Fabrication and fundamental errors 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 flags multiple severe problems in Hussien et al. (Scientific Reports, 2025; DOI: 10.1038/s41598-025-15604-6). Findings include: (1) In Table 2, VGG19 and ResNet152 on the Brain Tumor dataset share identical F1, Precision, and Recall values (0.7388, 0.6723, 0.8199) but different accuracies—an implausible coincidence for distinct architectures. (2) Equations for Precision and Recall on page 11 incorrectly use TN (true negatives) in the numerator instead of TP, indicating fundamental misunderstanding; the term 'Percsion'/'Percison' also appears. (3) Dataset references are mismatched: citations [38], [39], [40] do not correspond to the claimed IQ-OTH/NCCD, brain tumor, or leukemia datasets. (4) The abstract lists 4 datasets but reports 5 percentage values, and the Introduction repeats the same sentence verbatim across adjacent paragraphs. (5) Table 1 cites Agnes et al. with accuracies of 0.937% and 0.989%, which are below chance for medical classification. (6) Equation numbering restarts on page 15, suggesting text was spliced from multiple sources. Verdict: confirmed data fabrication and gross academic negligence. Confidence is high based on internal textual evidence; independent verification of the original PDF is recommended.

Verdict

Confirmed fabrication and extreme academic sloppiness. The paper contains fundamental mathematical errors, implausibly duplicated metrics, misattributed references, and copy-paste artifacts that collectively indicate the reported results cannot be trusted. Immediate editorial and institutional investigation is warranted.

Key findings

Identical metrics across different architectures (Table 2): VGG19 and ResNet152 on the Brain Tumor dataset produce identical F1 (0.7388), Precision (0.6723), and Recall (0.8199) but divergent accuracies (0.7531 vs 0.8119), which is mathematically incoherent and strongly suggests data copying or fabrication.
Wrong metric formulas (Eqs. 5 & 6, p. 11): Precision and Recall are defined using TN in the numerator (TN/(TN+FP) and TN/(TN+FN)) instead of TP, betraying a basic misunderstanding of evaluation metrics. The word "Percsion"/"Percison" also appears in the text.
Misattributed dataset citations: References [38], [39], and [40] do not match the claimed lung cancer (IQ-OTH/NCCD), brain tumor, and leukemia datasets; cited works concern semantic search, embeddings, and vector databases.
Logical inconsistencies in the Abstract: Four datasets are listed but five accuracy percentages are given (100%, 99%, 99%, 100%, 95%).
Duplicated paragraph in Introduction: The sentence "Supervised models often overfit to narrow labelled distributions, whereas SSL learns domain-invariant features through pretext tasks" is repeated verbatim in consecutive paragraphs.
Implausible literature-review numbers (Table 1): Agnes et al. is cited with accuracies of 0.937% and 0.989%, values that are below chance and indicate the cited paper was not actually read.
Equation numbering resets on p. 15: Cosine Similarity and Precision@k are labeled as Equation (5) and (6), duplicating earlier numbering, a hallmark of stitched-together drafts.

Evidence highlights

DOI: 10.1038/s41598-025-15604-6
Table 2: VGG19 and ResNet152, Brain Tumor — F1 0.7388, Precision 0.6723, Recall 0.8199 (identical); Accuracy 0.7531 vs 0.8119.
Page 11, Eqs. (5) and (6): Precision = TN/(TN+FP); Recall = TN/(TN+FN).
Refs. [38] Seifert et al. (semantic/similarity search); [39] Joulin (One embedding space to bind them all); [40] IEEE vector-database paper.
Abstract accuracies: "100%, 99%, 99%, 100 and 95%" attached to four datasets.
Table 1, Agnes et al.: Acc = 0.937% and 0.989%.

Notes

Confidence is high for findings 1–4 based on internal textual evidence (formula definitions, abstract arithmetic, duplicated sentence). Finding 5 depends on accurate transcription of Table 1 from the original PDF, which should be cross-checked. The editorial process at Scientific Reports is a clear failure point, as multiple elementary errors survived peer review. Any final determination of misconduct should come from an institutional or publisher investigation.