Geng Integrity Report: ShiftMorph — A Fast and Robust Convolutional Neural Network for 3D Deformable Medical Image Registration (Anonymous Submission, ACM MM 2024)

Academic fraud report · Geng Detector

Summary

This report evaluates the anonymous submission titled 'ShiftMorph: A Fast and Robust Convolutional Neural Network for 3D Deformable Medical Image Registration' for indicators of academic misconduct. The assessment covered six dimensions: image reuse, data fabrication, image splicing, statistical anomalies, publication productivity, and citation/methodological soundness. No anomalies were identified in any category. The paper is purely algorithmic/computational, so traditional biological image manipulation risks (Western blot, microscopy, flow cytometry) are not applicable. Tabulated performance metrics (Dice, HD95, TRE, Folds %) on OASIS, IXI, and Lung250M-4B exhibit realistic variance consistent with published baselines (e.g., TransMorph, VoxelMorph). Figures are code-generated visualizations or architecture diagrams with no splicing artifacts. Hardware specifications (Intel Xeon Silver 4314, 24GB NVIDIA RTX 3090) and references from 2021–2023 are internally consistent. Verdict: cleared / no evidence of misconduct. Limitations: confidence is moderate because journal-side peer review materials were not independently verified; the DOI is a placeholder. Final determination requires institutional investigation.

Verdict

Cleared — No evidence of academic misconduct detected.

Across six analytical categories (image reuse, data fabrication, image splicing, statistical anomalies, publication productivity, citations/methodology), no irregularities were identified. Recommended actions: no author contact, no PubPeer post, no journal or institutional report required at this time.

DOI: 10.1145/nnnnnnn.nnnnnnn (placeholder, not officially assigned)

Key findings

Image reuse: Not applicable in the traditional biology sense. All figures are algorithm-style architecture diagrams (Fig. 1, Fig. 2), conceptual illustrations (Fig. 3), performance plots (Fig. 4, Fig. 7), or qualitative registration examples (Fig. 5, Fig. 6). No visual duplication detected.
Data fabrication: Tabulated results in Tables 1–4 show realistic dispersion. Example: OASIS Dice scores range from 0.7947 to 0.8266 across compared methods, with natural leading-digit distribution and no suspicious arithmetic progression.
Image splicing: No splicing artifacts. Most plots are vector outputs from plotting libraries; no biomedical gel/blot images are present.
Statistical anomalies: Reported gains (~3× speedup with comparable accuracy) align with plausible architectural trade-offs in deformable image registration. No p-value tampering because the paper uses standard deep-learning benchmarking rather than biological hypothesis testing.
Productivity anomaly: Single, focused methodological contribution on 3D DIR; no signs of mass-produced "paper-mill" output.
Citations / methodology: Architecture grounded in explicit equations (Eqs. 1–10). Hardware (Intel Xeon Silver 4314 CPU, 24GB NVIDIA RTX 3090 GPU) and 2021–2023 references are temporally coherent with an iterative research timeline.

Evidence highlights

Quantitative anchor: OASIS Dice range 0.7947–0.8266 across baselines including TransMorph and VoxelMorph.
Hardware disclosure: Intel Xeon Silver 4314 + 24GB NVIDIA GeForce RTX 3090.
Datasets referenced: OASIS, IXI, Lung250M-4B.
All four inspection tables (Tables 1–4) audited; metrics include Dice, HD95, TRE, Folds %, and runtime.
Six figures (Fig. 1–7 inclusive of sub-figures) inspected for reuse and splicing; none flagged.

Notes

DOI status: 10.1145/nnnnnnn.nnnnnnn is a placeholder identifier associated with the source PDF filename; no canonical ACM DOI has been assigned at the time of review.
Submission context: Pre-print / under-review version for ACM Multimedia 2024 with anonymous authorship; formal peer-review artifacts were not available for cross-checking.
Limitations: This is an AI-assisted screening. The verdict rests on textual, tabular, and figure-level heuristics. Failure modes (false negatives) cannot be excluded, particularly for issues beyond the scope of automated inspection.
Disclaimer: Final academic-misconduct determinations require adjudication by the relevant institutional or editorial bodies.