Fabrication and fundamental errors in "Explainable self-supervised learning for medical image diagnosis based on DINO V2 model and semantic search"

Academic fraud report · Geng Detector

Summary

This report assesses serious integrity concerns in Hussien et al. (2025), published in Scientific Reports (DOI: 10.1038/s41598-025-15604-6). The verdict is a confirmed finding of substantive academic fraud, primarily driven by mathematical contradictions and likely fabricated tabular results. The most damaging issue is the incorrect formulation of Precision and Recall in Equations 5 and 6: Precision is given as TN/(TN+FP) and Recall as TN/(TN+FN), which are textbook errors that invert the standard definitions and actually correspond to Specificity and the true negative rate. Compounding this, Table 2 reports F1 scores that violate the harmonic mean constraint—e.g., ResNet152 on the Retina dataset shows Precision 0.435 and Recall 0.8423 but claims F1 = 0.8420, which is mathematically impossible. Table 3 contains implausibly perfect 1.00 scores across all metrics for DINO V2 (20 epochs) on Leukemia, contradicting the text. References are malformed or misattributed. Confidence in the verdict is high for the mathematical errors, which are independently verifiable; lower confidence for intent, which requires institutional investigation. Recommended actions include journal notification, PubPeer posting, and institutional review.

Verdict

Confirmed substantive fraud / severe integrity failure (🔴). The paper contains independently verifiable mathematical errors and tabular data that violate basic arithmetic constraints, consistent with fabrication. The author's institution (Kafr El-Sheikh University) and Scientific Reports should initiate formal investigation and consider retraction.

Key findings

Erroneous metric formulas (Eqs. 5–6): Precision is written as TN/(TN+FP) and Recall as TN/(TN+FN). These correspond to Specificity and true negative rate, respectively. The standard definitions are Precision = TP/(TP+FP) and Recall = TP/(TP+FN).
Mathematically impossible F1 scores in Table 2: F1, as the harmonic mean of Precision and Recall, cannot exceed the smaller of the two values multiplied by a bound. For ResNet152 on Retina (Precision 0.435, Recall 0.8423), maximum F1 ≈ 0.575, yet 0.8420 is reported. For EfficientNet_b7 on Retina (Precision 0.258, Recall 0.8139), maximum F1 ≈ 0.39, yet 0.8160 is reported.
Implausible identical metrics in Table 2 (Brain tumor): VGG19 and ResNet152, with different accuracies (0.7531 vs 0.8119), show identical Precision (0.6723) and identical Recall (0.8199).
Implausibly perfect scores in Table 3: DINO V2 (20 epochs) on Leukemia reports Accuracy, F1, Precision, and Recall all equal to 1.00.
Text–table inconsistency in DINO V1/V2 section: The prose claims "accuracy and F1-score of 99%" for lung cancer, whereas Table 3 shows DINO V1 (20 epochs) Lung cancer Accuracy = 0.90 and F1 = 0.99.
Malformed or misattributed references: Reference 5 cites "Huang, G. et al. DINO-Mix…" with no year or journal as the source for DINO V2; Reference 38 (Lung cancer dataset) cites an unrelated semantic-search paper with no year or journal; Reference 39 (Brain tumour dataset) cites a Joulin/Ai/Retrieval paper with a truncated title.
Language and editorial defects: Broken sentences (e.g., "…rapid case The results affirmed…"), ungrammatical conclusion ("Unless the superiority of our proposed model."), and misspelling of technical terms (Accuarcy, Percsion, percsion@k).

Evidence highlights

Formula check (Page 11, Eqs. 5–6): TN in the numerator of both Precision and Recall—definitionally incorrect; equivalent to inverting the confusion-matrix roles of positives and negatives.
F1 bound check (Page 11, Table 2): Harmonic mean upper bound = 2·P·R/(P+R). Applied to (0.435, 0.8423) yields ≈ 0.575, not 0.8420. Applied to (0.258, 0.8139) yields ≈ 0.390, not 0.8160.
Cross-model equality (Page 11, Table 2): VGG19 vs ResNet152, Brain tumor: Precision identical to 4 decimal places (0.6723); Recall identical to 4 decimal places (0.8199)—statistically near-impossible for independently trained models.
Perfect 1.00 block (Page 12, Table 3): DINO V2 (20 Epochs, Leukemia): Accuracy = F1 = Precision = Recall = 1.00.
Reference anomalies (References list): Ref. 5 (no year/journal), Ref. 38 (off-topic, no year/journal), Ref. 39 (truncated title).

Notes

All numeric values cited are reproduced exactly from the source report and the underlying article (DOI: 10.1038/s41598-025-15604-6).
The formula and F1-bound errors are independently verifiable from public text; the conclusion of fabrication rests on the convergence of multiple independent anomalies, but formal attribution of intent requires institutional investigation.
The affective commentary ("耿同学辣评") is the source author's editorial opinion and is not included as a finding.