Integrity Screening 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

Verdict: Questionable (yellow flag). The 2025 ACM MM '25 paper by Lihong Qiao et al. presents a vision-language pre-training framework for chest X-rays. Three concerns are raised. First, the authors repeatedly claim significant improvements (e.g., 96.0% accuracy on COVIDx) yet report no standard deviation, standard error, or significance test across Tables 1–3 and Sections 5.1–5.2, making ~1% gains (e.g., RSNA mAP 78.2 → 79.3) indistinguishable from seed variance. Second, Section 3.1 shows a notation inconsistency: Eqs. 1–2 use the variable M_V while Eq. 3 silently switches to P_V, and the compute claim of pre-training ViT-B/16 plus multimodal BERT and GAT on >370k MIMIC-CXR images using only 2× RTX 4090 GPUs is implausible without undisclosed distributed training or extreme gradient accumulation. Third, Section 4.1 uses MIMIC-CXR, CheXpert, and RSNA but provides no IRB, DUA, or HIPAA compliance statement. No image manipulation or data fabrication is alleged. Confidence: moderate for statistical and notation issues, lower for the compute claim which requires code verification.

Verdict

Questionable (🟡). Three methodological and reporting concerns were identified. No evidence of image manipulation or data fabrication is asserted at this stage.

Key findings

Missing statistical dispersion across all main tables. Tables 1, 2, 3 and Sections 5.1–5.2 report point accuracies/mAP only. The COVIDx result of 96.0% (Table 1) and the RSNA mAP lift from 78.2 (MRM) to 79.3 (Ours) are framed as significant improvements with no standard deviation, standard error, p-value, or t-test across seeds. A ~1% delta lies within typical random-seed variance for deep-learning pre-training.
Notation inconsistency in the pathology-aware prior module. Section 3.1, Eqs. (1) and (2) define and normalize M_V, but the reconstruction loss in Eq. (3) abruptly switches to P_V without definition, indicating a typographical or write-up error that affects reproducibility.
Implausible compute budget for MIMIC-CXR pre-training. Section 4.3 states pre-training was performed on 2× NVIDIA RTX 4090 (24 GB each) for ViT-B/16 (200 epochs) together with multimodal BERT and a GAT, over 370k images and 220k reports. Such a configuration is unlikely to support a sufficient batch size for convergence on dual-path 448×448 / 224×224 inputs without undisclosed distributed training or extreme gradient accumulation.
Absent ethics / data-use statement. Section 4.1 (Datasets) uses MIMIC-CXR, CheXpert, and RSNA but provides no IRB approval, de-identification note, or Data Use Agreement (e.g., PhysioNet Credentialed Access / HIPAA) statement, an unusual omission for clinical AI work at a top venue.

Evidence highlights

DOI: 10.1145/3746027.3755336
Table 1: COVIDx (100% data) accuracy = 96.0%, claimed "significant improvement."
Table 2/3: RSNA mAP progression cited as 78.2 → 79.3 (≈1.1 percentage points), no error bars.
Section 3.1, Eq. (1)–(2): variable M_V; Eq. (3): variable P_V (undefined transition).
Section 4.3: compute claim — 2× RTX 4090, ViT-B/16, 200 epochs, >370k images, >220k reports, BERT + GAT.
Code repository cited in suggested actions: https://github.com/Felix1118/PADKB (to be inspected for full pre-training scripts).

Notes

This report does not allege data fabrication or image manipulation; it flags reporting and reproducibility weaknesses.
The compute claim cannot be conclusively falsified without inspecting the released code and training logs; verification is recommended via the GitHub repository.
Final determination of any misconduct requires investigation by an authorized body.