Integrity Review Report: "Explainable self-supervised learning for medical image diagnosis based on DINO V2 model and semantic search" (DOI: 10.1038/s41598-025-15604-6)

Academic fraud report · Geng Detector

Summary

Verdict: RED — confirmed academic integrity failures. The 2025 Scientific Reports paper by Hussien et al. exhibits multiple severe defects consistent with fabricated or AI-generated content lacking human oversight. Key issues include (1) fundamental metric formulas written backwards (Precision given as TN/(TN+FP), Recall as TN/(TN+FN)), invalidating downstream results; (2) an internal logical contradiction in Table 2 where ResNet152 and VGG19 share identical F1, Precision, and Recall to four decimals despite differing Accuracy, mathematically impossible without fabrication; (3) reference misassignment where dataset citations point to unrelated semantic-search and vector-database papers, suggesting LLM hallucination; (4) self-plagiarised near-duplicate paragraphs between Introduction and Background; (5) broken cross-references to figure numbers; and (6) fabricated clinical validation percentages (92% clinical coherence, 89% radiologist agreement) with no radiologists, IRB approval, or methodology described. Confidence is high for findings 1, 2, and 3; clinical-data fabrication (finding 6) requires institutional confirmation. Limits: no raw data or code was inspected; analysis is text and table based only.

Verdict

🔴 Confirmed. The paper displays multiple compounding integrity failures — incorrect core formulas, internally contradictory tabulated metrics, mismatched references, self-plagiarism, and unsubstantiated clinical claims — consistent with AI-generated text assembled without scholarly review. Retraction inquiry is warranted.

Key findings

Swapped Precision/Recall formulas (Page 11). Formula (5) gives Precision = TN / (TN + FP) (this is specificity/TNR); Formula (6) gives Recall = TN / (TN + FN) (not a standard metric). The correct forms are TP / (TP + FP) and TP / (TP + FN). All downstream metrics in Tables 2–3 derived from these formulas are therefore unreliable.
Logically impossible table values (Page 11, Table 2). For Brain tumor, VGG19 lists Accuracy = 0.7531 with F1 = 0.7388, Precision = 0.6723, Recall = 0.8199, while ResNet152 lists Accuracy = 0.8119 with the same F1, Precision, and Recall to four decimals. Because Accuracy changed while F1, Precision, and Recall remained identical, the entries appear copied from a single row with only Accuracy edited.
Reference–content mismatch (Pages 7, 21). Dataset citations [38], [39], and [40] point to papers titled *Combined semantic and similarity search in medical image databases*, *One embedding space to bind them all*, and *Analyzing embedding models for embedding vectors in vector databases*, none of which are lung-cancer, brain-tumor, or leukemia imaging datasets. Strong indicator of LLM hallucination uncorrected by the authors.
Self-plagiarism / padded prose (Pages 1 and 4). The Introduction and the Background open with near-identical paragraphs restating the same idea with minor lexical changes.
Broken figure cross-references (Pages 12 and 15). The text refers to figures that do not correspond to their stated content (e.g., "Figures 8 and 9 show a comparison…"; sequential jump from Fig. 13 to Fig. 14 to Fig. 15 referencing query images and semantic search results without clear figure–text alignment).
Fabricated clinical evaluation data (Page 20, Table 4). Reports of "Clinical coherence 92%" and "Radiologist agreement 89%" appear with no mention of any radiologist involvement, sample size, recruitment, or ethics approval anywhere in the manuscript.

Evidence highlights

Formula errors: Eq. (5) TN / (TN + FP), Eq. (6) TN / (TN + FN) versus correct TP / (TP + FP) and TP / (TP + FN).
Table 2 Brain-tumor rows: VGG19 (Accuracy 0.7531, F1 0.7388, Precision 0.6723, Recall 0.8199) vs ResNet152 (Accuracy 0.8119, F1 0.7388, Precision 0.6723, Recall 0.8199).
Reference mismatches: dataset claims for IQ-OTH/NCCD lung cancer, a 7,023-image MRI brain-tumor set, and acute lymphoblastic leukaemia mapped to references whose titles concern semantic search and vector embeddings, not medical imaging datasets.
Table 4 figures: Clinical coherence 92%, Radiologist agreement 89% — no methods, IRB, or radiologist participants described.
DOI: 10.1038/s41598-025-15604-6.

Notes

This report is based on text and tabular inspection only; source code, raw predictions, and image files were not examined. Verification of the dataset identities, training/test splits, and confusion matrices would strengthen findings 1 and 2.
Confidence is high for findings 1–3; confidence for clinical-data fabrication (finding 6) is high but requires confirmation of authorship institutional records (listed as Kaferelshikh University, Egypt).
Authors and editors should be contacted for raw data and a corrigendum or retraction consideration.