Academic Integrity Review: 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 review concludes with a strong verdict of confirmed (实锤) academic fraud for the paper DOI 10.1038/s41598-025-15604-6. Multiple independent lines of evidence indicate systematic data fabrication and fundamental misunderstandings of machine learning basics. Key issues include: (1) numerically impossible self-consistency in Table 2, where F1 scores contradict reported Precision and Recall values (e.g., ResNet152 on Retina: reported F1 = 0.8420 versus computed F1 ≈ 0.573 from P=0.435, R=0.8423); (2) verbatim duplicate Precision/Recall/F1 triplets across VGG19 and ResNet152 on the Brain tumor dataset despite differing accuracy, indicating copy-paste fabrication; (3) mathematically wrong definitions of Precision as TN/(TN+FP) and Recall as TN/(TN+FN) in equations 5–6; (4) textual claims (EfficientNet accuracy 0.8671, F1 0.8913) that match a different model (VGG19 on Leukemia) rather than Table 2; and (5) mis-matched or irrelevant citations, including a Bose–Einstein condensate physics paper cited for DINO V2, a word-embedding NLP paper cited for the Brain tumour dataset, and a vector-database paper cited for the Leukaemia dataset. A minor anomaly in Table 3 records an impossible value of '0.97.80'. Confidence is high for findings 1–4 due to internal contradiction; finding 5 relies on reference inspection and should be confirmed by the editorial office. Limits: all numeric values are reproduced exactly as published and were not independently recomputed against raw data.

Verdict

🔴 Confirmed (实锤) academic misconduct. The paper exhibits systematic, internally contradictory evidence of data fabrication, formula misuse, and reference misattribution that go beyond honest error. Independent corroboration by the editorial office (requesting code, raw predictions, and confusion matrices) is warranted.

Key findings

Mathematically impossible metrics in Table 2 (supervised learning).
*Retina / ResNet152*: reported Precision 0.435, Recall 0.8423, F1 Score 0.8420. Standard F1 = 2·P·R/(P+R) ≈ 0.573, not 0.8420.
*Retina / VGG19*: Precision 0.432, Recall 0.8723, F1 Score 0.8422. Computed F1 ≈ 0.578.
*Leukemia / ResNet152*: Precision 0.3054, Recall 0.9052, F1 Score 0.9054. Computed F1 ≈ 0.456.
In each case the reported F1 essentially equals the reported Recall, strongly suggesting a coding error or fabricated values.
Copy-paste duplication across models. For the Brain tumour dataset, VGG19 and ResNet152 share *identical* Precision = 0.6723, Recall = 0.8199, and F1 = 0.7388, while their accuracies differ (0.7531 vs 0.8119). This statistical coincidence is effectively impossible.
Wrong definitions of Precision and Recall. Equations 5 and 6 in "Experimental setup" define Precision = tn / (tn + fp) and Recall = tn / (tn + fn). Standard definitions require True Positives (TP), not True Negatives (TN), in the numerator. This is a fundamental conceptual error that would invalidate any metrics derived from these formulas.
Text contradicts Table 2. The text on page 11 ("Supervised models") states EfficientNet attained accuracy 0.8671 and F1 0.8913 for brain tumours. Table 2 shows EfficientNet_B7 on Brain tumour has Accuracy = 0.8099 and F1 = 0.7383. The cited numbers 0.8671 and 0.8913 match VGG19's Leukemia results in Table 2, indicating copy-paste between sections.
Mis-matched citations.
Reference [39] ("One embedding space to bind them all"), a word-embedding NLP paper, is cited for the Brain tumour MRI dataset.
Reference [40] ("Analyzing embedding models for embedding vectors in vector databases"), an IEEE vector-database paper, is cited for the Leukaemia dataset.
Reference [6] ("Localization and delocalization of ground states of Bose–Einstein condensates under disorder"), a quantum/condensed-matter physics paper, is cited in the DINO V2 description.
Non-valid numeric value in Table 3. DINO V1 at 10 Epochs reports Brain tumour Accuracy = 0.97.80, which is not a valid decimal number.

Evidence highlights

| Location | Reported values | Expected / Cross-check | Issue | |---|---|---|---| | Table 2, Retina, ResNet152 | P=0.435, R=0.8423, F1=0.8420 | F1 should ≈ 0.573 | Mathematically inconsistent | | Table 2, Retina, VGG19 | P=0.432, R=0.8723, F1=0.8422 | F1 should ≈ 0.578 | Mathematically inconsistent | | Table 2, Leukemia, ResNet152 | P=0.3054, R=0.9052, F1=0.9054 | F1 should ≈ 0.456 | Mathematically inconsistent | | Table 2, Brain tumour | VGG19 vs ResNet152 share P=0.6723, R=0.8199, F1=0.7388 | Identical metrics, different accuracy | Copy-paste duplication | | Eqs. 5–6 | Precision = tn/(tn+fp), Recall = tn/(tn+fn) | Should use TP in numerator | Fundamental definitional error | | p. 11 text | EfficientNet: Acc=0.8671, F1=0.8913 (brain tumour) | Table 2 EfficientNet_B7: Acc=0.8099, F1=0.7383 | Numbers match VGG19 Leukemia row | | Table 3 | Accuracy = 0.97.80 | Not a valid number | Typographical / fabrication artefact |

Notes

All numeric values above are reproduced exactly from the published article; no value has been altered.
Recalculated F1 values were computed using the textbook formula F1 = 2·P·R/(P + R) and rounded for readability; the underlying discrepancy is robust to rounding.
DOI: 10.1038/s41598-025-15604-6.
Timeline per journal metadata: Received 7 October 2024; Accepted 8 August 2025.
Recommended actions: request raw confusion matrices and training/inference code from the authors; verify reference [39], [40], and [6] contexts in the published PDF; notify the editorial office and the authors' institutional research integrity bodies.
This review is an automated-assisted assessment; final determinations of misconduct are reserved for the publisher and the authors' institutions.