Integrity Assessment: Contrastive Masked Image-Text Modeling for Medical Visual Representation Learning (MICCAI 2023)

Academic fraud report · Geng Detector

Summary

Verdict: No academic fraud detected. This report evaluates the paper 'Contrastive Masked Image-Text Modeling for Medical Visual Representation Learning' by Chen et al., published in MICCAI 2023 (DOI: 10.1007/978-3-031-43904-9_48). Three checks were performed. First, text-data consistency: the claimed numerical improvements in the text (NIH +1.0%, COVIDx +1.5%, CheXpert 1% CMITM 89.0 vs. MRM 100% 88.7) match Table 1 (p. 499) exactly, with no suspiciously perfect integers or unnaturally uniform distributions. Second, citation and methodological timeline: references to MAE (CVPR 2022), MRM (ICLR 2023), and MGCA (NeurIPS 2022) are temporally plausible, and reported hardware (Tesla V100) and architectures (ViT-B/16, BERT) fit 2022-2023 norms. Third, image-manipulation analysis could not be performed because only plain-text extraction was available, so Figures 1-3 remain uninspected at the pixel level. Confidence in the cleared text and numerical dimensions is moderate-to-high; image-level integrity remains unverified.

Verdict

No academic misconduct detected at the level examined. The paper passes text-data consistency, citation timeline, and methodological plausibility checks.

Key findings

Text-data self-consistency (clear): The narrative claims in the paper align exactly with Table 1 (p. 499): CMITM (1%) - MRM (1%) on NIH = 80.4 - 79.4 = 1.0; on COVIDx = 79.5 - 78.0 = 1.5; on CheXpert, CMITM 89.0 (1% labels) exceeds MRM 88.7 (100% labels). No suspiciously rounded figures or implausible variance patterns were observed.
Citation and timeline plausibility (clear): All cited works (MAE CVPR 2022, MRM ICLR 2023, MGCA NeurIPS 2022) and the hardware/architectures (Tesla V100, ViT-B/16, BERT) are consistent with a 2023 publication date.
Image-level integrity (untested): Figures 1-3 could not be inspected for reuse or tampering due to plain-text-only extraction; this remains an open verification gap.

Evidence highlights

DOI: 10.1007/978-3-031-43904-9_48
Table 1 NIH: CMITM (1%) = 80.4 vs. MRM (1%) = 79.4 → +1.0
Table 1 COVIDx: CMITM (1%) = 79.5 vs. MRM (1%) = 78.0 → +1.5
Table 1 CheXpert: CMITM (1%) = 89.0 > MRM (100%) = 88.7
Source file: 978-3-031-43904-9_48 (1).pdf, pp. 499-503

Notes

Limitations: the review was performed on extracted text only; Figures 2 and 3 should be inspected in the original PDF for any duplicated panels, copy-pasted axes, or overlapping error bars.
No contact with the authors was deemed necessary based on text-level evidence.
This is an AI-assisted assessment and does not constitute a formal institutional investigation.