Critical Logic Inversion in Core Masking Formulation of MedIM (MICCAI 2023)

Academic fraud report · Geng Detector

Summary

Verdict: Substantive finding (red flag). A detailed review of the MICCAI 2023 paper 'MedIM: Boost Medical Image Representation via Radiology Report-Guided Masking' (DOI: 10.1007/978-3-031-43907-0_2) by Yutong Xie, Lin Gu, Tatsuya Harada, Jianpeng Zhang, Yong Xia, and Qi Wu identifies a fundamental contradiction between the paper's stated methodology and its mathematical formulation. The authors claim to introduce a 'high-activated masking' strategy that masks the most informative image regions identified by a MeSH-derived activation map, forcing the model to reconstruct clinically relevant areas. However, Equation (3) and its binary indicator definition implement the exact opposite operation: high-activation pixels are preserved (m=1 yields the original feature), while low-activation pixels are replaced with [MASK]. The ablation in Figure 3 (left), which reports that the authors' 'High-activated masking' outperforms random and 'Low-activated masking' baselines, appears to contradict the formal definition unless the released code does not match Eq. (3). Confidence is high for the textual/formal contradiction; uncertainty remains regarding whether the released GitHub code actually implements the high-activation masking as described in prose, which would resolve or worsen the concern.

Verdict

🔴 Substantive finding — core method description and core mathematical formulation are mutually contradictory. Recommend editorial erratum or retraction review pending code inspection.

Key findings

Text–formula inversion (Eq. 3): The prose defines the contribution as 'selectively mask then restore the high-activated regions' / 'high-activated masking … to remove the discovered attention regions'. The binary mask definition m_{(i,j)} = I(C^{MeSH}_{(i,j)} ≥ C^{MeSH}_{[γ·N_img]}) assigns m=1 to high-activation pixels. Substituting into M = { z_{(i,j)} | m·R(E_img) + (1−m)·[MASK] } gives: high-activation → R(E_img) (kept); low-activation → [MASK] (masked). This implements low-activated masking, the inverse of the claimed method.
Ablation appears inconsistent with the formula: Figure 3 (left) reports that the authors' 'High-activated masking' beats Random and 'Low-activated masking'. If the code followed Eq. (3), the configuration labelled 'High-activated masking' would in fact be performing low-activation masking, raising the possibility of prose–code mismatch or selective reporting.
Severity: The contradiction concerns the paper's headline contribution in a MICCAI 2023 main-conference paper; even if unintentional, it undermines reproducibility and the validity of the reported gains.

Evidence highlights

Stated method (Section 2.2, Report-Guided Mask Generation): 'selectively mask then restore the high-activated regions'; 'high-activated masking … to remove the discovered attention regions'.
Formal definition: m_{(i,j)} = I(C^{MeSH}_{(i,j)} ≥ C^{MeSH}_{[γ·N_img]}), with mask assembly given by M = { z_{(i,j)} | m_{(i,j)}·R(E_img)_{(i,j)} + (1−m_{(i,j)})·[MASK] }.
Algebraic evaluation: m=1 (high activation) → output equals R(E_img) (no masking); m=0 (low activation) → output equals [MASK].
Ablation (Section 3.3, Figure 3, left): 'our High-activated masking performs better than the Random and Low-activated masking'.
DOI: 10.1007/978-3-031-43907-0_2; venue: MICCAI 2023, LNCS 14220.

Notes

The finding relies solely on the published text and equations; no image-manipulation or statistical forensics were applied.
Resolution requires inspecting the official GitHub release to determine whether the implementation masks high- or low-activation patches. If the code matches the prose, Eq. (3) is wrong and an erratum is warranted. If the code matches Eq. (3), the prose and the ablation labelling are misleading, which is a separate integrity concern.
No evidence is presented regarding authorship misconduct, data fabrication, or peer-review manipulation; conclusions are limited to the methodological inconsistency described above.
Marked as AI-assisted preliminary assessment; final determination should be made by the relevant conference/integrity body.