Geng Integrity Report: ShiftMorph — A Fast and Robust Convolutional Neural Network for 3D Deformable Medical Image Registration

Academic fraud report · Geng Detector

Summary

This report assesses 'ShiftMorph: A Fast and Robust Convolutional Neural Network for 3D Deformable Medical Image Registration' (ACM MM 2024 submission, DOI: 10.1145/nnnnnnn.nnnnnnn) for academic integrity. The overall verdict is CLEAN (清白). No credible evidence of data fabrication, image manipulation, or methodological misconduct was found. Key positive indicators include: (1) unusually high numerical consistency between the main results in Table 1 and the ablation study in Table 4, suggesting genuine experiments rather than independently fabricated figures; (2) mathematically self-consistent derivations in the smoothing regularizer (Eq. 8), demonstrating genuine understanding rather than AI-generated padding; and (3) an unusually honest disclosure by the authors that their 'lossless downsampling' module is in fact 'pseudo-downsampling' that does not reduce computational complexity. One minor cosmetic concern was noted: ShiftMorph-diff reports a Folds (%) of exactly 0.0000 on both Validation and Testing sets in Table 2, which is theoretically plausible given diffeomorphic guarantees from the Squaring and Scaling step but appears slightly 'too clean.' Limitations: image-based checks (e.g., Western Blot duplication) are not applicable to this deep-learning paper, and high-resolution figures were unavailable for pixel-level analysis.

Verdict

Clean (清白). No substantive evidence of academic fraud, data fabrication, image manipulation, or methodological misconduct was identified. The paper passes the integrity checks performed.

Key findings

Strong cross-table numerical consistency (positive signal). Table 4 ablation entries for the ShiftMorph baseline with Self-consistency (λ=0.01) match Table 1 exactly: Dice 0.8116, HD95 2.1124, Folds 0.7786%. ShiftMorph×3 also shows close agreement (0.8266 / 1.9804 / 0.5052 vs 0.8266 / 1.9771 / 0.4736), consistent with genuine experimental runs rather than independently fabricated tables.
Mathematically rigorous derivations (positive signal). The smoothing regularizer in Section 3.7, Eq. (8), correctly expresses the diffusion penalty as ‖∇ₓφ − 1‖₂². Because φ = id + u, this is precisely ∇ₓu, indicating the authors understand the underlying calculus rather than copying formulas mechanically.
Transparent, honest module description (positive signal). Section 3.1 explicitly admits that the proposed 'lossless downsampling' is actually 'pseudo-downsampling' that trades spatial resolution for batch size without reducing computational complexity, before introducing the genuine Group Merging strategy that does reduce complexity.
Minor cosmetic anomaly (negligible). In Table 2 (Supervised Lung CT Image Registration), ShiftMorph-diff reports Folds (%) of exactly 0.0000 on both Validation and Testing splits. This is theoretically defensible under the Squaring and Scaling (SS) diffeomorphism guarantee (possibly due to floating-point underflow or threshold truncation), but the four-decimal precision uniformity is slightly 'too clean.'
Text-extraction artifacts are not a paper defect. Symbols such as !!"#$%& near Figures 1 and 2 are rendering artifacts from the PDF/text extraction tool applied to special fonts and embedded architecture-diagram labels, not errors in the original manuscript.

Evidence highlights

Table 1 vs. Table 4 (Dice / HD95 / Folds):
ShiftMorph baseline, λ=0.01: 0.8116 / 2.1124 / 0.7786% — identical across both tables.
ShiftMorph×3: 0.8266 / 1.9804 / 0.5052% (Table 4) vs. 0.8266 / 1.9771 / 0.4736% (Table 1).
Eq. (8), Section 3.7: penalty term ‖∇ₓφ − 1‖₂² is the correct equivalent of ∇ₓu under φ = id + u.
Section 3.1: explicit admission that 'The complexity of network computation remains unchanged' for the pseudo-downsampling module.
Table 2: ShiftMorph-diff Folds (%) = 0.0000 on both Validation and Testing splits.
DOI: 10.1145/nnnnnnn.nnnnnnn (placeholder; formal DOI not yet assigned at submission time).

Notes

The paper concerns 3D deformable medical image registration using deep learning; it contains no biological experimental imagery (e.g., Western blots, microscopy), so image-duplication checks of that kind are not applicable.
High-resolution original figures were not available, preventing pixel-level forensic analysis of the architecture diagrams in Figures 1 and 2.
The verdict of 'clean' is based on the evidence available in the supplied PDF and standard integrity heuristics; it is not a formal institutional investigation.
Recommended action: treat as a methodological reference and consider reproducing its Self-consistency component; no further escalation is warranted.