Integrity Review Report: 'Breast Cancer Whole Slide Image Classification Method Based on Glandular Discriminative Features' (DOI: 10.1109/ACAIT67930.2025.11522617)

Academic fraud report · Geng Detector

Summary

This review assesses a 2025 IEEE ACAIT conference paper proposing a glandular-discriminative-feature MIL method for breast cancer whole-slide-image classification. The verdict is Highly Suspicious (Orange). Two critical quantitative inconsistencies undermine the credibility of the reported results. First, in Table II (Camelyon17 ablation), the reported F1-Scores are mathematically incompatible with the paired Precision and Recall values according to F1 = 2·P·R/(P+R); for example, Experiment 1 reports P=59.41, R=84.51, yielding a true F1 of 69.78 versus the paper's 70.68, and Experiments 2 and 3 show comparable ~1-point discrepancies. Second, in Tables I and III (Camelyon16), reported Recall values such as 87.50%, 89.58%, 91.67%, and 93.75% correspond to non-integer true-positive counts over the fixed 77 positive WSIs in the official test set (e.g., 0.8958×77≈68.97), implying an unreal denominator (~48 or 96) inconsistent with Camelyon16. Ablation metrics further exhibit implausibly monotonic, noise-free progression across modules. No image-duplication evidence could be derived from the text-only figure descriptions. Confidence is moderate-to-high for the arithmetic and denominator issues, low for image reuse due to lack of pixel data.

Verdict

🟠 Highly Suspicious — The paper contains internally inconsistent numerical reporting that strongly suggests fabricated or hand-computed results, though final determination requires access to source code, training logs, and raw logits.

Key findings

Arithmetic collapse of F1 in Table II (Camelyon17 ablation). Reported F1 values cannot be derived from the paired Precision and Recall using the paper's own formula (14)–(17). Three independent rows show errors of roughly +0.88 to +1.09 points; the Camelyon16 table does not exhibit the same systematic offset, indicating selective data fabrication rather than a typographical issue.
"Half-tumor" denominator paradox in Tables I and III (Camelyon16). Recall values map to non-integer true positives over the official 77 positive WSIs in the Camelyon16 test split. The values 87.50%, 89.58%, 91.67%, and 93.75% are exactly consistent with a denominator of 48 (43/48, 44/48, 45/48), strongly implying the authors computed metrics against an incorrect positive-slide count.
Implausibly monotonic ablation curves. Across Baseline → FCM → GSAFormer → FASA in Tables I–II, multiple metrics are either identical to the previous row (e.g., ACC 77.38 in two consecutive rows) or rise in lockstep, an unlikely pattern for stochastic deep-learning training.
No image-based evidence available. Figure 4 (Confusion Matrix) is plausibly hand-rendered given the upstream numerical problems, but no pixel-level duplication analysis was possible from the provided text.
No methodological red flags beyond module stacking. References, equipment (RTX 3090), and timeline (2025) are internally coherent; no fabricated citations or impossible hardware claims.

Evidence highlights

Table II (Camelyon17), Experiment 1: P=59.41, R=84.51 → computed F1 = 69.78; paper reports 70.68 (Δ ≈ +0.90).
Table II (Camelyon17), Experiment 2: P=60.61, R=84.51 → computed F1 = 70.59; paper reports 71.68 (Δ ≈ +1.09).
Table II (Camelyon17), Experiment 3: P=60.82, R=83.10 → computed F1 = 70.10; paper reports 70.98 (Δ ≈ +0.88).
Table I / III (Camelyon16) Recall arithmetic: 0.8958 × 77 ≈ 68.97 and 0.9167 × 77 ≈ 70.58 TP slides — fractional pathology cases are physically impossible.
Inferred denominator: 43/48 = 0.8958, 44/48 = 0.9167, 45/48 = 0.9375 — the reported recalls are exact fractions with denominator 48, not 77.

Notes

DOI under review: 10.1109/ACAIT67930.2025.11522617 (IEEE, 2025 Asian Conference on Artificial Intelligence Technology).
Authors listed: Junfeng Zhang, Wei Jia, Defeng Kong, Xuefen Zhao. Affiliation inferred as Ningxia University per the report.
Recommended follow-up: request raw source code, training logs (log.json), and raw confusion-matrix outputs from the authors; raise the F1-mismatch and denominator concerns on PubPeer; notify the ACAIT / IEEE program committee; consider an institutional report to Ningxia University's research ethics committee.
Limitations: image duplication, Western-blot-style splicing, and peer-review-process integrity could not be assessed from text alone. This report is AI-assisted and intended for academic-discussion purposes only; final adjudication rests with the relevant institutional and editorial bodies.