Abstract
BACKGROUND AND OBJECTIVES: Presently, the radiomics and deep learning have achieved notable advancements in the automatic detection, stability assessment, and rupture risk prediction of intracranial aneurysms (IAs). However, there are still certain challenges in data quality, model robustness, and applicability among the published studies. Therefore, this study aims to provide a systematic review of the roles of radiomics and deep learning in the automatic detection, stability assessment, and rupture risk prediction of IAs, with providing insights for the individual stratified management of IAs patients. METHODS: Between January 2015 and December 2024, the literatures were retrieved in PubMed, Web of Science, Embase and the Cochrane Library using a combination of subject headings and keywords. Two researchers independently screened the literature of radiomics or deep learning in the automatic detection, stability and rupture risk prediction of IAs. The QUADAS-2 tool was used to assess the methodological quality of the included studies. The study characteristics and area under the curve (AUC) were summarized using tables to clearly present the research progress in this field. RESULTS: Ultimately, 28 original research of a total of 32,991 IAs were included. Notably, 89% of the publications (25/28) appeared between 2021 and 2024. Six studies for IAs automatic detection, all of them established deep learning frameworks, while five conducted multicenter analyses and only two carried out prospective validations. Six studies for assessing IAs stability underwent manual segmentation in data processing, three underwent multicenter analysis, and only three underwent model calibration and DCA analysis. 19 studies for IAs rupture risk, 15 studies underwent manual segmentation, while only two carried out fully automatic segmentation; regrettably, all the studies were retrospective studies, and only two underwent prospective validation. Importantly, in validation set, the AUC range for IAs automatic detection was from 0.791 to 0.930, the AUC range for assessing the stability was from 0.720 to 0.969, and the AUC range for rupture risk prediction was from 0.720 to 0.969. Moreover, approximately 50% of studies were based on small sample, single-center data, and most studies lacked model calibration and cross-platform validation. CONCLUSIONS: The models based on radiomics and deep learning for the automatic detection, stability assessment, and rupture risk prediction of IAs achieve excellent performance, and may be a potential non-invasive tool to help the clinical stratification management of high-risk IAs patients. However, current studies are predominantly retrospective, single-center, and heterogeneous, with limited calibration and external validation. Hence, there is a critical need for large-scale, prospective, long-term follow-up and multicenter studies to further establish the roles of these techniques. Future research should focus on building an easy-to-use and open-source dynamic online tool that combines standardized and normalized multi-center IAs radiomics systems with interpretable deep learning methods to enhance the accuracy and reliability of IAs risk assessment and management.