Geng Academic Fraud Report: Morphology-Guided Muscle Cell Detection & Counting (DOI: 10.1109/TCSVT.2026.3670914)

Academic fraud report · Geng Detector

Summary

Verdict: Highly suspicious (🟠). This IEEE TCSVT paper by Hongbing Qian, Meiyun Zuo and Lixi Zhao proposes a YOLO11-based muscle cell detector enhanced by transfer learning from blood-cell models, FFD augmentation, and a density-aware loss. Three major concerns are documented: (1) Implausible cross-domain transfer — the authors claim mAP50 jumps from 0.470 to 0.803 (a +33.3 point gain) when fine-tuning a blood-cell pretrained model on morphologically very different rat muscle histology images, violating typical transfer-learning expectations. (2) Overfitting/data leakage suspicion — the custom muscle dataset contains only 28 images total (8 test, ~20 train). The proposed YOLO11+TFD model reports mAP50 = 0.906, Precision = 0.885, Recall = 0.861 on 8 test images, an unusually high score for such a small, visually complex set, raising concerns of test-set leakage or result manipulation. (3) Internal inconsistency — the narrative claims 'substantial improvements in all indicators' yet Table III reportedly shows Precision deltas of −0.2% and −0.1%, suggesting data gaps or post-hoc fabrication. Pixel-level image forensics were not performed. These are indicators, not proven misconduct.

Verdict

🟠 Highly suspicious. Three substantive textual/data anomalies strongly question the validity of the reported results, and pixel-level image analysis could not be performed. No determination of fraud is made here; formal investigation by the journal or institution is recommended.

Key findings

Implausible cross-domain transfer gain. Table V (ablation) shows baseline YOLO11 mAP50 = 0.470 rising to 0.803 after transfer learning from blood-cell images to rat muscle histology — a +33.3 point jump across visually unrelated domains, which is inconsistent with typical transfer-learning behavior.
Tiny-dataset 'perfect' performance. Self-built muscle dataset contains only 28 images total (8 held out as test, ~20 used for training). On these 8 test images, the proposed YOLO11+TFD model achieves mAP50 = 0.906, Precision = 0.885, Recall = 0.861 (Table IV), an implausibly high score for a few-shot, dense, overlapping-cell detection task that warrants scrutiny for test-set leakage or result fabrication.
Text–table contradiction. The paper states that 'substantial improvements were observed in all indicators' when introducing the density-aware loss, but Table III reportedly shows Precision deltas of −0.2% and −0.1%, and many Δ values appear missing in the extracted text — consistent with data being retroactively assembled or a typesetting/OCR error masking worse results.
Image forensics not performed. Figures 1–7 could not be checked at the pixel level for duplication, splicing, or reuse because raw images were unavailable.

Evidence highlights

DOI: 10.1109/TCSVT.2026.3670914 (IEEE TCSVT, listed 2026).
Dataset size: 28 images total; 8 used as test set (Section IV.A).
Reported performance (Table IV): YOLO11+TFD, mAP50 = 0.906, Precision = 0.885, Recall = 0.861.
Reported ablation (Table V): Baseline mAP50 = 0.470 → with transfer learning mAP50 = 0.803 (Δ = +33.3 points).
Reported Table III anomalies: Precision deltas of −0.2% and −0.1% appearing alongside a narrative claim of uniform improvement across all metrics.

Notes

The blood-cell → muscle-cell transfer described is cross-domain in morphology (round/biconcave blood cells vs. tightly packed polygonal/elongated muscle fibers in H&E-stained sections), making a +33.3-point mAP50 gain anomalous.
FFD (free-form deformation) augmentation is a geometric transform that does not add new biological variability; it is unlikely to overcome a 20-image training bottleneck on its own.
A negative Δ for Precision in Table III directly contradicts the in-text claim of uniform improvement, which could reflect OCR/PDF extraction artifacts but, in a 2026 IEEE submission, is itself an editorial red flag.
Confidence in findings 1–3: moderate-to-high (textual/tabular evidence is internally consistent). Confidence in fraud intent: low — these are red flags, not proof of misconduct.
Recommended next steps: request raw labels and predictions for the 8 test images from the authors; verify dataset partitioning; check Figures 6–7 (density maps, prediction overlays) for image reuse or splicing; compare Table III source PDF against extracted text.
Limitations: no pixel-level image analysis, no author response, no independent replication attempted.