Integrity Review Report: "Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for CheXray Vision-Language Pre-training"

Academic fraud report · Geng Detector

Summary

This report assigns a verdict of Highly Suspicious to the ACM MM '25 paper (DOI: 10.1145/3746027.3755336) by Lihong Qiao et al. The analysis identifies multiple inconsistencies across ablation tables, reported metrics, mathematical definitions, and compute claims. In Table 4, the full model underperforms an individual submodule (81.0 → 80.8 AUC on ChestX-ray14; 20.2 → 19.9 mAP on RSNA), which violates standard ML expectations for cumulative architectural gains. In Table 1, the CheXpert ViT-based results report identical AUC values (89.5) at 1% and 10% training data, an implausible coincidence at one-decimal precision. Additionally, the variable Np is used inconsistently in Section 3.1 for high-resolution and low-resolution patches, suggesting copy-paste from related work (e.g., MAE) without reconciling the 4× patch-count difference. The 4090-GPU compute claim for 200-epoch pre-training on 370k+ MIMIC-CXR images is flagged as implausible. Findings are reported with explicit confidence levels; final determination requires institutional investigation.

Verdict

Highly suspicious. Multiple independent inconsistencies across experimental tables, mathematical formulation, and reported compute environment suggest probable data fabrication or incoherent experimental execution. The paper has not yet been formally published (expected October 2025), and no retraction or official response is on record at the time of this report.

Key findings

Ablation logic violation (Table 4): Adding both PAR and DKBA (full model) yields 80.8 AUC on ChestX-ray14, lower than DKBA alone at 81.0%. On RSNA detection, the baseline reaches 20.2 mAP while DKBA alone drops to 19.9 mAP—contradicting the principle that adding modules to a complete model should not degrade performance below a sub-configuration.
Implausible identical metrics (Table 1): CheXpert ViT-based AUC reported as 89.5 (1% data), 89.5 (10% data), and 89.7 (100% data). A 10× increase in training data producing an identical one-decimal AUC is statistically implausible. The COVIDx column shows a 10.0-point jump between 1% and 10% data, inconsistent with the CheXpert curve.
Mathematical inconsistency in Np (Section 3.1): The same symbol Np is applied to both high-resolution (448×448) and low-resolution (224×224) images. With equal patch size, the counts differ by a factor of 4 (e.g., 784 vs. 196). This suggests formulas were adapted from prior work (likely MAE) without proper rederivation.
Compute claim concern (Section 4.3): The paper reports using 2× NVIDIA RTX 4090 GPUs (~48 GB total) for 200 reconstruction-pretraining epochs plus 15 alignment-pretraining epochs on >370,000 MIMIC-CXR images under a ViT-B/16 backbone. Independent reviewers note this scale is difficult to reconcile with the stated hardware; while not impossible, the description lacks wall-clock estimates or memory optimization details.

Evidence highlights

Table 4 (ChestX-ray14): full-model AUC 80.8 < DKBA-only 81.0; (RSNA): DKBA-only mAP 19.9 < baseline 20.2.
Table 1 (CheXpert ViT-based): AUC = 89.5 at 1% data and 89.5 at 10% data (identical to one decimal).
Section 3.1 text: "The high-resolution image is divided into Np patches … The low-resolution one is then divided into Np patches."
Section 4.3: "two NVIDIA RTX 4090 GPUs" for MIMIC-CXR pretraining.

Notes

All quoted numeric evidence is preserved verbatim from the original report; readers should verify against the official proceedings when available.
Finding 4 (compute plausibility) is marked with reduced confidence in the original analysis ("依据不足"/insufficient basis) and should be treated as a soft flag pending author clarification on training duration and memory strategy.
This report does not constitute a formal finding of misconduct; final determination requires investigation by the ACM MM 2025 program committee, the publisher, or the authors' institution (Chongqing University of Posts and Telecommunications).
The paper remains pre-publication; recommended actions include contacting authors for raw logs and training records, posting concerns on PubPeer, and notifying the program committee.