Integrity review: "An MRI–pathology foundation model for noninvasive diagnosis and grading of prostate cancer" (Nature Cancer, 2025)

Academic fraud report · Geng Detector

Summary

Verdict: Severe integrity concerns flagged as a confirmed ("实锤") case by the detector, pending independent institutional verification. The Nature Cancer paper (DOI: 10.1038/s43018-025-01041-x) by Lizhi Shao and colleagues reports a multimodal MRI-pathology foundation model for prostate cancer diagnosis. Four issues are highlighted. (1) Implausibly perfect performance: ROC analyses in the PSA gray zone (4–10 ng/ml) yield AUC = 1.000 on External test sets #1 and #3 and on the Prospective test set, while the PSA-only comparator sits at AUC ≈ 0.500, suggestive of data leakage or curated test sets. (2) Suspected image reuse: T2WI, DWI and ADC panels in Figure 6c appear visually identical in contour, anatomy, edge crop and noise, inconsistent with physics of distinct MRI contrasts; the same attention heatmap may have been overlaid on a single base image. (3) Implausible clinical timeline: 100 prospective patients supposedly accrued and processed to whole-mount histopathology after radical prostatectomy within roughly two months. (4) Date inconsistency: Figure 2d caption reads "Feb 2023 and Mar 2022." Confidence is moderate-to-high based on visible text and figures; original DICOM, surgical records and code were not independently inspected. Outcomes require editorial and institutional investigation.

Verdict

🔴 Confirmed integrity concerns by automated review; independent editorial and institutional investigation required before any formal finding of misconduct.

Key findings

Implausibly perfect AUCs in the PSA gray zone (4–10 ng/ml): Figure 5a / Figure 3a report AUC = 1.000 on External test set #1, External test set #3, and the Prospective test set, with the PSA-only baseline at AUC ≈ 0.500.
Suspected multimodal image duplication (PS-style reuse): In Figure 6c, T2WI, DWI, and ADC panels appear to share identical prostate contours, internal anatomy, edge cropping, and noise patterns, with a single attention heatmap overlaid across modalities.
Implausible prospective collection timeline: The prospective cohort of 100 patients at JSPH is stated to have been collected within ~2 months (text: "Feb 2023 to Mar 2023"; figure caption variant: "Feb 2023 and Mar 2022"), yet ground truth requires whole-mount histopathology after radical prostatectomy.
Temporal logic typo: Figure 2d caption lists "JSPH between Feb 2023 and Mar 2022 (n = 100)", an inverted date range.

Evidence highlights

AUC = 1.000 reported simultaneously on multiple independent external and prospective test sets is statistically extraordinary for MRI/pathology-based prostate cancer prediction and is consistent with data leakage between training and evaluation cohorts or post-hoc test set curation.
T2WI reflects anatomy while DWI/ADC reflect water diffusion; genuine acquisitions show distinct SNR, artifacts, and geometric distortions. Identical contours and noise across the three modalities in Figure 6c indicate possible single-source image manipulation (grayscale remapping) rather than true multimodal acquisition.
Whole-mount prostate pathology after radical prostatectomy typically requires biopsy, diagnosis, surgical scheduling, recovery, sectioning, staining, scanning, and pathologist review. A continuous pipeline producing 100 such cases in ~2 months at a single center is operationally unrealistic.
The "Feb 2023 to Mar 2022" date inversion is a low-level but informative inconsistency that often arises when authors edit text to obscure timeline problems.

Notes

DOI: 10.1038/s43018-025-01041-x
Source PDF referenced: "An MRI–pathology foundation model for noninvasive.pdf"
Claims are based on text and figures visible in the published article; no original DICOM, surgical records, pathology timestamps, or code repository were independently examined.
Recommended actions: request raw MRI DICOM, surgical and pathology timestamps, and training/test split documentation from the authors; raise concerns on PubPeer; report to the Nature Cancer editorial office and to the corresponding author's institutional research integrity committee, with emphasis on prospective-study ethics and data authenticity.