Geng 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: Questionable but largely unsupported by current evidence (🟡). This Analytical Chemistry paper proposing a cascaded CSAM-ResUNet architecture for Raman spectral unmixing and classification of mixed microplastics was examined across image integrity, data/logic consistency, and timeline/equipment provenance. No figures suitable for pixel-level duplication analysis were available (the paper contains only neural-network schematics and Raman spectral plots), so image-reuse checks remain inconclusive. Conversely, numerical self-consistency was strong: macro-average precision and recall in Table 5 reproduce the stated F1-score of 0.9680, and the abstract's '32% further reduction in MSE' was independently verified by back-calculation from Table 1 (CSAM-ResUNet gain ≈1.98 vs SE-ResUNet gain ≈1.49). Timeline checks also passed: the cited RTX 4070 Ti Super GPU (released Jan 2024) is consistent with a 2025 submission, and the TensorFlow 2.10 / Python 3.8 stack is plausible. Confidence is moderate; limits include no pixel-level image evidence and no access to raw spectra or GitHub code.

Verdict

🟡 Questionable — no clear fraud indicators detected, but pixel-level image analysis was not possible due to insufficient raw figures. Textual, statistical, and provenance checks all passed.

Key findings

Image integrity: Could not be verified at pixel level; paper contains no Western blots or microscopy images typical of manipulation. Raman spectra and architecture diagrams show no obvious duplication.
Statistical consistency: Macro-average precision (0.9727), recall (0.9678), and accuracy (0.9678) in Table 5 reproduce the stated F1-score of 0.9680 exactly.
Abstract claim verification: The '32% further reduction in MSE' wording is convoluted but mathematically correct — CSAM-ResUNet MSE reduction (~1.98) exceeds SE-ResUNet reduction (~1.49) by approximately 32%.
Combinatorial logic: The Methods section correctly states C(7,2) = 21 binary mixtures, consistent with the 21-class classification target.
Hardware timeline: RTX 4070 Ti Super (released Jan 2024) is consistent with a July 2025 submission and March 2026 publication.
Software stack: Python 3.8 with TensorFlow 2.10 is plausible (TF 2.10 was the last native-Windows-GPU version, after which WSL2 is required).

Evidence highlights

DOI: 10.1021/acs.analchem.5c04049
Table 5 confusion-matrix macro metrics: Precision = 0.9727, Recall = 0.9678, Accuracy = 0.9678, F1 = 0.9680 (under 150 mW / 1000 ms condition).
Table 1 MSE comparison: ResUNet = 9.21, SE-ResUNet = 7.72, CSAM-ResUNet = 6.57; incremental gains 1.49 and 2.64 respectively, with the latter being ~1.32× the former.
GitHub repository referenced: https://github.com/V1S10NAL/CSAM-ResUNet (not independently audited).
Authors: Weixiang Huang, Jiajin Chen, Hao Xiong, Ligang Shao, Guishi Wang, Kun Liu, Chilai Chen, Xiaoming Gao.

Notes

Pixel-level image analysis was not performed; no raw TIFFs or high-resolution supplementary figures were available to the investigator.
Verification of code reproducibility via the public GitHub repository is recommended but was not carried out in this assessment.
The convoluted abstract phrasing around percentage reductions is stylistic rather than fraudulent, though clearer language would benefit readers and reviewers.
All numerical checks were performed by the AI assistant; independent reproduction by a qualified statistician or spectroscopist is advisable before any formal action.