Image Reuse and Suspected Cross-Modality Input Fabrication in "Structure-Driven Unsupervised Domain Adaptation for Cross-Modality Cardiac Segmentation" (IEEE TMI, 2021)

Academic fraud report · Geng Detector

Summary

This report examines two concerns raised about a 2021 IEEE Transactions on Medical Imaging paper (DOI: 10.1109/TMI.2021.3090432) by Cui et al. on unsupervised cross-modality cardiac segmentation. The first concern is that Figure 4 (page 3611), which claims to display both MRI→CT and CT→MRI adaptation directions, may contain reused images where only grayscale/contrast transformations were applied, rather than genuine cross-modality inputs. The second concern argues that, if the Figure 4 inputs are fabricated, the quantitative results in Table I lose their experimental basis. Verdict: Questionable (yellow flag). The visual-analysis summary provides circumstantial hints, but no textual evidence in the paper directly corroborates fabrication, and the source pixel data cannot be verified from the PDF alone. Both findings are therefore classified as "insufficient evidence" rather than confirmed misconduct. Confidence in the flagged anomaly is moderate-low; the conclusions rest on visual inspection only. Recommended next steps include requesting raw DICOM data from the authors and soliciting independent image-forensic review on PubPeer.

Verdict

Questionable (🟡) — Figure 4 exhibits visually suspicious cross-modality inputs that may be reuse with grayscale/contrast transformation, but no textual or pixel-level evidence in the PDF confirms fabrication. Both flagged issues remain unverified.

Key findings

Figure 4 (Page 3611) — Suspected cross-modality input fabrication: Images labeled as MRI→CT and CT→MRI adaptation may be the same anatomical slice with only linear grayscale/contrast adjustment, per a visual-analysis summary. The caption states "Qualitative results of different methods for unsupervised MRI to CT domain adaptation (top three rows) and CT to MRI domain adaptation (bottom three rows)," which describes the layout but does not disprove the visual anomaly.
Experimental-logic risk: If Finding 1 holds, the quantitative comparisons in Table I and the claim "Our method achieves the best performance in both directions by a large margin" (Section IV.C) would lose physical grounding. However, this conclusion depends entirely on Finding 1, which is itself unverified.
Severity rating: Both issues marked 🔴 by the original reporter, but reclassified to ⚠️ Insufficient Evidence after textual cross-check.

Evidence highlights

Figure 4 caption text explicitly distinguishes "top three rows" (MRI→CT) and "bottom three rows" (CT→MRI), consistent with bidirectional qualitative comparison.
No paragraph in the paper text explicitly describes image acquisition, preprocessing pipelines, or DICOM provenance for the specific slices in Figure 4 that would allow textual corroboration.
DOI: 10.1109/TMI.2021.3090432; journal: IEEE Transactions on Medical Imaging, Vol. 40, No. 12, December 2021.
Authors of record: Zhiming Cui, Changjian Li, Zhixu Du, Nenglun Chen, Guodong Wei, Runnan Chen, Lei Yang, Dinggang Shen, Wenping Wang.

Notes

Limitations: This assessment is based solely on the PDF and a visual-analysis summary; no raw pixel-level forensic comparison (e.g., ELA, PRNU, DCT analysis) was performed by the reviewer.
Recommended follow-up: (1) Request original DICOM data for Figure 4 panels from the corresponding author; (2) solicit independent image-forensic review via PubPeer; (3) escalate to the journal editorial office only if independent pixel-level forgery evidence is obtained.
Disclaimer: This AI-assisted report is for academic discussion only and does not constitute a formal finding of misconduct.