Integrity Review: Cascaded Improved Neural Network for Reconstruction, Classification, and Unmixing of Raman Spectra of Mixed Microplastics

Academic fraud report · Geng Detector

Summary

Verdict: CLEAN (no substantive integrity concerns identified). This review examines the article by Huang et al. in Analytical Chemistry (DOI: 10.1021/acs.analchem.5c04049), a deep-learning/Raman-spectroscopy algorithm paper. Key checks: (1) image reuse — figures could not be subjected to pixel-level analysis because high-resolution originals were not provided, but captions and experimental conditions are internally consistent (no mismatched content observed); (2) data plausibility — reported metrics (e.g., R² up to 0.9996, MSE on the order of 10^-7) are appropriate because models were trained and tested on simulated spectra, while real-sample tests at degraded conditions (e.g., 8.97%–82.51% accuracy at 50 mJ) include deliberately imperfect results, which argues against fabrication; (3) statistical anomalies — none, the paper relies on ML/computer-vision metrics rather than inferential tests; (4) timeline/methodology — hardware (RTX 4070 Ti Super, released early 2024) and cited 2026 Talanta reference are consistent with publication timing. Optional follow-up: reproduce from the public GitHub repository. Limits: AI-assisted review, no pixel-level image inspection, no laboratory auditing.

Verdict

CLEAN. No credible evidence of academic fraud was detected. The paper appears methodologically self-consistent, and observed "imperfect" real-sample results actually increase confidence in the integrity of the data.

Key findings

No indication of image duplication or reuse; pixel-level image forensics were not possible due to absence of high-resolution source files, but figure–caption alignment is consistent.
Performance metrics on simulated spectra (Table 6, R² = 0.9996; MSE ≈ 10^-7) are reasonable for synthetic-dataset ML benchmarks.
Real-sample evaluation at low laser energy (50 mJ) reports markedly degraded accuracy (Raw 8.97%; up to 82.51% with processing), which is inconsistent with a fabricated, uniformly perfect dataset.
No inferential statistical tests (p-values, ANOVA) are used or misused; evaluation relies on standard computer-vision/ML metrics (SNR, PSNR, SSIM, F1-score).
Hardware (NVIDIA RTX 4070 Ti Super) is temporally compatible with 2024–2025 model training; cited 2026 reference (Talanta) is plausible given current publication timeline.
Reported conflicts are minor and non-indicative of fraud.

Evidence highlights

DOI: 10.1021/acs.analchem.5c04049
Submission: 2025-07; Publication: 2026-03.
Table 6: R² = 0.9996, MSE on the order of 10^-7 (simulated spectra training/testing, explicitly stated by authors).
Table 4 / Dataset 9 real-sample accuracy at 50 mJ: 8.97% (Raw), up to 82.51% (with model).
Training hardware: NVIDIA GeForce RTX 4070 Ti Super (released early 2024).
Public code repository referenced for optional reproduction: https://github.com/V1S10NAL/CSAM-ResUNet

Notes

This is an AI-assisted screening report; it is not an institutional investigation.
Image-level verification was limited by the lack of original high-resolution figures; such checks are recommended if formal review is required.
Final responsibility for any integrity determination rests with the relevant institutional or publisher process; appeals should follow official channels.
Both false positives and false negatives are possible.