Performance of radiomics analysis in ultrasound imaging for differentiating benign from malignant adnexal masses: A systematic review and meta‐analysis

Abstract

Introduction

We present the state of the art of ultrasound‐based machine learning (ML) radiomics models in the context of ovarian masses and analyze their accuracy in differentiating between benign and malignant adnexal masses.

Material and Methods

Web of Science, PubMed, and Scopus databases were searched. All studies were imported into RAYYAN QCRI software. All studies that developed and internally or externally validated ML models using only radiomics features extracted from ultrasound images were included. The overall quality of the included studies was assessed using the QUADAS‐AI tool. Summary sensitivity and specificity analyses with corresponding 95% confidence intervals (CIs) were reported.

Results

12 studies developed ML models including only radiomics features extracted from ultrasound images, and six of them were included in the meta‐analysis. The overall sensitivity and specificity for differentiating benign from malignant adnexal masses were 0.80 (95% CI 0.74–0.87) and 0.86 (95% CI 0.80–0.90), respectively, in the validation set. All studies demonstrated a high risk of bias in subject selection (e.g., lack of details on image sources or scanner models; absence of image preprocessing), and the majority also showed a high risk in the index test (e.g., models were not validated on external datasets) domain. In contrast, the risk of bias was generally low for the reference standard (i.e., most studies used a reference that accurately identified the target condition) and the testing workflow (i.e., the time interval between the index test and reference standard was appropriate) domains.

Conclusions

The good performance of ultrasound‐based radiomics models in the validation set supports that radiomics is worth exploring to improve the diagnosis of adnexal masses. So far, the studies have a high risk of bias due to the small sample size, single‐setting design, and no external validation included.

Currently, the IOTA‐ADNEX model and the O‐RADS are the most reliable methods for calculating the risk of malignancy in adnexal masses. However, some skill is required to accurately recognize the ultrasound variables included in these methods. In recent years, several authors have demonstrated that radiomic analysis, which does not rely on expert interpretation or the recognition of ultrasound variables, performs almost as well as the methods currently in use in discriminating between benign and malignant adnexal masses. With further development and validation, it has the potential to be integrated into ultrasound imaging systems, enhancing clinical decision‐making and potentially evolving into a more autonomous diagnostic tool in the future.

Abbreviations

ADNEX

Assessment of Different Neoplasia in the adnexa

AI

artificial intelligence

CT

computed tomography

IOTA

International Ovarian Tumor Analysis group

ML

machine learning

MRI

magnetic resonance imaging

O‐RADS

Ovarian‐Adnexal Reporting and Data System

QUADAS‐AI

Quality Assessment Tool for Artificial Intelligence‐Centered Diagnostic Test Accuracy Studies criteria

Key message.

The IOTA‐ADNEX model and O‐RADS reliably assess adnexal masses but require expert interpretation. Radiomics, which does not require expert input, shows promise in distinguishing between benign and malignant adnexal masses and could, with further development, evolve into an autonomous ultrasound tool.

1. INTRODUCTION

Ultrasound is the first imaging modality for the diagnosis of gynecological diseases ¹ and plays a crucial role in the characterization of ovarian neoplasms. ² Since 1999, ultrasound has achieved a high degree of accuracy in the diagnosis of ovarian masses when performed by an experienced examiner. ³ Over the past decade, modern tools have been developed and tested by the International Ovarian Tumor Analysis (IOTA) group to achieve the same performance as an experienced examiner in correctly differentiating benign from malignant ovarian masses. In particular, the IOTA‐Assessment of Different Neoplasia in the adneXa (ADNEX) model, which combines clinical and simple ultrasound variables, ⁴ has been proposed in international guidelines for the diagnosis and management of patients with adnexal masses. ² The model showed high accuracy (AUC 0.94, sensitivity 0.97 and specificity 0.71) in distinguishing between benign and malignant adnexal masses.

In 2020, the IOTA group and the American College of Radiology jointly published a consensus guideline on the Ovarian‐Adnexal Reporting and Data System (O‐RADS) for ultrasound. ⁵ O‐RADS categorizes adnexal masses into six risk groups ranging from a normal ovary to high risk of malignancy and guides the management of adnexal masses. According to the consensus statement, malignancy risk can be estimated using either the ADNEX model or the examiner's interpretation of ultrasound findings based on the O‐RADS lexicon.

Both ADNEX and O‐RADS lexicons are practical diagnostic tools that can be integrated into ultrasound devices, providing an objective evaluation of ovarian masses without requiring extensive experience. However, some skill is necessary to accurately identify the ultrasound variables included in these methods.

