Reviewer Note on 'Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for CheXray Vision-Language Pre-training' (DOI: 10.1145/3746027.3755336)

Academic fraud report · Geng Detector

Summary

This report evaluates a 2025 ACM Multimedia paper by Lihong Qiao and colleagues proposing a pathology-aware, alignment-based framework for chest X-ray vision-language pre-training. The overall verdict is inconclusive ('questionable'). No direct evidence of data fabrication or image manipulation was found; on the contrary, several reported ablation results (notable performance drops in Tables 4 and 6, and an explicit acknowledgement in Section 5.1 that the model underperforms Med-Unic on CheXpert) are atypical of fabricated benchmarks, lending credibility to the experimental pipeline. However, the manuscript shows serious editorial weaknesses: in Section 2.1 multiple citations appear mismatched to their reference list entries (e.g., [24] cited for MedUnic is actually a SAT paper, and [12] cited for ReCO is the MAE paper), and reference [13] lists 'J Healthcare' as an author, suggesting a journal name was mistakenly used in place of an author name. These point to careless writing or citation cross-wiring rather than systematic fraud. Pixel-level image forensics could not be performed because raw figures were unavailable.

Verdict

Questionable. The experimental data show signs of authenticity, but the manuscript contains severe citation mismatches and bibliographic errors that suggest very careless preparation. No conclusive evidence of research misconduct (data fabrication, image manipulation, or plagiarism) was identified.

Key findings

In Section 2.1 (Related Work), the citation for the description of MedUnic points to reference [24], which corresponds to a SAT paper by Bo Liu et al. — a clear citation mismatch.
In Section 2.1, the description associated with 'ReCO [12]' cites reference [12], which is the MAE (Masked Autoencoders) paper, not ReCO.
Reference [13] lists 'J Healthcare' as the author, evidently a journal name mistakenly placed in the author field; the entry reads "[13] J Healthcare. 2020. Object-cxr-automatic detection of foreign objects on chest x-rays."
Ablation tables show realistic, non-monotonic behaviour: in Table 4, adding only PAR reduces AUC on ChestX-ray14 from 79.1 to 78.9; in Table 6, the jointly trained row underperforms the Alignment-only row on COVIDx.
Section 5.1 explicitly states that the proposed model underperforms Med-Unic on CheXpert, which is unusual for a paper presenting a new method.
Pixel-level image forensics could not be performed because raw, uncompressed figures were not available; Figure 5 (t-SNE) and Figure 6 (attention maps) in particular could not be examined.

Evidence highlights

Reference list vs. in-text citations: [24] = SAT (Bo Liu et al.), [12] = MAE — both inconsistent with claims in Section 2.1.
Reference [13] author field anomaly: "J Healthcare" is not a person.
Table 4 row 2 (PAR only, ChestX-ray14): 78.9 vs. baseline 79.1.
Table 6 row 3 (joint training, COVIDx): lower than row 2 (Alignment only).
Section 5.1 self-critical statement regarding CheXpert vs. Med-Unic.
DOI: 10.1145/3746027.3755336.

Notes

The combination of non-monotonic ablations and self-acknowledged underperformance is consistent with real multi-task optimisation dynamics, including gradient conflicts and competing loss terms.
Citation cross-wiring and bibliographic errors may stem from sloppy copy-paste workflows or unreliable reference-manager exports; they are indicative of poor scholarship but do not by themselves constitute fraud.
A formal misconduct complaint is not warranted on current evidence.
A constructive PubPeer comment requesting the authors to verify Section 2.1 citations and reference [13] would be appropriate.
Request to authors for uncompressed originals of Figure 5 and Figure 6 would enable a complete image-integrity review.
This report is AI-assisted and intended for academic discussion only; final determinations rest with appropriate institutional or editorial investigations.