Integrity Review: Improving the classification performance of microplastics by noise reduction and baseline correction of Raman spectra with a neural network-based algorithm (DOI: 10.1364/OE.597337)

Academic fraud report · Geng Detector

Summary

This review evaluates a 2026 Optics Express paper by Chen et al. that proposes an SE-ResUNet for denoising and baseline correction of microplastic Raman spectra, then classifies six polymer types. The overall verdict is highly suspicious. The most serious concern is a physical inconsistency in Table 3: Dataset 4 (29.63 mJ laser energy, 1000 ms integration) yields a post-processing Mean SNR of 1446.07, while Dataset 5, which uses clearly lower excitation energy (22.22 mJ) and shorter integration (750 ms), paradoxically reports a higher Mean SNR of 1857.15. This inverse relationship between input signal quality and output SNR contradicts basic photon statistics and signal-processing theory, suggesting possible fabrication or a calculation/transcription error. A second concern involves an internal narrative contradiction: Section 3.3 states that ~30 mW and 1000 ms is the lower stability limit, yet Section 3.4 showcases Dataset 5 (below that limit) as a high-accuracy benchmark at 96.90%. Additional minor issues include a missing right bound in the variable N range of Eq. (4), and a publication pattern of closely related network variants in 2025–2026 suggesting incremental output. Confidence is moderate pending inspection of raw spectra and SNR computation code.

Verdict

🟠 Highly suspicious. Primary concerns are (1) a physically implausible SNR inversion between Datasets 4 and 5 in Table 3, and (2) an internal contradiction between the stated operational limits (Section 3.3) and the highlighted benchmark result (Section 3.4 / Table 4). Findings 3 and 4 are secondary indicators of editorial or methodological sloppiness rather than proven misconduct.

Key findings

Physically implausible SNR inversion (Table 3, p. 17249): Dataset 4 (29.63 mW, 1000 ms, 29.63 mJ) → post-SE-ResUNet Mean SNR = 1446.07. Dataset 5 (29.63 mW, 750 ms, 22.22 mJ) → post-SE-ResUNet Mean SNR = 1857.15. Lower excitation energy and shorter integration producing ~28% higher SNR after denoising contradicts photon-counting statistics and conventional signal-processing behavior.
Internal methodological contradiction (Section 3.3 vs. Section 3.4 / Table 4): Section 3.3 asserts the effective stability threshold is ~30 mW and 1000 ms integration. Table 4 nevertheless presents Dataset 5 (29.63 mW, 750 ms) as a flagship result with 96.90% classification accuracy, while Section 3.3 acknowledges that shorter integrations introduce artifacts and degraded reconstruction.
Publication pattern suggestive of incremental output (References 26 & 27): Multiple 2025–2026 papers from the same group apply sequentially renamed neural-network variants (Improved NN → Cascaded Improved NN → SE-ResUNet) to the same microplastic Raman dataset (PC, PE, PET, PP, PS, PVC), raising concerns about data reuse and "salami-style" publication.
Manuscript typo in Eq. (4) definition (p. 17242): The range for N is written as "randomly selected within the range of [5, ]" with the upper bound missing.

Evidence highlights

DOI: 10.1364/OE.597337 (Optics Express, Vol. 34, No. 9, 4 May 2026, oe-34-9-17239.pdf)
Table 3 (p. 17249): Dataset 4 Mean SNR = 1446.07 at 29.63 mJ; Dataset 5 Mean SNR = 1857.15 at 22.22 mJ.
Section 3.3 (p. 17248) quote: "The results indicate that the lower limit for maintaining effective and stable operation of this method is approximately 30 mW of laser power and an integration time of 1000 ms."
Table 4 (p. 17250): Dataset 5 classification accuracy reported as 96.90% despite being below the stated operating limit.
Equation (4) variable definition (p. 17242): N ∈ [5, ] with no upper bound supplied.
References 26 (2026) and 27 (2025) document rapid back-to-back publication of related network-architecture variants on the same dataset.

Notes

The SNR anomaly could in principle stem from (a) different number of averaged spectra per dataset, (b) a non-standard SNR definition, or (c) post-processing artifacts; only inspection of the raw spectra and the SNR computation script can discriminate these possibilities from outright fabrication.
The classification-vs-limit contradiction may reflect selective reporting (highlighting only the favorable subset of Dataset 5 spectra), but the paper does not state that the 96.90% result is restricted to a filtered subset.
The findings regarding publication cadence are circumstantial and do not, by themselves, constitute evidence of misconduct.
Recommended follow-up: request raw spectra and SNR/accuracy code from the authors; raise the specific Table 3 inversion and Section 3.3/3.4 contradiction on PubPeer; notify the Optics Express editorial office. Institutional investigation is optional pending author response.