Investigation Report: "Coupled Diffusion Models for Metal Artifact Reduction of Clinical Dental CBCT Images" (IEEE TMI, 2025)

Academic fraud report · Geng Detector

Summary

This report assesses two serious allegations of image manipulation against the paper DOI 10.1109/TMI.2025.3587131 by Zhang et al., while clearing the paper on timeline and citation integrity. The first concern is that Figure 5 (t-SNE visualization) appears to show suspiciously regular parallel-line grid patterns, which is highly atypical for t-SNE projections of real high-dimensional features; such geometry is more consistent with procedurally generated points or non-real data. The second concern is that background noise and texture patterns in the zoom-in insets of Figure 6 appear pixel-identical across results from fundamentally different algorithms (NMAR, Restormer, InDuDoNet+), suggesting shared background compositing rather than independent inference. Both allegations are marked as unverified pending inspection of the original high-resolution figures. The overall verdict is "Questionable." Timeline, hardware (NVIDIA A100 40GB), and 2024 references are consistent. Confidence is limited because the review is text-based only.

Verdict

🟡 Questionable. Two severe image-integrity allegations are flagged but require visual verification of the original high-resolution figures before any formal conclusion of misconduct.

Key findings

⚠️ Finding 1 — Possible fabricated t-SNE visualization in Figure 5 (Page 6). The reported scatter pattern resembles a regular oblique-line grid rather than the irregular cloud-like clusters expected from t-SNE of genuine neural features. Severity: 🔴. Verification status: ⚠️ Unverified — text-only review cannot confirm pixel-level geometry; the paper does describe a t-SNE analysis.
⚠️ Finding 2 — Possible background compositing / image reuse in Figure 6 (Page 9). Background soft-tissue noise and grayscale gradients in the red-box zoom-ins allegedly appear pixel-identical across qualitatively different methods (NMAR, Restormer, InDuDoNet+), which is implausible for independent model outputs. Severity: 🔴. Verification status: ⚠️ Unverified — text confirms Figure 6 is a comparison panel, but visual confirmation is required.
✅ Finding 3 — Timeline, hardware, and citations are internally consistent. Submitted 2025; published July 2025; references include 2024 works ([27], [28], [38], [47]); reported hardware is NVIDIA A100 (40GB). No anachronisms detected.

Evidence highlights

DOI: 10.1109/TMI.2025.3587131
Journal: IEEE Transactions on Medical Imaging, Vol. 44, No. 12, December 2025
Authors: Zhouzhuo Zhang, Juncheng Yan, Yuxuan Shi, Zhiming Cui, Jun Xu, Dinggang Shen
Figure 5 alleged anomaly: parallel-line / grid structure inconsistent with real t-SNE projection; gray points in (b) reported as linear streaks.
Figure 6 alleged anomaly: identical background noise patterns across heterogeneous algorithms in the red-box regions.
Timeline check passed: 2024 references present, A100 40GB hardware plausible.

Notes

Both primary allegations rely on visual judgment of figure pixels and cannot be conclusively confirmed from text alone. The report explicitly flags them as "依据不足" (insufficient evidence pending image inspection).
No textual evidence of data fabrication, statistical manipulation, or plagiarism was identified in this review pass.
Recommended follow-up: request raw high-dimensional embeddings and plotting scripts for Figure 5; request original inference outputs for Figure 6; post concerns on PubPeer; notify the journal editorial office and the authors' institutions (Nankai University; ShanghaiTech University).
This report is AI-assisted and intended for academic discussion only; final determination of misconduct requires institutional investigation.