This systematic review was performed in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. ¹⁷ The protocol was registered in the International Prospective Register of Systematic Reviews (PROSPERO), and the checklist is available in supplementary file 1. Ethical approval was not required as this study analysed previously published articles for which approvals had already been obtained.

This literature review search was conducted using four databases such as PubMed, Embase, Scopus, and Web of Science which included the following keywords: adnexal masses, ovarian lesions, machine learning models, radiomics and computed tomography. The detailed search strategy and Boolean operator combinations are given in supplementary file 2.

This review includes both prospective and retrospective studies of adnexal masses, mainly assessing the diagnostic, staging and prognosis of lesions through computed tomographic imaging and radiomics models. Original research articles that are ethically approved from the respective institutions and from peer-reviewed journals containing enough amount of automatic segmentation using radiomics models were included. Reviews, editorials, conference abstracts, nonhuman studies, case reports, small case series with fewer than 10 participants, and studies concerning predictive modelling were excluded.

Two reviewers independently performed the literature screening. The duplicate articles were removed using Rayyan. ¹⁸ Full texts of potentially relevant articles were retrieved and reviewed in detail. Discrepancies were resolved through discussion with a third reviewer. We also developed an extraction template to standardize extraction by including information on the studies, their participants, and diagnostic performance measures like AUC, sensitivity, and specificity. Meta-analysis was not performed in this review due to the heterogeneity between the models used in the included studies.

The quality and risk of bias of the studies were independently evaluated using the quality assessment of diagnostic accuracy studies 2 (QUADAS-2) ¹⁹ tool. This established framework looks at potential bias and applicability issues in four main areas: patient selection, index test, reference standard, and flow and timing. Each area was rated as having low, high, or unclear risk of bias. This process ensured a clear and organized evaluation of the studies’ validity and clinical importance.

After duplicate and abstract removal, a total of 1107 original studies were retrieved, and 12 were found to be eligible for full-text screening. Of these, articles are part of the review that fulfilled the inclusion criteria. This process of selection is elaborated in Figure 1.

This systematic review combines the results of 11 retrospective studies on CT radiomics, and machine learning published between 2021 and 2025, with a total sample of 4439 patients. The sample size of each study varied from 149 to 1329 patients, and one study included 185 tumors. The studies were mostly carried out in chine (n = 9), with two multicenter studies including patients from the UK, Germany, USA. ^{20– 23} Most of the studies were single -center studies (n = 6), with five studies including two to three centers, which improved external validity.

Clinical tasks included differentiation of benign and malignant ovarian tumors (n = 4), differentiation of serous borderline and malignant tumors (n = 2), prediction of FIGO stage (n = 1), detection of peritoneal metastases (n = 1), and prediction of overall survival (n = 2). All studies were performed using contrast -enhanced CT scans, mainly in the portal venous phase, with 3D VOI segmentation in most cases.

Machine learning algorithms used were logistic regression, support vector machine, random forest, K-nearest neighbors, XGBoost, LightGBM, and deep learning models like CNN and U-Net networks. Validation methods used were train-test split validation, internal validation, leave-one-out cross-validation, and external multi-cohort validation. The diagnostic performance reported was excellent, with AUC ranging from 0.79 to 0.96, accuracy of up to 87%, specificity of up to 89%, and prognostic C-index of up to 0.73, thereby confirming the stability of CT-based radiomics models for the characterization of ovarian cancer. The detailed study characteristics of articles included in the review are provided in Table 1.

Among the studies, radiomics and deep learning models showed moderate to excellent diagnostic performance, with AUC values ranging from 0.72 to 0.99 ( Table 2). The highest accuracy was reported by Chen et al., ²⁴ achieving an AUC of 0.98 to 0.99. This was followed by Li et al. ²⁵ with an AUC of 0.96, and Liu et al. ²⁶ with an AUC of 0.951. Liu et al. integrated radiomics with a ResNet-18 deep learning framework. Most studies reported AUC values above 0.85, indicating strong discriminative ability. Sensitivity ranged from 68% to 91.7%. The highest sensitivity was observed in Liu et al. ²⁶ at 91.7% and in Li et al. ²³ at 90%, indicating good detection performance. Specificity varied from 75% to 99%, with Leng et al. ²⁷ achieving the highest specificity at 99%. Overall, models that included wavelet-transformed features, higher-order texture metrics, and deep learning architectures consistently achieved better accuracy. This highlights the advantages of improved feature extraction and hybrid radiomics-DL strategies. These findings confirm the high diagnostic potential of radiomics and AI-based models for characterizing lesions, although differences in feature selection, modeling methods, and validation protocols led to varying performance across studies.

The quality of the studies included was assessed using the QUADAS-2. Generally, there was a low risk of bias in the domains of patient selection, index test, and reference standard, which is an indication of high methodological quality showed in Figure 2.

All the studies included in the review, namely Yu et al., ²¹ Li et al., ²⁵ Jan et al., ²⁸ Li et al., ²³ Linton-Reid et al., ²⁰ Leng et al., ²⁷ Chen et al., ²⁴ Yu et al., ²⁹ Su et al., ³⁰ and Liu et al., ²⁶ had a low risk of bias in the domain of patient selection, which is an indication that the studies had appropriate study populations and that there was no selection bias. The domains of index test and reference standard also had low risks of bias, which is an indication that the studies applied the tests appropriately and that they used accepted diagnostic reference standards. Low risk was found in most studies in the flow and timing domain, reflecting appropriate intervals between the index test and reference standard. However, a moderate risk was found in this domain by Leng et al., ²⁷ which could be attributed to differences in follow-up or reporting.

However, concerns regarding applicability were found to be high in most studies, primarily because of the single-center study nature, lack of heterogeneity in the population, and differences in imaging protocols and model validation approaches. Only Li et al. ²³ and Leng et al. ²⁷ reported a moderate level of concerns regarding applicability. Thus, although the internal validity was excellent, external validity is poor, and there is a need for multicenter, externally validated studies.