Fabrication and Quality Control Failures in "Explainable Self-Supervised Learning for Medical Image Diagnosis Based on DINO V2 Model and Semantic Search" (Scientific Reports, 2025)

Academic fraud report · Geng Detector

Summary

This report documents multiple serious integrity issues in the article DOI 10.1038/s41598-025-15604-6, published in Scientific Reports. The verdict is strongly against the paper. Key issues include (1) fundamental errors in machine-learning metric definitions: Equations 5 and 6 misrepresent Precision and Recall, transposing them with Specificity and NPV formulae; (2) implausible numeric identities in Table 2, where VGG19 and ResNet152 on Brain Tumor share identical F1 (0.7388), Precision (0.6723) and Recall (0.8199) despite a ~6 percentage-point Accuracy gap (0.7531 vs 0.8119); (3) abstract-vs-results inconsistency (4 datasets cited, 5 percentages listed); (4) a misattributed reference to a quantum-physics paper (Bose–Einstein condensates) used to support SSL; (5) DINOv2 core citation placed at Reference 33 instead of the standard primary source. Confidence is high for the formula and table-identity findings (verifiable from text); medium for AI-generation suspicion (circumstantial). Limits: confusion matrices and raw data were not obtained, so fabrication cannot be 100% proven without institutional investigation.

Verdict

Severe concerns indicating data integrity and authorship-quality failures. Likely fabricated or copy-pasted metric values combined with incorrect ML formulae. The paper should be investigated by the journal editorial board.

Key findings

Incorrect metric definitions (Eqs. 4–7): Precision written as tn/(tn+fp) (= Specificity) and Recall as tn/(tn+fn) (= NPV). True Precision is tp/(tp+fp); true Recall is tp/(tp+fn).
Implausible identical metrics in Table 2 (Brain Tumor): VGG19 and ResNet152 report Accuracies of 0.7531 and 0.8119, yet identical F1_score = 0.7388, Precision = 0.6723, Recall = 0.8199. A ~6 pp Accuracy difference cannot reconcile with identical precision/recall/F1.
Recall equals Accuracy in multi-class task: In Table 2 (Retina, 4 classes), ResNet152 and EfficientNet show Recall numerically equal to Accuracy (0.8423 and 0.8139), which is statistically implausible.
Abstract–Results mismatch: Abstract lists accuracies "100%, 99%, 99%, 100 and 95%" (5 values) for only 4 datasets (Lung cancer, brain tumour, leukaemia, Eye Retina Disease).
Mislabelled figure: "Fig. 8" denotes the proposed framework on page 10 and is reused on page 12 for supervised-model comparison.
Erroneous numeric format: Table 3 lists "0.97.8" for DINO V1 (10 Epochs) Brain-tumor Accuracy.
Reference misattribution: Reference 6 (Altmann, R., 2022) on "Localization and delocalization of ground states of Bose–Einstein condensates under disorder" is cited to support self-supervised learning.
DINOv2 citation misplaced: The actual DINOv2 paper (Oquab et al.) appears only at Reference 33, while Reference 5 (Huang, G. et al., "DINO-Mix") is incorrectly foregrounded.
Weak statistical reporting: Friedman and Nemenyi tests are mentioned but no χ² values or p-values are provided; "critical difference" diagrams are not substantiated.

Evidence highlights

DOI: 10.1038/s41598-025-15604-6
Table 2 Brain Tumor: VGG19 Acc 0.7531, F1 0.7388, Precision 0.6723, Recall 0.8199; ResNet152 Acc 0.8119, F1 0.7388, Precision 0.6723, Recall 0.8199.
Table 2 Retina: ResNet152 and EfficientNet Recall numerically equals Accuracy (0.8423 / 0.8139).
Equation 5 (Precision = tn/(tn+fp)) and Equation 6 (Recall = tn/(tn+fn)).
Abstract accuracies: "100%, 99%, 99%, 100 and 95%" against 4 datasets.
Reference 6 title contains "Bose–Einstein condensates under disorder" (physics).
Received: 7 October 2024; Accepted: 8 August 2025 (~10-month review).

Notes

Findings are reproducible from the published PDF; deeper proof of data fabrication requires the authors' raw confusion matrices and training logs, which should be requested via PubPeer or editorial correspondence.
AI-generation suspicion is circumstantial (formula errors, figure-number drift, tone issues) and cannot be confirmed without forensic text analysis or author response.
Recommendation: file a formal concern with Scientific Reports editorial board and post detailed queries on PubPeer; institutional ethics investigation is warranted if the journal declines action.