Forensic review of '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 forensic review of a 2025 Scientific Reports article on DINO V2–based self-supervised medical image diagnosis concludes that the paper contains multiple severe, mutually reinforcing anomalies consistent with data fabrication and methodological incompetence. The verdict is strongly adverse. Key findings: (1) the Precision and Recall formulas in the evaluation section are inverted (using TN in the numerator instead of TP), raising the question of how any reported metrics were actually computed; (2) Table 2 values fail basic arithmetic — e.g., ResNet152 on Retina gives Precision 0.435 and Recall 0.8423, yielding an expected F1 of ~0.574, but the table reports 0.8420; (3) VGG19 and ResNet152 on Brain tumor report identical Precision, Recall and F1, suggesting copy-paste fabrication; (4) the text claims DINO V1 (20 epochs) is best for lung cancer, brain tumor and leukemia, while Table 3 actually shows DINO V2 (20 epochs) outperforming V1 (e.g., lung cancer 0.99 vs 0.9); (5) the vector database 'Qdrant' is repeatedly misspelled 'Quadrant', and a fabricated term 'Qudra Net dataset' is invented, suggesting the authors do not understand their own pipeline; (6) the Introduction and Background duplicate passages verbatim and Table 3 contains malformed numbers (e.g., '0.97.8'). Confidence in these findings is high because they are verifiable from text and tables; limitations are that original code/training logs were not independently obtained.

Verdict

🔴 Substantiated — Multiple independently verifiable errors in formulas, table arithmetic, narrative consistency, and tool nomenclature point strongly toward data fabrication and/or deep methodological negligence. The paper does not meet basic standards for publication in its present form.

Key findings

Wrong evaluation formulas (Eqs. 4–6, p. 11): Precision is written as TN/(TN+FP) and Recall as TN/(TN+FN). These actually compute specificity, not the claimed metrics.
Arithmetically impossible F1 values (Table 2): Reported F1 scores contradict the harmonic mean of the listed Precision and Recall in at least two rows (ResNet152/Retina, EfficientNet_b7/Retina).
Identical metrics across distinct models (Table 2, Brain tumor): VGG19 and ResNet152 share identical Precision (0.6723), Recall (0.8199) and F1 (0.7388) — a hallmark of copy-paste fabrication.
Text contradicts Table 3 (pp. 13–14): The narrative states DINO V1 with 20 epochs is best for lung cancer, brain tumor and leukemia, yet Table 3 shows DINO V2 (20 epochs) with higher accuracy in each case (e.g., lung cancer 0.99 vs 0.90; brain tumor 0.994 vs 0.99).
Tool-name confusion and invented terminology (pp. 10, 14–15): The vector database Qdrant is repeatedly written 'Quadrant', and the phrase 'Qudra Net dataset' is fabricated; Qdrant is a database, not a dataset.
Self-plagiarism and sloppy formatting: The Introduction and Background duplicate the radiologist workload paragraph nearly verbatim; Table 3 contains malformed entries such as '0.97.8' and '98.2' without a decimal prefix.

Evidence highlights

Equation (4): Precision = TN/(TN+FP); Equation (5): Recall = TN/(TN+FN). Correct: TP/(TP+FP) and TP/(TP+FN).
Table 2 (Retina, ResNet152): P=0.435, R=0.8423 ⇒ expected F1 ≈ 0.574; reported F1 = 0.8420.
Table 2 (Retina, EfficientNet_b7): P=0.258, R=0.8139 ⇒ expected F1 ≈ 0.391; reported F1 = 0.8160.
Table 2 (Brain tumor): VGG19 and ResNet152 share (Precision, Recall, F1) = (0.6723, 0.8199, 0.7388).
Table 3 (lung cancer, 20 epochs): DINO V1 Accuracy = 0.9 vs DINO V2 Accuracy = 0.99; text asserts V1 is best.
Misnaming: 'Quadrant' used for Qdrant; 'Qudra Net dataset' used for what is actually a vector DB.
Formatting anomalies in Table 3: '0.97.8', '98.2'.
DOI: 10.1038/s41598-025-15604-6.

Notes

All numeric quotes above are taken directly from the reported tables and equations; values were not recomputed beyond the stated harmonic-mean checks.
No original training code, logs, or datasets were independently accessed; the conclusions rest on the manuscript itself and on arithmetic/textual consistency.
The combination of an inverted formula, impossible F1 values, identical cross-model metrics, a narrative that contradicts its own table, and invented tool terms cannot be plausibly dismissed as mere typos.
Recommended actions: request raw data and code from the authors, post a structured query on PubPeer, and notify the editorial office of Scientific Reports and the authors' institution.
This report summarizes publicly observable inconsistencies; formal misconduct determinations remain the responsibility of the journal and institutional review bodies.