Geng Integrity Report: FICMD-Med: A Novel Framework for Imbalanced Class and Modality Distributions of Medical Images (DOI: 10.1109/BIBM66473.2025.11356481)

Academic fraud report · Geng Detector

Summary

This Geng integrity report flags the paper 'FICMD-Med: A Novel Framework for Imbalanced Class and Modality Distributions of Medical Images' (DOI: 10.1109/BIBM66473.2025.11356481), presented at the 2025 IEEE BIBM conference, as highly suspicious. Three core issues are identified. First, a methodological inconsistency in Section II.C.3 (Joint Strategy): the paper defines the hyperparameter w_mix in equation (7), but the surrounding text refers to an ablation on λ, with no corresponding ablation study found in Section III or the appendix, suggesting copy-paste or AI hallucination. Second, a statistical anomaly: the experimental setup states that three independent runs were conducted and mean ± standard deviation were reported, yet Tables I, II, and III show only single point values (e.g., 83.04, 95.35) for Acc, Recall, F1, and AUC, with no ± symbols or variance. Third, methodological gaps in the Modality-focus Expert Enhancement module: Figure 2 mentions RAG using VLM and LLM, but the paper lacks details on how these are integrated, deferring to GitHub, and Figure 1 caption text shows garbled concatenated phrases. Limitations of the review include reliance on textual and visual inspection of the PDF; verification of reproducibility is not possible without the original code and raw data.

Verdict

🟠 Highly suspicious. Multiple internal inconsistencies between stated methodology and reported experimental evidence suggest either careless assembly of content from heterogeneous sources, inadequate self-review, or AI-assisted writing without proper verification. No direct evidence of data fabrication is established, but the statistical and methodological discrepancies warrant author clarification and independent verification.

Key findings

Hyperparameter name mismatch (Section II.C.3): Equation (7) defines w_mix ∈ [0,1] as the mixture weight, but the same paragraph claims an ablation study was conducted on "different λ values," which cannot be located in Section III or the appendix.
Missing standard deviations (Section III.A.4, Tables I–III): The setup states three independent runs were performed and mean ± standard deviation are reported, yet all numeric entries in Tables I, II, and III (Acc, Recall, F1, AUC; e.g., 83.04, 95.35) are bare single numbers with no ± symbols or variance information.
Undetailed RAG/VLM/LLM integration (Section II.B, Figures 1–2): Figure 2 labels "RAG (Using VLM&LLM)" and the text references BioGPT or DistilBERT for caption generation, but the paper provides no substantive methodological description of the RAG mechanism and defers to GitHub.
Garbled figure text (Figure 1): Caption text contains concatenated fragments such as "gradient optimizationLack adaptive weighting Introduce noise Coarse-grained gradient optimization," suggesting unpasted and unreviewed content.

Evidence highlights

Equation (7) text: "Here, w_mix ∈ [0,1] is the mixture weight" followed by "we conduct an ablation study on different λ values" — no matching ablation table or figure for λ is present.
Experimental setting verbatim claim: "we conduct three individual runs with different seeds and report the mean and standard deviation."
Tables I, II, and III contain only point values for Acc, Recall, F1, and AUC (examples: 83.04, 95.35) with no variance notation.
Figure 2 explicitly labels "RAG (Using VLM&LLM)"; Section II.B mentions BioGPT or DistilBERT but provides no RAG implementation detail.
Figure 1 caption displays compressed/joined fragments: "gradient optimizationLack adaptive weighting Introduce noise Coarse-grained gradient optimization."
DOI: 10.1109/BIBM66473.2025.11356481

Notes

Severity ratings in the original report: Finding 1 🔴, Finding 2 🟠, Finding 3 🟡; all three have ✅复核状态 (re-verification status) marked as 成立 (confirmed).
The analysis is limited to the manuscript PDF; claims about reproducibility cannot be verified without access to source code, training logs, and raw experimental outputs.
The author(s) should be asked to provide the claimed ablation study on λ (or clarify the relationship between λ and w_mix), supply the standard deviations for Tables I–III, and document the RAG/VLM/LLM pipeline.
Final determination of academic misconduct requires investigation by the conference organizers, IEEE, or the authors' institutional ethics committee.
The original report includes an explicit AI-assisted generation disclaimer; this translation preserves that caveat.