Geng Fraud-Detection Report: ShiftMorph — A Fast and Robust Convolutional Neural Network for 3D Deformable Medical Image Registration (ACM MM 2024 submission)

Academic fraud report · Geng Detector

Summary

Verdict: Clean (✅). This report evaluates a double-blind ACM Multimedia 2024 submission proposing ShiftMorph, a CNN-based framework for 3D deformable medical image registration evaluated on OASIS, IXI, and Lung250M-4B. Because no source images were available for pixel-level inspection, the analysis focused on tabular data consistency, methodological descriptions, references, and timeline plausibility. Table 1 (Dice, HD95, Folds) shows internally coherent rankings with no suspicious last-digit clustering. Cross-table consistency is confirmed: ShiftMorph's self-consistent Dice in Table 4 (0.8116) matches Table 1. Two minor editorial defects were noted: missing VoxelMorph++ values in Table 2 and unconventional '%'/'!' symbols in Table 4 (LaTeX rendering error), which appear to be careless typesetting rather than data fabrication. The 2024 submission date is consistent with a 2026 review window, and cited methods (TransMorph, TransMatch) and the corrfield challenge link are real and current. No image manipulation, fabricated citations, or timeline anomalies were detected. Confidence is moderate due to the inability to perform image-based forensic checks.

Verdict

✅ Clean — No substantive evidence of academic fraud. Findings are limited to minor editorial and typesetting oversights. Final determination should still be made by the conferring institution or a formal investigation.

Key findings

Tabular data in Table 1 is internally consistent: Dice scores span 0.7947 (rank 9) to 0.8266 (rank 1) on OASIS with coherent HD95 and Folds values; no last-digit anomalies (no clustering on 0 or 5).
Cross-table consistency holds: ShiftMorph's self-consistent Dice of 0.8116 in Table 4 matches the ShiftMorph value in Table 1.
Two non-fraud editorial defects were observed:
Table 2: VoxelMorph++ entries for TRE opt. (mm) and Folds (%) are reported as - with no in-text justification.
Table 4: Header uses % and ! as placeholders for β=0 and β=0.01, almost certainly a LaTeX macro / compilation error rather than deliberate misrepresentation.
Timeline and citation logic are plausible: 2024 ACM MM submission is consistent with a 2026 review date; TransMorph, TransMatch, and the cited grand-challenge link (https://grand-challenge.org/algorithms/corrfield/) are real and current.
Ablation study (Table 3) supports the architectural design of Shifted Embedding and Group Merging with a logical, iterative contribution pattern.

Evidence highlights

Table 1 (pp. 575–580): Manual cross-check of OASIS Dice column against author-supplied superscript ranks shows complete agreement.
Table 4 (p. 928): Header reads % represents β=0; ! represents β=0.01. Self-consistent (!) Dice for ShiftMorph = 0.8116, equal to Table 1.
Table 2 (p. 810): VoxelMorph++ cells for TRE opt. (mm) and Folds (%) are blank (-); no textual explanation provided in the manuscript body.
DOI: 10.1145/nnnnnnn.nnnnnnn (placeholder; not yet assigned).

Notes

Pixel-level image-reuse and splicing checks were not possible because raw figures were unavailable to the detection pipeline; any image-based concerns must be assessed by manual comparison of original source files.
The two editorial issues identified (Table 2 blanks, Table 4 symbols) should be flagged to the authors for revision but are not indicative of fraud.
A reviewer recommendation would be: request the authors (1) explain or fill in the missing VoxelMorph++ entries in Table 2, and (2) correct the LaTeX symbol rendering in Table 4.
No follow-up is warranted on PubPeer based on current findings.