Integrity Concerns in 'Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Chest X-ray Vision-Language Pre-training' (ACM MM '25, DOI: 10.1145/3746027.3755336)

Academic fraud report · Geng Detector

Summary

This report flags multiple issues in a chest X-ray vision-language pre-training paper submitted to ACM Multimedia 2025 by authors at Chongqing University of Posts and Telecommunications. The overarching verdict is highly suspicious, marked by five confirmed concerns. Two stand out as serious: (1) misattributed citations, where references labeled as MGCA and Med-UniC actually point to MAE (He et al.) and SAT (Liu et al.); and (2) unsupported comparative claims, as the paper asserts MRM outperforms Med-UniC although Table 3 omits any Med-UniC data. Additional flagged problems include malformed numeric entries in Tables 1 and 5, an ablation in Table 4 where the touted PAR module lowers linear classification accuracy, and a lack of standard deviations or seed repetitions despite claiming significant gains of roughly 1%. Confidence in textual evidence is high; however, claims about the validity of underlying numerical results remain conditional until authors release source code and raw logs.

Verdict

Highly suspicious (🟠). Multiple confirmed integrity issues including incorrect citations, claims unsupported by presented tables, formatting anomalies, and self-contradictory ablation results. Statistical rigor is also lacking.

Key findings

Misattributed citations (Finding 1, 🟠): In Section 2.1 and Section 5.2/5.4.2, reference [12] is cited as MGCA but actually corresponds to He et al.'s MAE; reference [24] is cited as Med-UniC but is Liu et al.'s SAT, while Med-UniC is reference [34].
Unsupported comparative claim (Finding 2, 🔴): Section 5.2 asserts that reconstruction-based MRM [46] outperforms MGCA [12] and Med-UniC [34] on fine-grained tasks, yet Table 3 contains no Med-UniC [34] row, making the comparison unverifiable from the paper itself.
Data formatting artefacts (Finding 3, 🟡): Table 1 (Ours row) shows numbers concatenated as 89.5 89.589.792.1 93.2 93.8; Table 5 (KAD row) contains the malformed entry 66.020.9, indicating careless data transfer into LaTeX.
Negative ablation of headline module (Finding 4, 🟠): In Table 4, adding PAR reduces linear classification accuracy on CXR14 from 79.1 → 78.9 without DKBA and from 81.0 → 80.8 with DKBA, contradicting the paper's framing of PAR as a core contribution.
Missing statistical reporting (Finding 5, 🟡): Tables 1–3 report no standard deviations or multi-seed results; claims such as RSNA (1%) improvement to 92.1 vs. MRM's 90.9 (Δ ≈ 1.2%) are presented as significant without variance estimates.

Evidence highlights

Table 3 (Detection and Segmentation): no Med-UniC [34] row present despite the textual claim of direct comparison.
Reference list: [12] = He et al., MAE; [24] = Liu et al., SAT; actual MGCA = [35]; actual Med-UniC = [34].
Table 1 literal text contains 89.5 89.589.792.1 93.2 93.8; Table 5 KAD row contains 66.020.9.
Table 4 (CXR14 linear classification, %): baseline 79.1 → +PAR 78.9 (without DKBA); baseline 81.0 → +PAR 80.8 (with DKBA).

Notes

DOI preserved as 10.1145/3746027.3755336; venue is ACM MM '25 (33rd ACM International Conference on Multimedia).
Findings 1–4 are based on direct text/table inspection and are marked as confirmed by the reporter. Finding 5 reflects the absence of variance reporting rather than proven data fabrication.
Verification of reproducible experimental gains requires the authors' released code, training logs, and raw prediction outputs, particularly the MRM vs. Med-UniC comparison.
All conclusions are conditional pending an official institutional investigation.