Fraud Detection Report: Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training (DOI: 10.1145/3746027.3755336)

Academic fraud report · Geng Detector

Summary

This is a Chinese 'Geng' academic-fraud detection report concerning the paper 'Pathology-Aware Reconstruction with Discriminative Knowledge Boosting Alignment for Che-Xray Vision-Language Pre-training' by Lihong Qiao et al., accepted at ACM Multimedia 2025 (DOI: 10.1145/3746027.3755336). The overall verdict is 'questionable' (yellow/ambiguous). After re-review, one substantive concern is confirmed: a logical inconsistency in the ablation study. Table 4 shows that adding the PAR module to the DKBA baseline on ChestX-ray14 actually decreases AUC from 81.0 (DKBA alone) to 80.8 (PAR + DKBA), contradicting the authors' claim of complementary gains. Four other allegations—self-contradictory captions, an alleged fatal formula error, infeasible compute, and image manipulation—were judged insufficiently supported upon re-examination. Confidence in the confirmed issue is reasonable; broader fraud allegations remain unsubstantiated.

Verdict

🟡 Questionable — one substantive concern was confirmed; four additional allegations were judged insufficiently supported upon re-review.

Key findings

Ablation logic inconsistency (CONFIRMED). In Table 4 (Ablation analysis, Page 6), on ChestX-ray14, DKBA alone reports AUC = 81.0, while PAR + DKBA reports AUC = 80.8. Section 5.3.1 explicitly claims PAR significantly improves performance and that PAR and DKBA provide complementary benefits, but adding PAR to DKBA produces a regression rather than an improvement on this dataset.
Caption vs. table inconsistency (UNSUPPORTED). The caption notes Med-UniC metrics on some datasets are excluded due to missing official checkpoints, and Table 1 indeed shows '-' entries for Med-UniC across several configurations (e.g., CXP, RSNA, COVIDx under CNN-based; CXR14 under ViT-based), which is consistent with the caption statement.
Loss-function formula (UNSUPPORTED). Equation 3 defines L_par = (1/Nv)·Σ( S(i)·‖X_h − X_rec‖² ) with a min-max normalized weight matrix S(i). The report's claim that this constitutes a severe mathematical error is not generally established; per-element weighted MSE losses without explicit sum-normalization are common in deep learning practice.
Compute feasibility (UNSUPPORTED). The paper states training on 2× NVIDIA RTX 4090 GPUs over MIMIC-CXR (≈377k images) with ViT-B/16 and BERT. This configuration is plausible in current deep-learning engineering practice; the claim of impossibility lacks direct evidence of fabrication.
Image manipulation (UNSUPPORTED). No pixel-level analysis of Figures 1–6 was possible from text-only input; suspicion of overly perfect t-SNE clusters or other visual artifacts lacks substantive evidence.

Evidence highlights

Table 4 ablation values on ChestX-ray14: DKBA-only = 81.0; PAR + DKBA = 80.8 (a decrease of 0.2 AUC after adding the headline PAR module).
Section 5.3.1 verbal claim of complementary gains from PAR and DKBA, including stated gains of ≈2% on ChestX-ray14, which does not reconcile with the observed regression.
Table 1 (Page 5): Med-UniC rows show '-' entries across multiple datasets and architectures, supporting the caption's 'some datasets excluded' statement.
Section 4.3 (Page 5): Implementation details describe 2× RTX 4090, ViT-B/16, BERT backbone, 200 epochs of reconstruction pre-training plus 15 epochs of alignment on MIMIC-CXR.
DOI: 10.1145/3746027.3755336

Notes

The single confirmed issue may reflect (a) genuine model instability or high variance in ablation runs, (b) selective reporting, or (c) post-hoc tuning of the full model. It does not by itself prove systematic fraud.
Recommended actions: request authors release full training logs and per-seed variance for Table 4, particularly the ChestX-ray14 row; reproduce the ablation under fixed seeds; cross-check error bars or significance tests.
Other allegations are recorded for transparency but should not be cited as evidence of misconduct absent further verification (e.g., image-forensic analysis, code release, independent reproduction).
Limitations: text-only review cannot adjudicate image-based claims; statistical checks (e.g., equivalence testing) are not possible from the reported numbers alone.