Forensic Review: Cascaded Improved Neural Network for the Reconstruction, Classification, and Unmixing of the Raman Spectra of Mixed Microplastics (Analytical Chemistry, DOI: 10.1021/acs.analchem.5c04049)

Academic fraud report · Geng Detector

Summary

This automated forensic review evaluates the paper by Huang et al. in Analytical Chemistry proposing CSAM-ResUNet for Raman spectral processing of microplastics. The machine-forensic image analysis flagged eight figures (img-000, 002, 003, 004, 005, 006, 008, 009) with large numbers of same-offset copy-move block matches (e.g., img-003 at offset (32,0) with 116 pairs; img-008 at offset (0,32) with 149 pairs) accompanied by inconsistent local noise variance (img-008 CV=0.914). The paper does not describe a unified image-export pipeline, vector-format archival, or scanning-illumination gradients that would explain these signals, and the magnitude of matches exceeds what repeated axes/ticks would naturally produce. Six other findings (Tables 3-5 uniformity, Figure 3g-h simulated-data regularity, Figure 5h-j visual similarity, selective reporting of 35%/80% gains) were each reviewed and found to have plausible benign explanations grounded in the paper's stated methods (denoising pipeline, SNR-driven energy binning, distinctive PTFE Raman peaks, controlled simulation parameters). Timeline, equipment models, and open-source code were verified as consistent. Overall rating: highly suspicious (orange) primarily on the image-forensic evidence. Confidence is high on the machine-forensic signals but final judgment requires author-supplied raw figures and CSV data.

Verdict

Highly suspicious (orange) — primarily driven by machine-forensic image analysis (copy-move and noise-variance anomalies) across multiple figures. Several tabular and visual concerns were identified but have plausible benign explanations consistent with the paper's stated methodology. No definitive conclusion of misconduct is warranted without author-produced raw artifacts.

Key findings

Multiple figures show large counts of same-offset copy-move block matches: img-003 at offset (32,0) with 116 pairs; img-008 at offset (0,32) with 149 pairs; img-005 at offset (64,0) with 57 pairs; img-000 at offset (216,0) with 49 pairs; img-002, 004, 006, 009 also flagged.
Local noise variance inconsistency accompanies the copy-move signals (img-008 CV=0.914, the highest in the set).
No methodological justification in the manuscript for these signals: no mention of unified image-export pipeline, vector archival, or scanning/illumination gradients that could benignly produce hundreds of matched blocks at identical offsets across multiple figures.
Table 3 processed-column compression (e.g., PC processed range 0.9986→0.9602→0.9776, span 0.0384; raw span 0.6821) — partially explained by the paper's own denoising/baseline-correction pipeline (CSAM-ResUNet); MSE drops from 9.21×10⁻⁵ to 6.57×10⁻⁵ and R² rises to 0.9428 per Table 1.
Table 4 processed column clustered 82.51–99.68 across nine energy bins — consistent with SNR/energy-driven monotonic behavior, but processed values remain flatter than the raw column at the same energy (Dataset 8 raw=51.84 vs Dataset 3 raw=60.41 at 150 mJ).
Table 5 multiple 1.0000 precision/recall/F1 entries for PTFE-containing mixtures — physically plausible due to distinctive PTFE Raman features (~731/1382 cm⁻¹), but the frequency of perfect scores remains atypical.
Figure 3(g,h) scatter appears overly uniform — data are simulated via controlled generative process (formulas 5–11) with bounded parameters (λ∈(0.4,0.6), A∈(0,100), N∈[5,20]), so regularity is expected.
Figure 5(h,i,j) PE&PP spectra visually near-identical across three experimental conditions — caption describes distinct laser power/integration time settings; no raw spectra provided.
Figure 4 / abstract 35% PSNR and 80% SSIM gains reproduce Table 2 values (WTD+AirPLS PSNR=24.2648 vs CSAM-ResUNet PSNR=32.9637, ≈35.85%; SSIM 0.4372→0.8004, ≈83.1%); Figure S4 provides full noise-range comparison, mitigating cherry-picking concern.
Timeline and equipment consistency: submission 2025-07-03, revision 2026-01-27, acceptance 2026-02-26, publication 2026-03-09 — consistent. Equipment models (FC-D-785, RPB-D-A, NOVA2S) are real products. RTX 4070 Ti Super (2024 release) consistent with 2025 submission. NSFC grant numbers (42205148, 42305141) follow standard format. Code released at https://github.com/V1S10NAL/CSAM-ResUNet.

Evidence highlights

DOI: 10.1021/acs.analchem.5c04049
Copy-move detections (logLR=1.46 each, machine-forensic):
img-000: offset (216,0), 49 pairs
img-002: offset (32,0), 21 pairs
img-003: offset (32,0), 116 pairs
img-005: offset (64,0), 57 pairs
img-008: offset (0,32), 149 pairs
img-004, img-006, img-009 also flagged
Noise-variance anomaly: img-008 CV=0.914
Table 3 (PC row): processed 0.9986 / 0.9840 / 0.9602 / 0.9776; raw 0.9916 / 0.6911 / 0.4608 / 0.3095
Table 4 processed column: 99.68, 97.71, 96.78, 97.94, 97.60, 93.00, 97.43, 91.17, 82.51
Table 5: precision/recall/F1 = 1.0000 for PC&PTFE, PE&PTFE, PP&PTFE, PVC&PTFE, PMMA&PTFE
Table 2: PSNR 24.2648 → 32.9637; SSIM 0.4372 → 0.8004
Table 1: MSE 9.21×10⁻⁵ → 6.57×10⁻⁵; R² = 0.9428

Notes

The copy-move evidence across eight distinct figures is the principal concern and is not adequately explained by the manuscript's methodology. Benign explanations (repeated axes, ticks, gridlines in stacked spectra / confusion matrices / Grad-CAM heatmaps) do not plausibly account for 100+ matched pairs at uniform offsets.
All tabular concerns (Tables 3, 4, 5) have at least partial benign explanations grounded in the paper's described methods (denoising pipeline, SNR-driven energy binning, distinctive PTFE Raman peaks). These were downgraded to yellow but retained as residual concerns.
The automated Bayesian synthesis reports a posterior ≥99.9% and BF≈1.72×10¹⁰, but this assumes conditional independence among evidence items and may overstate joint strength given the correlated image-forensic signals.
Recommended follow-up: request raw vector files (.svg/.ai/.pdf) for img-002/003/005/008, raw per-spectrum CSVs for Tables 3–5, point-coordinate data for Figure 3(g,h) and Figure 4(g-j), independent raw spectra for Figure 5(h,i,j), and full training logs/seeds from the GitHub repository.
This report is AI-assisted and intended for academic-discussion purposes only; final misconduct determinations require investigation by qualified institutions.