Integrity Review Report: MedIM — Boost Medical Image Representation via Radiology Report-Guided Masking

Academic fraud report · Geng Detector

Summary

Verdict: No integrity issues identified (clean). This review examines the MICCAI 2023 paper by Yutong Xie et al. presenting MedIM, a self-supervised learning framework for medical imaging that uses radiology-report-guided masking. Key checks performed: (1) image reuse and splicing analysis was not possible because only extracted text was provided, no high-resolution figures were available; (2) numerical tables (Tables 1–3) show normal non-linear variation typical of deep-learning experiments, with no suspicious last-digit regularity; (3) datasets (MIMIC-CXR-JPG, CheXpert, COVIDx, SIIM-ACR), backbones (ViT, BioClinicalBERT), and baselines (MAE, GLoRIA, MGCA) are publicly known and temporally consistent with 2023 literature. The authors also released code at https://github.com/YtongXie/MedIM, reducing fabrication risk. Limitations: image-level forensics could not be conducted; conclusions are confined to textual and numerical analysis. Overall confidence in the clean verdict is moderate-to-high given the corroborating methodological details, but a definitive ruling requires full image inspection and institutional review.

Verdict

Clean — no substantive evidence of academic misconduct detected within the scope of this analysis. Confidence: moderate-to-high for textual and numerical checks; image-level checks were not performed.

Key findings

Image forensics not possible: Only extracted text was supplied, so pixel-level reuse and splicing detection on Figures 1–3 could not be executed. This is an information limitation, not an indication of paper fault.
Numerical data show realistic variation: Reported metrics exhibit the non-linear, seed-dependent fluctuations expected of deep-learning experiments, with no suspicious uniformity in terminal digits.
Methodology and citations are self-consistent: Datasets, backbones, and baselines are standard, publicly available resources aligned with the 2023 state of the art.
Open-source release reduces fabrication risk: The authors provide a public code repository, enabling independent verification of results.

Evidence highlights

Table 1: CheXpert mAUC for MedIM = 88.91 (1%), 89.25 (10%), 89.65 (100%); SIIM results = 63.50 (10%), 81.32 (100%). Comparable baseline MGCA* shows 88.11, 88.29, 88.88 — distributions consistent with normal experimental variance.
Datasets used (Section 3.1): MIMIC-CXR-JPG [12], CheXpert [10], COVIDx [18], SIIM-ACR [1] — all established public medical imaging benchmarks.
Models used (Section 2.1): ViT [7] and BioClinicalBERT [2] — standard architectures for the period.
Evaluation metrics (mAUC, Dice, R@k) match the norms of the medical image analysis community.
Code release: https://github.com/YtongXie/MedIM.
DOI: 10.1007/978-3-031-43907-0_2.

Notes

Image-level analysis (duplication, splicing, Photoshop artifacts) remains pending and should be conducted if original high-resolution figures become available.
The CheXpert 88.91/89.25/89.65 series and SIIM 63.50/81.32 pair were treated as illustrative only; no claim of independent re-verification is made.
This review is based solely on the text-extraction report supplied by the user and does not constitute an institutional investigation.
If concerns emerge from figure inspection or code audit, a follow-up review is recommended.