Investigation 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

This report assesses a 2025 ACM MM '25 paper by Lihong Qiao et al. on chest X-ray vision-language pre-training, reaching a verdict of "highly suspicious." Three core concerns are documented. First, Table 1 reports an identical AUC of 89.5 on CheXpert for the proposed method at both 1% and 10% training data, with suspicious mirror-style last-digit symmetry (e.g., COVIDx 1% ACC 82.3 vs. 10% ACC 92.3), contradicting expected variation across baselines. Second, the claimed pre-training setup—two NVIDIA RTX 4090 GPUs running 200 epochs over the MIMIC-CXR corpus (~377k images, ~227k reports) at 448×448 resolution with ViT-B/16 and cross-attention—is implausible given typical 8× A100 (80GB) hardware for similar work. Third, ablation gains across Table 4 are reported without standard deviations, undermining statistical interpretability. Image-based forensics could not be performed. Confidence is moderate-to-high for findings 1–3; image reuse (specifically Figure 5 t-SNE) remains unverified pending access to raw vectors.

Verdict

Highly suspicious. Three independent anomalies—data implausibility, computational infeasibility, and missing statistical reporting—co-occur in a way that substantially weakens confidence in the reported results. The cumulative pattern is consistent with fabrication, copy-paste errors, or exaggerated experimental scope. However, no single finding is individually conclusive, and image forensics were not completed.

Key findings

Implausible training-data scaling in Table 1: The proposed method (Ours) is reported to achieve AUC = 89.5 on CheXpert at both 1% and 10% training data, identical to one decimal place. All baselines (MAE, MGCA, GLoRIA, MRM) show clear variation between these data fractions, violating the expectation of monotonic, non-zero change.
Suspicious numeric symmetry: On COVIDx, ACC is reported as 82.3 at 1% data and 92.3 at 10% data—an unusually tidy mirror on the last digit, raising concerns about fabricated or hand-tuned results.
Questionable compute claims: Section 4.3 states that full pre-training on MIMIC-CXR (~377,000 images, ~227,000 reports, 448×448 resolution, ViT-B/16, 112-token text, cross-attention, 200 epochs) was performed on two NVIDIA RTX 4090 GPUs (24 GB each). Comparable published setups typically use 8× A100 (80 GB) or equivalent. The 4090 claim is asserted as implausible.
Ablation gains without variance: Table 4 reports point estimates for ablation steps (e.g., ChestX-ray14 baseline 79.1 → +DKBA 81.0; COVIDx 77.0 → 81.9) without any standard deviations or seed counts, making the reported 0.1–0.5 increments uninterpretable statistically.

Evidence highlights

Table 1 (p. 5): CheXpert "Ours" AUC = 89.5 at both 1% and 10% training data; COVIDx "Ours" ACC = 82.3 (1%) vs. 92.3 (10%).
Section 4.3 Implementation Details (p. 6): two RTX 4090 GPUs, 200 reconstruction-pre-training epochs, 448×448 inputs, 112-token text, MIMIC-CXR corpus.
Table 4 and Section 5.3.1 (p. 6): ablation gains reported as bare means without ±std; no variance reported across the paper.
DOI: 10.1145/3746027.3755336 (ACM MM '25).

Notes

Image-based forensics (Figure 1–6, including the t-SNE plot in Figure 5) could not be performed at the pixel level; reuse or post-hoc manipulation of visualizations cannot be ruled out or confirmed.
The exact AUC duplicate (89.5 at 1% and 10%) is the strongest independent indicator of either fabrication or a copy-paste error and should be the priority for author clarification.
Recommended actions: request raw training logs, a std-augmented supplemental table, and clarification on hardware/time investment for the 200-epoch MIMIC-CXR run. Concerns may be raised with the ACM MM '25 program committee and the authors' institution.
This report is AI-assisted and intended for academic discussion only; final determination of misconduct requires institutional investigation.