Academic Integrity Review Report: Patient-Level Prediction of Multi-Classification Task at Prostate MRI Based on End-to-End Framework Learning From Diagnostic Logic of Radiologists

Academic fraud report · Geng Detector

Summary

This report assesses potential academic integrity concerns in a 2021 IEEE Transactions on Biomedical Engineering paper (DOI: 10.1109/TBME.2021.3082176) on end-to-end prostate MRI classification. The overall verdict is 'highly suspicious' (orange). Four issues were identified. The most severe finding concerns a statistical reporting anomaly: the Discussion claims significance of p<0.05 versus DCNN and radiomics baselines, but no p-values, test names, or statistical methodology appear anywhere in the results tables. A second key issue is an internal numerical contradiction in the dataset description (187 patients vs. N=178 in the same sentence, leaving 9 patients unaccounted for). A third issue involves suspiciously rounded and uniform accuracy values across independent external test sets, with TeS1 reporting exactly 0.800 and consistent standard deviation patterns. A fourth, lower-confidence finding notes unnatural oscillation in training curves and visually similar case-study images. Confidence is high for findings 1-3 based on the published text; finding 4 remains limited by lack of pixel-level evidence. Final determination requires institutional investigation.

Verdict

🟠 Highly suspicious. Multiple independent concerns are documented within the published text alone, with the most serious being a statistical-significance claim unsupported by any reported test or p-value. No allegation of fraud is made here; the findings are consistent with sloppy data handling, insufficient statistical reporting, or selective presentation.

Key findings

Statistical-significance claim without supporting evidence (Finding 3, 🔴): The Discussion states that the proposed method outperformed DCNN (p<0.05) and radiomics (p<0.05), yet Tables II, III, IV, and VIII contain no p-values and no statistical test is named anywhere in the methods.
Internal numerical contradiction in the dataset description (Finding 1, 🟠): Section IV.B refers to "187 patients of PUTH-p2 (N=178)" in a single clause, leaving 9 patients unaccounted for.
Suspiciously uniform and rounded accuracy figures (Finding 2, 🟠): Reported ACC-case values across independent external sets include 0.849±0.023 (TS), 0.824±0.051 (VS), 0.800±0.029 (TeS1), and 0.824±0.044 (TeS2); TeS1 lands exactly at 0.800 and standard-deviation patterns appear unusually regular.
Visual irregularity concerns (Finding 4, 🟡): Figure 6 training curves show unnatural sawtooth oscillation and the five "typical cases" in Figure 7 appear visually similar; flagged but not confirmed due to absence of pixel-level analysis.

Evidence highlights

Direct quote (Section IV.B, p. 3695): "Then, 187 patients of PUTH-p2 (N=178) without slice annotations are selected as an external test set (TeS1)." — The pair (187, N=178) is internally inconsistent.
Direct quote (Section VI, p. 3699): "In the experiment, our proposed method outperformed DCNN (p-value<0.05) and radiomics (p-value<0.05) significantly." — No corresponding p-values or statistical test procedures appear in Tables II–IV or VIII.
Quantitative pattern (Section V.B, Table III): ACC-case = 0.849±0.023 (TS); 0.824±0.051 (VS); 0.800±0.029 (TeS1); 0.824±0.044 (TeS2). TeS1 mean equals exactly 0.800, an unusually clean value for cross-center medical AI evaluation.
DOI: 10.1109/TBME.2021.3082176 — retained verbatim for traceability.

Notes

Findings 1–3 are derived solely from the published PDF text and tables and are verifiable by any reader; confidence is high.
Finding 4 relies on visual-summary heuristics applied to figures without high-resolution pixel inspection and is therefore marked as evidence-insufficient; it should not be treated as a confirmed irregularity.
The dataset discrepancy (187 vs. N=178) could reflect a typographical error, an inclusion/exclusion criterion not described, or a mislabeled subset; clarification from the authors is required.
The p-value claim may indicate that tests were performed but omitted from the manuscript — a remediable reporting defect — or that the claim is unsupported, which would be more serious.
No determination of misconduct is made here. Institutional review and author response (raw data, statistical scripts) are necessary before any formal conclusion.