Integrity Assessment: 'Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training' (ACM MM '25)

Academic fraud report · Geng Detector

Summary

This report raises concerns about an ACM Multimedia 2025 paper proposing PAR and DKBA modules for chest X-ray vision-language pre-training. The verdict is 'Questionable.' The central finding is a logical inconsistency in the ablation study (Table 4): the authors claim the two modules provide 'complementary benefits,' yet on ChestX-ray14, adding PAR to DKBA alone decreases AUC from 81.0 to 80.8, contradicting the stated conclusion. A secondary finding shows a citation error in Section 5.2, where the MRM/alignment baseline 'MGCA' is attributed to reference [12], which is actually He et al.'s MAE paper; the correct MGCA reference is [35]. This points to careless scholarship. A third claim about possible image manipulation could not be verified due to lack of figure files. Confidence is high for the two verified textual/logical issues; the image-analysis component remains unverified. The overall pattern undermines confidence in the rigor of the submission but does not by itself constitute definitive misconduct.

Verdict

Questionable — Two independent textual/logical issues are verified (ablation self-contradiction and citation misattribution); image-level analysis could not be performed. No conclusive proof of fraud, but the documented inconsistencies significantly weaken confidence in the paper's scholarly rigor.

Key findings

Ablation logic contradicts stated conclusion: on ChestX-ray14, DKBA alone yields AUC 81.0, while PAR + DKBA drops to 80.8 — inconsistent with the authors' claim of 'complementary benefits.'
Citation misattribution in Section 5.2: 'MGCA' is cited as [12], but [12] corresponds to He et al.'s 'Masked autoencoders are scalable vision learners' (MAE); the correct MGCA citation is [35] (Fuying Wang et al.).
Image-level analysis (Figures 1–6) could not be conducted because only plain text was available; no pixel-level duplication or fabrication checks were possible.
The paper reports gains on other metrics, so the concerns are localized rather than systemic, pending verification.

Evidence highlights

Table 4 (Section 5.3.1): ChestX-ray14 row, PAR=✗ / DKBA=✓ → AUC 81.0; PAR=✓ / DKBA=✓ → AUC 80.8. The full-configuration underperforms the DKBA-only setting on this dataset.
Section 5.3.1 prose: "These results collectively validate the complementary benefits of PAR for detailed pathological feature extraction and DKBA for discriminative representation learning." This claim is not supported by the ChestX-ray14 column.
Section 5.2 prose: "...the reconstruction-based MRM [46] outperforms MGCA [12] and Med-Unic [34]..." Reference [12] resolves to MAE (He et al.), not MGCA; MGCA appears at [35].
Reference list discrepancy: Confirms incorrect numeric mapping between textual mention and bibliography.
DOI: 10.1145/3746027.3755336 (ACM MM '25).

Notes

Severity ratings from the original report are preserved: Finding 1 = 🟠 (moderate-high), Finding 2 = 🟡 (moderate), Finding 3 = ⚠️ unverified.
The ablation contradiction may reflect (a) genuine negative interaction between modules, (b) selective reporting/cherry-picking, or (c) an authoring oversight; without raw logs, intent cannot be determined.
The citation error, while possibly a typo, is notable because MAE (reconstruction-based) and MGCA (alignment-based) belong to different methodological families — the swap is not a mere index error in flavor.
Recommended actions: request raw experiment logs and hyperparameter sweeps; query authors about the ChestX-ray14 regression; raise a PubPeer comment; notify the ACM MM '25 program chairs if additional irregularities emerge.
Limitations of this assessment: no figure files were available, so duplication/fabrication of visualizations (e.g., t-SNE, heatmaps) remains untested.