Concerns Report: 'Lung Cancer Whole Slide Image Classification Method Based on Dual-Branch Mamba' (ACAIT 2025)

Academic fraud report · Geng Detector

Summary

This report evaluates the paper 'Lung Cancer Whole Slide Image Classification Method Based on Dual-Branch Mamba' (DOI: 10.1109/ACAIT67930.2025.11522027) presented at the 2025 Asian Conference on Artificial Intelligence Technology (ACAIT). The overall verdict is highly suspicious, though no direct biomedical image manipulation (e.g., Western blot duplication) was found. The work is characterized as a likely hastily assembled 'salami-slicing' or 'repackaging' publication. Key issues include: (1) retained template artifacts, notably an unwiped default IEEE LaTeX reference about magneto-optical media; (2) internal data contradictions where the text states an ACC of 88.74% while TABLE V shows 89.74%, and an F1-score (90.78%) that is mathematically inconsistent with the reported Precision (92.11%) and Recall (89.50%); (3) suspiciously monotonic ablation results that lack the variance typical of real deep learning training; and (4) a thinly justified 'innovation' that largely repackages a 2024 arXiv method (Mamba2MIL, arXiv:2408.15032). Confidence in the textual/numerical inconsistencies is high, while concerns about data fabrication remain circumstantial pending release of code and training logs.

Verdict

Highly suspicious. No direct image-based fraud (e.g., figure duplication) was detected, but the manuscript shows extensive signs of rushed, careless assembly and possible data fabrication. The paper was likely published in its current form without adequate proofreading or experimental rigor.

Key findings

Retained IEEE template reference: Reference [7] is an unwiped default example citation from the IEEE LaTeX template (*IEEE Transl. J. Magn. Japan*, vol. 2, pp. 740–741, 2021), strongly indicating copy-paste assembly.
Duplicate citation: References [2] and [9] are identical bibliographic entries, indicating sloppy reference management.
Text-table numerical conflict: Section III.D reports an ACC of 88.74% for the DBMC model, while TABLE V lists 89.74%.
Mathematically inconsistent F1-score: Given Precision = 92.11% and Recall = 89.50%, the correct F1-score is 2 × (0.9211 × 0.8950) / (0.9211 + 0.8950) = 90.76%, not the reported 90.78%.
Implausibly smooth ablation curves: TABLES II and III show ablation metrics (e.g., 89.87 / 86.51 / 85.54 / 85.32 / 85.42) that monotonically increase in suspiciously uniform increments (0.1–0.5), with no regressions or variance across runs.
Thin novelty / 'repackaging' concern: The core technique largely adapts the Mamba2 Structured State Space Duality (SSD) approach introduced in the 2024 preprint *Mamba2MIL* (arXiv:2408.15032, Aug. 2024), rebranded as 'Dual-Branch Mamba' with components MLDMamba and GSPMamba.

Evidence highlights

DOI: 10.1109/ACAIT67930.2025.11522027
Authors: Qi Zhang, Wei Jia, Defeng Kong, Xuefen Zhao
Venue: 2025 Asian Conference on Artificial Intelligence Technology (ACAIT)
Exact text excerpt: *"Specifically, it achieves an AUC of 94.03%, an ACC of 88.74%, and Precision, Recall, and F1-score values of 92.11%, 89.50%, and 90.78% respectively."*
TABLE V DBMC row ACC listed as 89.74%.
TABLE II base metrics: 89.87 / 86.51 / 85.54 / 85.32 / 85.42.
Default IEEE template reference: *Y. Yorozu, M. Hirano, K. Oka, and Y. Tagawa, 'Electron spectroscopy studies on magneto-optical media and plastic substrate interface,' IEEE Transl. J. Magn. Japan, vol. 2, pp. 740–741, 2021.*
Preprint cited: arXiv:2408.15032 (Aug. 2024).

Notes

Limitations: The analysis is based solely on the published PDF. No raw pathology images, code, or training logs were available, so claims about possible numerical fabrication remain circumstantial. Authors should be formally requested to release code and TensorBoard/WandB logs to verify the ablation and comparative results.
Image-based forensic analysis (e.g., Western blot duplication) was not applicable, as the paper provides no biological blot images and only minimal visualization.