Retractions of publications in radiomics: An underestimated problem?

Radiomics is increasingly explored as a tool for improving diagnosis, prognosis, and treatment planning. However, concerns exist about the reproducibility and methodological rigor of its studies. The integration of high-dimensional radiomic features and machine learning makes the field prone to unintentional errors that may warrant retraction. Despite a rising number of retractions in science overall, no dedicated study has examined retractions specifically within radiomics. Therefore, this study aimed to review retracted radiomics publications and identify the characteristics and reasons for their retraction. We systematically searched six databases (Crossref, Retraction Watch Database, OpenAlex, PubMed, Scopus, Web of Science) and identified 93 retracted radiomics publications, of which 20 were included. These articles were analyzed with respect to publisher, country of origin, dates, citation counts, and reasons for retraction. Retraction rates were then estimated and compared with those in general radiology. Our findings indicate that a disproportionate number of retractions are linked to specific publishers and countries (particularly China and India), with overall low citation counts (median 4.0 citations). Retractions peaked sharply in 2023, followed by a strong decline. Many retraction notes lack a clear explanation for the retraction. Estimated retraction rates in radiomics were lower than in general radiology (6.7 vs 7.4 per 10,000 publications). Notably, no major radiological or oncological journal appears to have retracted a radiomics publication. Given that radiomics demands higher, interdisciplinary expertise, this suggests a gap, implying that flawed research may yet have to be retracted. KEY POINTS: Question Considering the technical complexity of radiomics studies and their susceptibility to unintentional errors, how do their retraction rates compare to those in general radiology? Findings Retractions in radiomics were disproportionately linked to specific publishers and countries; however, no retractions appeared in major journals. Estimated retraction rates were lower than those for general radiology publications. Clinical relevance A potential gap in the number of retracted radiomics studies was identified, implying that flawed research in the field may not yet have been addressed.