In the current study, 3,193 studies were retrieved from online databases. Duplicates were removed, leaving 2,736 studies. After reviewing abstracts, 31 articles were obtained. After excluding 12 studies unrelated to radiomics, 2 studies lacking necessary data, 1 study with irrelevant outcome measures, and 1 study not related to model construction, a total of 15 studies were selected for this meta-analysis (Fig. 1).

Literature selection process

The 15 eligible articles were published between 2018 and 2024 and encompassed a total of 9,111 IA, of which 3,182 were ruptured IA. Of these studies, 10 [15] were from China; 3 [25] were from South Korea; the remaining two studies were from Japan [28] and the United States [29]. Among the included studies, 13 [15] focused on the diagnosis of ruptured IA and were case-control studies; 2 [17] were cohort studies predicting the risk of rupture. Of the included studies, 13 employed retrospective designs, and 2 used prospective designs. Nine studies were conducted across multiple centers (three of which utilized data from registries), while the remaining six studies were single-center ones. Among the studies, radiomics data sources included 5 from CTA [15], 10 from DSA [16], and 3 from MRA [28] Among these, 3 were externally validated, and the rest were internally validated through random sampling. The models used were primarily based on machine learning (Tables 1 and 2).

In the training set, we performed a meta-analysis of 10 diagnostic 2 × 2 tables. The pooled results for sensitivity, specificity, PLR, NLR, diagnostic OR, and SROC curve were 0.87 (95% CI: 0.78–0.92), 0.85 (95% CI: 0.79–0.90), 5.9 (95% CI: 4.0–8.6), 0.16 (95% CI: 0.09–0.27), and 37 (95% CI: 17–83), respectively (Figs. 2 and 3). No publication bias was observed in the detection of IA rupture status by radiomics in the training set according to Deek’s funnel plot (P = 0.31) (Fig. 4). Approximately 36% of patients in the included studies had ruptured IA, which was used as the prior probability. With a PLR of 5.9, the model predicted a 72% probability that a rupture status indicated by the model was indeed a ruptured IA. Conversely, if the model indicated a non-ruptured aneurysm, the probability that the aneurysm was truly non-ruptured was 93% (Fig. 5).

Forest plot of meta-analysis for sensitivity and specificity of radiomics in distinguishing IA rupture status in the training set

SROC curve of meta-analysis for radiomics in identifying IA rupture status in the training set

Deek’s funnel plot of meta-analysis for radiomics in detecting IA rupture status in the training set

Nomogram of meta-analysis for radiomics in determining IA rupture status in the training set

In the validation set, a meta-analysis of 16 validation cross-tables was conducted, yielding the following summary statistics for sensitivity, specificity, PLR, NLR, diagnostic OR, and SROC curve: 0.84 (95% CI: 0.76–0.90), 0.82 (95% CI: 0.77–0.86), 4.7 (95% CI: 3.7–5.8), 0.19 (95% CI: 0.13–0.29), and 24 (95% CI: 15–40), respectively (Figs. 6 and 7). Deek’s funnel plot indicated no publication bias in the assessment of radiomics for detecting IA rupture status in the validation set (P = 0.05) (Fig. 8). Approximately 36% of patients in the eligible studies had ruptured IA, which was used as the prior probability. Given a PLR of 4.7, if the model suggests a ruptured status, the probability of an actual ruptured IA is 72%; if the model suggests a non-ruptured aneurysm, the probability of an actual non-ruptured aneurysm is 90% (Fig. 9).

Forest plot of meta-analysis showing sensitivity and specificity of radiomics for differentiating IA rupture status in the validation set

SROC curve from meta-analysis assessing radiomics for differentiating IA rupture status in the validation set

Deek’s funnel plot from meta-analysis evaluating radiomics for IA rupture status in the validation set

Nomogram of meta-analysis for radiomics in determining IA rupture status in the validation set

Among the studies included, only two reported predictions of rupture risk for UIA based on radiomics. In light of the limited studies, only a summary could be provided. In the study by Chubin Ou et al. [19], radiomics showed a sensitivity of 0.865, specificity of 0.687, PLR of 2.764, NLR of 0.197, and diagnostic OR of 14.064 for predicting aneurysm rupture risk. The study by Yi Yang et al. [17] reported an AUC of 0.816 in the training set, with sensitivity and specificity of 0.7667 and 0.8658, respectively, demonstrating significant statistical relevance.

Of the 15 studies, nine did not offer detailed imaging protocols, thus scoring zero points for the first criterion. Seven studies did not mention methods for segmenting multiple images, resulting in zero points for the second criterion. None of the studies employed repeated measurements across different images or imaging at multiple time points, leading to zero points for the third, fourth, and fifth criteria. Six studies did not address multivariable analysis with non-radiomic features, thus scoring zero for the sixth criterion. None of the studies examined or discussed biological relevance, leading to zero points for the seventh criterion. Twelve studies did not perform cutoff value analysis, resulting in zero points for the eighth criterion. Twelve studies did not calibrate statistical data, thus scoring zero for the tenth criterion. Fourteen studies did not register prospective studies in trial databases, resulting in zero points for the eleventh criterion. All studies did not compare with a “gold standard”, leading to zero points for the thirteenth criterion. Twelve studies did not report potential clinical applications, thus scoring zero for the fourteenth criterion. None of the studies conducted a cost-effectiveness analysis, leading to zero points for the fifteenth criterion. Three studies did not publicly disclose corresponding codes and data, resulting in zero points for the sixteenth criterion. The final average score for the included studies was 7.4 points.