Integrity Review Report: Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Chest X-ray Vision-Language Pre-training

Academic fraud report · Geng Detector

Summary

This report assesses a paper published at ACM MM '25 (DOI: 10.1145/3746027.3755336) by Lihong Qiao et al., proposing a chest X-ray vision-language pre-training framework. The overall assessment is highly concerning. Five issues are documented: (1) systematic citation-number mismatches in Sections 2.1, 2.2, and 5.2, where methods such as Med-UniC, ReCO, and MGCA are linked to wrong bibliographic numbers (e.g., MAE [12] cited in place of MGCA); (2) contradictory ablation claims in Section 5.3.1 versus Table 4, where the full model scores 80.8 on ChestX-ray14, lower than DKBA alone at 81.0, contradicting the stated gains; (3) implausible baseline behavior in Table 2, where GLoRIA's RSNA Dice drops from 68.7 (10% data) to 68.3 (100% data); (4) irregular formatting in Table 1 suggesting manual editing; (5) image analysis was not performed due to lack of pixel data. Confidence is high for textual issues; image-based conclusions are unavailable. No definitive misconduct finding is made; institutional verification is recommended.

Verdict

🟠 Highly suspicious. Multiple textual and tabular inconsistencies suggest careless preparation, possible LLM-assisted drafting without verification, or selective reporting. No confirmed misconduct finding; further institutional review is recommended.

Key findings

Systematic citation-number mismatches: Methods cited in the text do not correspond to the listed references (e.g., MAE [12] is cited where MGCA should appear; ReCO is linked to MAE).
Self-contradictory ablation claims: ChestX-ray14 linear classification result for the full model (80.8) is lower than DKBA-only (81.0), contradicting the narrative of universal improvement.
Implausible baseline regression: GLoRIA's RSNA Dice coefficient decreases from 68.7 (10% data) to 68.3 (100% data), which is atypical for deep-learning segmentation.
Irregular table formatting in Table 1: Values such as "90.891.993.1" appear without spaces; some baseline differences are suspiciously uniform.
Image analysis not performed: Figures were not available in pixel form; visual evidence could not be examined.

Evidence highlights

Section 2.1 attributes Med-UniC to reference [24], but reference [24] is a Bo Liu et al. SAT paper; Med-UniC corresponds to reference [34] (Zhongwei Wan et al., 2024).
Section 2.1 cites ReCO as reference [12]; reference [12] is He Kaiming et al.'s MAE paper.
Section 5.2 again cites MGCA as reference [12]; MGCA actually maps to reference [35].
Table 4 (ChestX-ray14): baseline 79.1; +PAR 78.9; +DKBA 81.0; Full (PAR+DKBA) 80.8 — full model underperforms DKBA-only by 1.9 points.
Table 2 (RSNA): CNN-based GLoRIA [14] Dice = 68.7 at 10% data vs. 68.3 at 100% data; the proposed model reports 79.3 in the same setting.
Table 1 formatting anomaly: "89.4 89.7 90.891.993.193.780.389.594.5" contains a missing space between 90.8 and 91.9; MRM values include suspiciously regular increments such as 88.5/88.5/88.7 and 90.9/92.3/92.7.

Notes

Confidence: high for findings 1–3 based on direct text and table evidence; moderate for finding 4 due to alternative benign explanations (LaTeX rendering artifact); unknown for image-related issues due to lack of pixel data.
DOI: 10.1145/3746027.3755336 — preserved exactly as provided.
The authors have open-sourced code at https://github.com/Felix1118/PADKB; reproducing the reported numbers is recommended as the next verification step.
A PubPeer comment and an Errata request to ACM MM '25 are advised if findings are confirmed.
This report is AI-assisted and intended for academic discussion only; final determination of misconduct requires formal investigation.