Critical flaws in 'A deep learning framework based on structured space model for detecting small objects in complex underwater environments' (Communications Engineering, 2025)

Academic fraud report · Geng Detector

Summary

This report flags multiple serious and verifiable mathematical/formula errors in the paper (DOI: 10.1038/s44172-025-00367-9) by Zhuang et al., published in Communications Engineering in 2025. The verdict is 'highly suspicious.' Confirmed findings include a wrong recall formula (written as TP/(TN+FN) instead of the correct TP/(TP+FN)) and contradictory percentage claims about parameter reduction (claiming a 390% reduction, which is mathematically impossible for absolute quantities and self-contradictory between abstract and Discussion). Two further concerns—alleged background-image reuse in Figure 4c and suspiciously uniform AP50 bar heights across heterogeneous models in Figure 6a—are flagged as plausible but insufficiently evidenced from the text alone. Overall, the systematic failure in basic math undermines confidence in all quantitative performance claims. Limitations: image and bar-chart suspicions require higher-resolution visual inspection; final determination of misconduct requires institutional investigation.

Verdict

🟠 Highly suspicious. The paper contains clear, reproducible mathematical and logical errors that are uncharacteristic of rigorous peer-reviewed work, and additional concerns about data presentation could not be independently confirmed from the published text alone.

Key findings

Wrong recall formula (Confirmed): Equation (4) defines Recall = TP / (TN + FN), which is mathematically incorrect; the standard definition is TP / (TP + FN). The presence of this error in a core evaluation metric is a serious flaw.
Impossible parameter-reduction claim (Confirmed): The abstract and Discussion state reductions of "40%, 67%, 21%, 41%, 227%, and 390%" relative to mainstream detectors. A reduction exceeding 100% in absolute parameter count is mathematically impossible. Numerical check: 32.66M vs 6.67M indicates RT-DETR is ~389% larger than UWNet, not that UWNet is "390% smaller."
Self-contradictory percentage logic (Confirmed): The text simultaneously states UWNet has "390% fewer parameters than RT-DETR" and that RT-DETR "has 32.66 million parameters—390% more than UWNet," exposing a systematic collapse in percentage-change logic that invalidates all percentage-based performance statements.
Suspected background-image reuse in Figure 4c (Insufficient evidence): Visual analysis suggests the third and fourth test-scene columns share the same seabed background, but the paper provides no textual acknowledgment and high-resolution confirmation is lacking.
Suspected fabricated uniformity in Figure 6a (Insufficient evidence): AP50 values across YOLOv7s, RT-DETR, YOLOv8s and others on the DUO echinus class appear near-identical (~0.93–0.94), but the paper text does not list the exact competitor values needed to substantiate the claim.

Evidence highlights

Equation (4) in the Methods section: Recall = TP / (TN + FN) — incorrect denominator.
Abstract / Discussion quoted percentages: "40%, 67%, 21%, 41%, 227%, and 390%."
Numerical cross-check: 32.66M (RT-DETR) vs 6.67M (UWNet) → difference ≈ 389% of UWNet, i.e., RT-DETR ≈ 4.89× UWNet, not a 390% reduction.
DOI: 10.1038/s44172-025-00367-9

Notes

Confirmed errors alone justify a formal erratum or correction request at minimum.
The two visual concerns (Figure 4c image reuse; Figure 6a uniform bars) should be re-examined against the published high-resolution figures and supplementary material.
Final determination of academic misconduct (data fabrication vs. mere sloppy writing) requires institutional investigation and access to raw experimental data.
Report limitations: AI-assisted analysis; image-based claims not fully verifiable from text; possible false positives.