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

Academic fraud report · Geng Detector

Summary

This report examines the MICCAI 2023 paper by Chen et al. (DOI: 10.1007/978-3-031-43904-9_48) for signs of data fabrication. The overall verdict is a strong indication of fraud (🔴 confirmed in the original assessment). The central finding concerns Table 1 (CheXpert dataset), where the proposed CMITM model and the MRM baseline exhibit a mathematically implausible pattern: both improve by exactly 0.2% from 1% to 100% labels, and CMITM leads MRM by a constant +0.5% across all data fractions—a degree of arithmetic regularity incompatible with independent stochastic training. A second finding shows suspiciously engineered micro-margins on NIH X-ray (0.1%–1.0%) and RSNA datasets (-0.1% to +0.3%), consistent with hand-tuned fabrication to dramatize low-data gains. Third, Table 1 and Table 2 share identical values to one decimal place under 1% labels (e.g., RSNA AUC 91.6; COVIDx ACC 79.5), suggesting the ablation table was copy-pasted from the main results. Confidence is high for findings 1–3 based on numeric evidence; a fourth concern about Figures 2–3 remains unverified without visual access.

Verdict

🔴 Strong indication of data fabrication (per original reviewer assessment). Three numeric anomalies in Tables 1–2 reach the threshold of evidence; one figure-related concern is flagged but unverified.

Key findings

Table 1 / CheXpert (Page 499): MRM goes 88.5 → 88.5 → 88.7 (Δ = +0.2 from 1% to 100%); CMITM goes 89.0 → 89.0 → 89.2 (Δ = +0.2). CMITM − MRM = +0.5 at every data fraction. Two distinct architectures across a 100× data-size sweep yielding identical deltas and a constant offset is statistically untenable.
Table 1 / NIH X-ray (Page 499): CMITM − MRM margins are +1.0 (1%), +0.1 (10%), +0.1 (100%)—artificially inflated at the headline 1% setting.
Table 1 / RSNA (Page 499): Margins are +0.3 (1%), −0.1 (10%), +0.1 (100%)—the negative point appears engineered to feign realism.
Table 1 vs. Table 2 reuse (Page 499–500): CMITM's RSNA AUC at 1% labels = 91.6 in both tables; COVIDx ACC at 1% labels = 79.5 in both tables. Bit-for-bit reuse across independent experiments under 1% data is highly improbable without copying.
Figures 2–3 (Page 500): Visualization claims cannot be cross-checked from text alone; flagged for further inspection.

Evidence highlights

CheXpert triple: (88.5, 88.5, 88.7) vs. (89.0, 89.0, 89.2) — constant Δ = +0.2 and constant offset +0.5.
NIH X-ray margins at 1% / 10% / 100%: +1.0 / +0.1 / +0.1.
RSNA margins at 1% / 10% / 100%: +0.3 / −0.1 / +0.1.
Cross-table exact match at 1% labels: RSNA AUC 91.6, COVIDx ACC 79.5.
DOI: 10.1007/978-3-031-43904-9_48

Notes

Authors: Cheng Chen, Aoxiao Zhong, Dufan Wu, Jie Luo, Quanzheng Li (affiliations include Massachusetts General Hospital / Harvard Medical School).
All numeric values are reproduced verbatim from the source report and paper extract; no new figures were derived.
Limitations: no visual inspection of Figures 2–3 was possible, so the fourth finding remains unverified. Final institutional adjudication requires access to training logs, seeds, and code.