Integrity Review Report: Contrastive Masked Image-Text Modeling for Medical Visual Representation Learning

Academic fraud report · Geng Detector

Summary

Verdict: No evidence of academic misconduct found. This report reviews a MICCAI 2023 paper (DOI: 10.1007/978-3-031-43904-9_48) by Cheng Chen, Aoxiao Zhong, Dufan Wu, Jie Luo, and Quanzheng Li. Three aspects were examined: (1) methodological consistency and citation timeline — all referenced works, including ICLR 2023 MRM, were publicly available prior to submission; (2) experimental data consistency — the identical AUC value of 89.0 at 1% and 10% training labels on CheXpert is plausible given threshold or convergence bottlenecks and follows the overall trend (89.0, 89.0, 89.2); (3) figures and tables — text-only description limits pixel-level forensic analysis, but ablation logic is consistent and code is declared open-source. Limitations: the review is constrained by lack of access to high-resolution images for duplication or PS detection, and no variance was reported (a common but noted gap in computer science venues). Confidence in the 'clean' verdict is moderate-high, pending independent image forensics and code reproducibility checks.

Verdict

No evidence of academic misconduct detected. The paper is assessed as likely legitimate, with logical methodology, reasonable data trends, and coherent ablation results. The review is constrained by text-only access to figures and the absence of variance reporting.

Key findings

Citation timeline and methodological references are fully consistent with submission timing of MICCAI 2023.
Identical AUC value (89.0) at 1% and 10% labels on CheXpert is plausible and not indicative of fabrication; values follow an internally consistent ascending trend (89.0 → 89.0 → 89.2).
Ablation study logic is internally consistent with conventional deep learning behavior (e.g., removing cascaded training causes notable performance drop).
Authors transparently attribute reused components (ViT-B/16, BERT, prior loss functions) to prior work.
Code declared open-source on GitHub, supporting reproducibility.

Evidence highlights

Location: Table 1 (Page 499). CheXpert AUC values: 89.0 (1%), 89.0 (10%), 89.2 (100%). No suspicious terminal-digit regularity observed.
Citations: MAE [8], MRM [28] (ICLR 2023), MGCA [22] — all temporally valid relative to MICCAI 2023 submission.
Ablation (Table 2): Performance drops align with conventional expectations when key components are removed; no reverse-ablation anomalies.
DOI: 10.1007/978-3-031-43904-9_48 (MICCAI 2023, LNCS 14224).

Notes

Image forensics: Not performed — figures were only available as text descriptions in the source PDF, preventing pixel-level checks for splicing, copy-paste, or prior-publication overlap.
Reproducibility: Author-declared open-source code was not independently executed in this review; results rely on reported numbers.
Variance reporting: No standard deviation or confidence intervals reported, which is common in MICCAI submissions but limits strength of claims.
Recommendation: No further action indicated. Future reviews with full-resolution figures and code execution may strengthen confidence.