Integrity Review: Contrastive Masked Image-Text Modeling for Medical Visual Representation Learning (MICCAI 2023, DOI: 10.1007/978-3-031-43904-9_48)

Academic fraud report · Geng Detector

Summary

This report assesses the paper 'Contrastive Masked Image-Text Modeling for Medical Visual Representation Learning' by Chen et al. (MICCAI 2023, LNCS 14224, DOI: 10.1007/978-3-031-43904-9_48). The verdict is CLEAN. Three checks were performed: (1) image reuse and splicing detection on Figures 1–3 (model architecture diagram, bar chart, line plot) found no evidence of pixel manipulation, reuse, flipping, or splicing, as the paper contains no wet-lab biological images; (2) statistical and data-fabrication checks on Tables 1 and 2 found numerical claims internally consistent, including the notable result that CMITM at 1% CheXpert labels achieves AUC 89.0 versus MRM at 100% labels (AUC 88.7); (3) timeline and methodological reasonableness checks confirmed citations to recent work (ICLR 2023 MRM, NeurIPS 2022 MGCA) and a plausible 4× Tesla V100 compute configuration for 2023. No follow-up action is required.

Verdict

CLEAN — No evidence of image manipulation, data fabrication, or methodological irregularity. The paper passes all six integrity checks applied.

Key findings

Image reuse/splicing (Figures 1–3): No anomalies. The figures are a model architecture diagram, a bar chart, and a line plot. No Western blots, microscopy, or gel images are present, so pixel-level duplication checks are not applicable.
Statistical consistency (Tables 1–2): All reported metrics are internally consistent with the narrative. The headline claim (CMITM at 1% CheXpert labels reaches AUC 89.0, exceeding MRM at 100% labels with AUC 88.7) is corroborated by the table entries. AUC decimal distributions look natural, not artificially uniform.
Timeline and methodology: Citations to ICLR 2023 (MRM) and NeurIPS 2022 (MGCA) are appropriate for a 2023 MICCAI submission. Reported compute (4× Tesla V100) is reasonable for ViT-B/16 + BERT multimodal pretraining in 2023. Use of MIMIC-CXR and CheXpert is standard. Cascaded training strategy and masking ratios are methodologically sound.
Open code: Authors released code at https://github.com/cchen-cc/CMITM, enabling full reproducibility.

Evidence highlights

DOI: 10.1007/978-3-031-43904-9_48
Table 1: CMITM on 1% CheXpert = 89.0 AUC; MRM on 100% CheXpert = 88.7 AUC — values match the text.
Reported improvements on NIH (+1.0%) and COVIDx (+1.5%) versus MRM are internally consistent with the table.
No $p$-values reported; AUC values across multiple seeds/epochs are presented in line with standard ML practice.

Notes

This is a pure deep-learning / computer-vision paper; traditional biomedical-image forensics (Western blot duplication, gel splicing, microscopy reuse) are not applicable.
The authors' code is open-sourced, which lowers the risk of unreproducible results. Interested parties may attempt independent reproduction via the GitHub repository.
Confidence: high for image and statistical checks; medium for novel-claim verification (would require full retraining to confirm). Limitations are inherent to any automated desktop review and do not constitute formal institutional findings.