Integrity Review of MedIM: Boost Medical Image Representation via Radiology Report-Guided Masking (MICCAI 2023)

Academic fraud report · Geng Detector

Summary

Overall verdict: CLEAN (no evidence of fraud). This review examined the MICCAI 2023 paper introducing MedIM (DOI: 10.1007/978-3-031-43907-0_2) for potential academic misconduct. Three findings were central. First, the report format precluded pixel-level image analysis of Figures 1–3, so Western blot/PS-tampering checks could not be performed. Second, internal numerical consistency is strong: the last row of the ablation table (Table 3, full MedIM with L_align, KWM, and SDM) reports CheXpert, COVIDx, and SIIM scores of 89.25, 90.34, and 63.50, which match the Table 1 figures for the 10% labeled-data setting precisely—a signature unlikely under fabricated data. Third, reported AUC, accuracy, and recall values (e.g., 88.91, 77.22, 5.74) show natural last-digit variation consistent with genuine stochastic training. The timeline of datasets (MIMIC-CXR-JPG 2019, CheXpert 2019, COVIDx 2020, SIIM-ACR 2019) and baselines (MAE 2022, MGCA 2022, MRM 2023) predates the 2023 publication. Code is released. Confidence: moderate pending access to raw figures and code execution.

Verdict

CLEAN – No actionable evidence of academic fraud was identified. The paper passes internal-consistency, timeline, and plausibility checks. A formal image-forensics audit was not possible.

Key findings

Pixel-level image analysis not performed: Figures 1–3 were not available in raw pixel form, so reuse, splicing, or PS manipulation in Western blots, gels, or microscopy images cannot be evaluated.
Strong numerical consistency between main and ablation tables: The complete-model row in Table 3 (L_align + KWM + SDM) yields 89.25 (CheXpert), 90.34 (COVIDx), and 63.50 (SIIM), which exactly match the 10% labeled-data figures in Table 1—an alignment inconsistent with fabricated data.
Plausible metric granularity: Reported AUC, accuracy, and recall values (e.g., 88.91, 77.22, 5.74) show natural last-digit variation rather than suspiciously round or monotonically progressing numbers; CheXpert performance scales plausibly from 1% to 100% labeled data.
Chronologically valid references: Datasets (MIMIC-CXR-JPG 2019, CheXpert 2019, COVIDx 2020, SIIM-ACR 2019) and baselines (MAE 2022, MGCA 2022, MRM 2023) all predate the 2023 publication.
Open code availability: The authors publicly release code, which is a positive credibility signal in computer-science research.

Evidence highlights

DOI: 10.1007/978-3-031-43907-0_2
Numerical match: Table 3 full-model row = Table 1 (10% labels) = 89.25 / 90.34 / 63.50 on CheXpert / COVIDx / SIIM.
Sample metric values: 88.91, 77.22, 5.74 across Table 1 and Table 2.
Datasets and baselines dated 2019–2023, all prior to publication year.

Notes

The text-only nature of the review limits forensic checks of Fig. 1–3 and any supplementary images; a pixel-level re-examination is recommended if high-resolution originals become available.
No statistical re-analysis (e.g., recomputation of variance, seed sensitivity) was performed; the natural last-digit variation is suggestive but not conclusive proof of authentic training runs.
The review is AI-assisted and intended for academic discussion only; final determinations of misconduct require institutional investigation.