Fifty-seven studies including 4590 patients were evaluated. Several questions were repeatedly addressed. These include the differentiation of benign and malignant lesions in 30% (22/57), determining the histological subtype in 26% (15/57), assessing treatment response, nuclear grade in 5% (3/57) and the presence of specific mutations in 4% (2/57) manuscripts.

The included studies achieved an average Radiomics Quality Score of 3.41 (10.8% of total) with a good inter-rater agreement (ICC: 0.96, 95%-CI: 0.93-0.98). Only 4/57 studies validated the proposed radiomics signatures on independent datasets.  

Analysis with the QUADAS-2 tool revealed that the risk of bias among the studies is particularly elevated due to heavy reliance on retrospective, surgical cohorts which were enriched in patients with malignant tumours. Additionally, machine learning methods were only infrequently described in sufficient detail to allow replication of the results.

The meta-analysis revealed an average diagnostic odds ratio of 5.90 (95% CI: 4.02 – 8.63 p < 0.001) for the differentiation of RCC and angiomyolipoma without visible fat (Figure 2). The trim and fill analysis indicated that one study with a diagnostic odds ratio below average was missing, suggesting possible publication bias. However, even after the addition of this study, the diagnostic odds ratios would have remained significant with an odds ratio of 5.55 (Figure 3).

The radiomics features selected for differntiating angiomyolipoma without visible fat and RCC differed strongly between the studies. The mean unenhanced intensity as well as entropy in the unenhanced and nephrographic phase were found to differentiate AMLwvf and RCC on CT twice independently.