Investigation Report: 'Cascaded Improved Neural Network for the Reconstruction, Classification, and Unmixing of the Raman Spectra of Mixed Microplastics' (Anal. Chem., DOI: 10.1021/acs.analchem.5c04049)

Academic fraud report · Geng Detector

Summary

Verdict: Highly suspicious (🟠). The report identifies two confirmed issues and one flag requiring further verification. (1) Statistical/marketing misstatement: the abstract claims a '32% further reduction in mean squared error' for CSAM-ResUNet, but direct calculation from Table 1 yields only ~21.5% (vs. ResUNet, MSE 7.23E-5 vs. 9.21E-5) and ~28.6% for the best variant CSAM-ResUNet-DHPL (6.57E-5). The 32% figure is reconstructed by comparing percentage point differences between successive relative reductions (21.50% − 16.18%), i.e., a second-order relative change, misrepresented as a direct MSE reduction. (2) Methodological concern: under extreme low-dose conditions (50 mW, 1 s, ~50 mJ total energy; Table 4 Dataset 9), raw classification accuracy is 8.97% (near chance) yet reconstructed spectra reach 82.51%, raising the possibility of feature hallucination rather than genuine denoising. (3) Unverified flag: the cited code repository (github.com/V1S10NAL/CSAM-ResUNet) could not be live-checked; ownership and contents are uncertain. Confidence in findings 1–2 is high based on the paper's own numbers; finding 3 is precautionary.

Verdict

🟠 Highly suspicious. Two substantive issues are confirmed from the published data; one additional concern (code availability) could not be verified and remains a flag.

Key findings

Misrepresented performance metric (confirmed): The abstract headline figure of a '32% reduction in MSE' is not reproducible from Table 1. It is a second-order calculation, not a direct percentage reduction.
Potential feature hallucination under extreme low-SNR conditions (confirmed): Classification accuracy jumps from 8.97% to 82.51% on Dataset 9 (50 mJ total laser dose), a regime where Raman features are plausibly noise-dominated.
Unverified open-source claim (flag): The cited GitHub repository https://github.com/V1S10NAL/CSAM-ResUNet could not be confirmed; account ownership relative to the authors' institutions (USTC / Hefei Institutes of Physical Science, CAS) is unclear.

Evidence highlights

Table 1 (MSE values, reported by the authors):
ResUNet baseline: 9.21 × 10⁻⁵
SE-ResUNet: 7.72 × 10⁻⁵ (≈ 16.18% reduction vs. ResUNet)
CSAM-ResUNet-MSE: 7.23 × 10⁻⁵ (≈ 21.5% reduction vs. ResUNet; ≈ 6.3% reduction vs. SE-ResUNet)
CSAM-ResUNet-DHPL (best): 6.57 × 10⁻⁵ (≈ 28.6% reduction vs. ResUNet)
Reconstruction of the '32%' claim: (21.50% − 16.18%) / 16.18% ≈ 32.8%. The authors appear to have reported the relative change between two relative reductions as if it were a direct MSE reduction, which is misleading.
Table 4, Dataset 9 (50 mW × 1 s, ≈ 50 mJ total): raw classification accuracy 8.97%; post-reconstruction accuracy 82.51%.
DOI preserved as reported: 10.1021/acs.analchem.5c04049.

Notes

Finding 1 is verified purely by arithmetic on the authors' own Table 1; no external data is required.
Finding 2 is not a definitive accusation of hallucination but a methodological flag consistent with over-reconstruction/overfitting in low-SNR regimes. Independent reproduction with blinded spectra is needed for a stronger conclusion.
Finding 3 is a precautionary flag; the repository link should be checked manually. A 404 page, an unrelated upstream project, or a near-empty fork would materially change its severity.
No additional forms of image manipulation, duplicate publication, or authorship anomalies are identified in this report.
This assessment is based solely on the publicly available text; it does not constitute a formal determination of misconduct.