Geng Integrity 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: CLEAR (no integrity concerns identified). This is an integrity review of the anonymous double-blind submission 'ShiftMorph: A Fast and Robust Convolutional Neural Network for 3D Deformable Medical Image Registration' submitted to ACM MM 2024. Four investigation axes were examined: (1) numerical fabrication in Tables 1–2 via last-digit distribution and rank-order consistency checks; (2) image reuse/manipulation across Figures 1–7; (3) citation, hardware, and methodological timeline consistency; (4) statistical anomalies and p-hacking in the ablation study (Section 4.5, Tables 3–4). No red flags were triggered on any axis. Inference times in Table 1 (e.g., 189, 282, 134, 158, 680, 52, 39, 40, 127 ms for GTPP) show natural noise without suspicious clustering on 0/5 endings, and ranking superscripts match ordering. Figures are consistent with their captions. Hardware (Intel Xeon Silver 4314, RTX 3090 24G) and references up to 2024 are plausible. Limitations: pixel-level image forensics could not be performed because only extracted text and captions were reviewed.

Verdict

CLEAR — No evidence of data fabrication, image manipulation, citation anomalies, or statistical malpractice was found. The submission passes the four integrity checks performed.

Key findings

Numerical integrity of Tables 1–2 (runtime, Dice, HD95) is consistent: last-digit distribution shows natural noise; no clustering on 0 or 5; rank superscripts match value ordering.
Figures 1–7 are coherent with their captions and surrounding text; no obvious mislabeling or cross-figure inconsistency was detected.
Hardware (Intel Xeon Silver 4314 CPU, NVIDIA RTX 3090 24G) and reference list (including 2023–2024 works such as TransMatch [8] and Distill-SODA [42]) are temporally plausible for a 2024 submission.
Methodology is transparent, including explicit disclosure that the OASIS test set has no labels and that evaluation was performed on the validation set using corrfield-generated correspondences.
The ablation study (Section 4.5, Tables 3–4) honestly reports cases where self-consistency yields limited or inconsistent gains, contradicting a p-hacking pattern.

Evidence highlights

Table 1 (GTPP inference times in ms): values 189, 282, 134, 158, 680, 52, 39, 40, 127 — distributed without artificial round-number clustering; rankings (1–9) align with sorted magnitudes.
Section 4.1 (Implementation Details): explicit hardware specification (Xeon Silver 4314, RTX 3090 24G) consistent with 2024 practice.
Section 4.2 (Datasets): candid statement that the OASIS test set lacks labels and evaluation was performed on the validation set via corrfield.
Section 4.5 (Ablation): authors acknowledge mixed outcomes ("In most cases, except the red one, self-consistency leads to better evaluation scores"), which is atypical of cherry-picked ablations.
References: inclusion of recent works [8] (2023), [42] (2024), [44] (2024) demonstrates up-to-date coverage with no anachronistic citations.

Notes

This review is based on extracted text and figure captions only; pixel-level forensic analysis of Figures 1–7 was not possible. Authors remain anonymous due to the double-blind review stage.
DOI listed in the source material (10.1145/nnnnnnn.nnnnnnn) is a placeholder and not a real identifier.
The finding is a negative result (no fraud detected); absence of evidence here is not equivalent to absolute proof of integrity, but no anomaly surfaced under the four checks performed.
Confidence: moderate-to-high for numerical and statistical checks; lower for image-related checks due to the text-only input. The reviewer recommends standard peer review continue normally.