (>370,000 images, 448×448 resolution) using ViT-B/16, a transformer text encoder, a 4-layer ViT decoder, two GAT layers, and multimodal cross-attention was completed on two NVIDIA RTX 4090 GPUs (24 GB each).
Citation mislabeling (Sections 2.1 and 5.2): Med-UniC is cited as [24], but [24] is "Bo Liu et al., SAT, IEEE TMI 2023". MGCA is cited as [12], but [12] is "He et al., MAE, CVPR 2022". Misnumbered citations suggest copy-paste from unrelated drafts.
Text truncation and data concatenation (Table 1 area): The paragraph following Table 1 begins with a fragmentary "t. we compute the cosine similarities...", indicating lost preceding text. Table 1 row for Med-UniC reads "89.4 89.7 90.891.993.193.780.389.594.5" with missing spacing between metrics.

Evidence highlights

Direct quote from Section 4.3: *"The entire pretraining process is completed on two NVIDIA RTX 4090 GPUs."*
MIMIC-CXR scale: >370,000 chest X-ray images at 448×448 input resolution with contrastive learning objective.
Confirmed mismatches:
[24] = Bo Liu et al., SAT, IEEE TMI 2023 (not Med-UniC)
[12] = He et al., MAE, CVPR 2022 (not MGCA)
Verbatim textual fragment: *"t. we compute the cosine similarities..."*
Verbatim numerical concatenation: *"89.4 89.7 90.891.993.193.780.389.594.5"*

Notes

The hardware feasibility concern (Finding 1) is marked ⚠️ insufficiently substantiated; reviewers recommend requesting full training logs, per-GPU batch size, gradient accumulation settings, and any OOM records to verify reproducibility.
The citation and formatting issues (Findings 2 and 3) are confirmed by direct comparison against the manuscript's reference list and tables.
No evidence of image manipulation or data fabrication is presented in this report; concerns are limited to compute feasibility, citation integrity, and editorial quality.
Final adjudication of academic misconduct requires institutional investigation by ACM MM '25 program committee or the authors' institution.
DOI: 10.1145/3746027.3755336