Image integrity and textual analysis report for 'Multi-Scale Feature-Based Classification Method for Colorectal Pathological Images' (Han Zhang et al., ACAIT 2025)

Academic fraud report · Geng Detector

Summary

This report evaluates the ACAIT 2025 paper by Han Zhang, Xuefen Zhao, Wei Jia, and Defeng Kong (DOI: 10.1109/ACAIT67930.2025.11522031) and assigns a HIGHLY SUSPICIOUS verdict. The most concrete issue is on Page 1006: Equation (8) defines Recall = TP/(TP+FN), and Equation (9) defines Precision identically as TP/(TP+FN), omitting FP—a fatal copy-paste error that contradicts the standard textbook definition. Additional concerns include (1) a probable data-leakage setup: random seed(0) with a flat 6:2:2 split instead of patient-level partitioning, coinciding with an implausibly high 99.07% accuracy and 98.99% F1 on MedMF_colon, a 6.41-point jump over VMamba; (2) garbled ablation narrative on Page 1007 that inconsistently swaps DBMA/DBAM module names; (3) mathematically inconsistent metrics in Table I (MHIST, ∂=0.5), where Recall=78.72 and Precision=78.77 yield an F1 of 78.54, which is below the harmonic mean bound. Image-level forensics could not be performed. Confidence in textual/findings is high; confidence in coding or image integrity claims is limited pending source code and figures.

Verdict

🟠 Highly Suspicious. Multiple converging textual and numerical anomalies strongly suggest hasty, poorly proofread manuscript preparation. The Precision formula error is a hard, verifiable defect; the data-leakage, ablation-text, and F1-metric inconsistencies raise serious methodological doubts.

Key findings

Equation (9) copy-paste error (fatal, confirmed): Page 1006, Equation (9) defines Precision as TP / (TP + FN), identical to Equation (8) for Recall. The standard definition is TP / (TP + FP).
Probable data leakage in pathological image splits: Random seed(0) with a flat 6:2:2 split is described, with no mention of patient-level partitioning. MedMF_colon results (ACC 99.07%, F1 98.99%, +6.41 vs. VMamba) are implausibly high for histopathology without such controls.
Garbled ablation text: Page 1007 ablation paragraph mixes up DBMA and DBAM and likely intends MSGC + DBAM, indicating copy-paste assembly between drafts.
Metric self-inconsistency in Table I: MHIST row at ∂=0.5 reports Recall=78.72 and Precision=78.77 but F1=78.54, which is mathematically impossible since F1 is bounded by min(Recall, Precision) when both are positive.
Single-seed, no cross-validation: Only one random seed is reported, with no k-fold or repeated runs, making the headline 99.07% result statistically unverified.

Evidence highlights

DOI: 10.1109/ACAIT67930.2025.11522031
Paper title: *Multi-Scale Feature-Based Classification Method for Colorectal Pathological Images*
Authors: Han Zhang, Xuefen Zhao, Wei Jia, Defeng Kong
Venue: 2025 Asian Conference on Artificial Intelligence Technology (ACAIT)
Page 1006, Equation (8): Recall = TP / (TP + FN)
Page 1006, Equation (9): Precision = TP / (TP + FN) (should be (TP + FP))
Page 1006, Training details: "The random seed is fixed at seed(0). The data were divided into training, validation, and test sets at a ratio of 6:2:2."
Table III (MedMF_colon): ACC 99.07%, F1 98.99%, improvement over VMamba +6.41.
Page 1007 ablation paragraph: "Combining the MSDP and DBMA modules resulted in approximately 78.02% and 97.61%, while integrating MSDP and DBAM modules achieved approximately 78.66% and 97.65%." (module names appear swapped; Table II header suggests MSGC + DBAM)
Table I, MHIST, ∂=0.5: Recall 78.72, Precision 78.77, F1 78.54 (F1 < min(Recall, Precision), impossible).

Notes

Image-level forensic checks (e.g., pixel-level duplication, rotation/reuse, splicing) could not be performed because the original high-resolution figures were not available for analysis.
No patient-level split verification is possible without access to the authors' code and dataset partitioning scripts.
The Precision formula error alone warrants an erratum or formal clarification from the authors; combined with the metric inconsistency, it points to insufficient quality control rather than a single typographical slip.
All numeric values are reproduced exactly from the source report; no additional findings have been invented.
Final determination of misconduct requires investigation by the publisher or the authors' institution.