In recent years, the scientific community has developed a great interest in artificial intelligence (AI) and radiomics applied to imaging, including ultrasound. ⁶ , ⁷ , ⁸ , ⁹ AI and radiomics‐based models are revolutionizing the current approach to diagnosis and patient management as they are objective and clinician‐independent. In medical imaging, AI systems help interpret images, detect anomalies, and predict clinical outcomes, enhancing diagnostic accuracy and treatment planning, including imaging data, clinical variables, and radiomics features. Radiomics is a technique used to extract, analyze, and interpret quantitative data from medical images. ¹⁰ The types of features commonly extracted include statistical features (which reflect the intensity distribution and heterogeneity of the image) and textural features (which capture patterns and variations in tissue architecture based on the spatial arrangement of gray‐level pixels). Additionally, other types of features such as morphological features (which describe the shape and size of structures) and wavelet features (which analyze the image at multiple scales and orientations) can also be utilized to gain a more comprehensive understanding of the underlying tissue characteristics.

Since 2020, the role of AI and radiomics has been increasingly explored in the field of ultrasound for both benign and malignant gynecological conditions. ⁶ Most of the studies in the literature have focused on the creation of classification tasks with the main interest in the differentiation between benign and malignant ovarian tumors. ⁶ , ⁷ However, there is still a lack of synthesis of the available evidence regarding radiomics models in the prediction of ovarian malignancies.

We aim to present the state of the art of ultrasound‐based machine learning (ML) radiomics models in the context of ovarian masses and analyze their accuracy in differentiating between benign and malignant adnexal masses.

2. MATERIAL AND METHODS

2.1. Search strategy

A comprehensive literature search was conducted using the Web of Science, PubMed, and Scopus databases to identify potentially eligible articles published up to 1 November 2024. A search strategy was developed using a combination of Medical Subject Headings (MeSH), keywords, and free‐text terms, including: “radiomics,” “ultrasound‐based radiomics,” “artificial intelligence,” “machine learning,” “deep learning,” “ultrasonography,” “gynecology,” “gynecological diseases,” “ovary,” “ovarian,” “ovaries,” and “fallopian tube.” The search was restricted to studies involving human participants and published in English. No additional filters or limitations were applied. The full search strategies for all databases are available in Supporting Information Appendix S1. This systematic review was conducted and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses (PRISMA) guidelines. The protocol was registered in PROSPERO (registration number CRD42024613425).

2.2. Inclusion criteria

Articles that developed ML models using only radiomics features extracted from ultrasound images to differentiate between benign and malignant adnexal masses were included. Articles that did not include an independent (internal or external) validation set were not considered for the meta‐analysis.

2.3. Exclusion criteria

Articles that do not consider an independent validation set (i.e. papers with cross‐validation: a type of validation in which the whole set is divided into a number of folds, usually 10, among which one is used as test set, while the other 9 were used as training set, this cycle of validation is repeated the same number of times as the number of sets, and in each specific cycle the test set is different from the previous one, over the 10 folds) were considered but not included for the meta‐analysis. In the case of papers with multiple ML models (i.e., a model with only radiomics features and models with both radiomics and other types of features and/or a deep learning model using images), we consider the model with only radiomics features for the analysis.

Systematic reviews, non‐empirical studies, animal studies, conference abstracts, editorials, commentaries, book reviews, and abstracts without corresponding full‐text articles were excluded from eligibility for inclusion in the present systematic review.

2.4. Study selection

All studies retrieved through the search strategy were imported into the RAYYAN QCRI software, and duplicates were removed. Two authors (M.C., M.S.) independently reviewed all abstracts, with consensus reached regarding potential relevance. The full‐text versions of the selected papers were acquired, and four reviewers (F.M., M.C., G.B., and M.S.) independently extracted relevant data on study characteristics. Any discrepancies were discussed among the reviewers, and consensus was achieved. In cases where multiple studies were published for the same cohort with identical endpoints, the report containing the most comprehensive information on the population was included to prevent overlap.

Reference lists of the included studies were searched to identify additional studies.

When the full text could not be retrieved online or specific data needed for the meta‐analysis (e.g., number of patients, images of benign tumors, images of malignant tumors, or the number of patients in the training and validation sets) were unavailable, the corresponding authors of the articles were contacted.

2.5. Data extraction and analysis

Data extraction was performed by four researchers (F.M., M.C., G.B., and M.S.) using a standardized data extraction form. The following information was extracted for each eligible study: (1) study identification (first author, publication year); (2) study characteristics (study period, country, population); (3) specific type of ML model assessed; and (4) performance outcomes expressed as sensitivity and specificity. A summary of the results is presented in a dedicated table.

The reported performance refers to either the external or internal validation set. Specifically, if a model was validated on an external cohort, its performance is based on the results obtained from the external validation set. In the absence of external validation, performance data from the internal validation set were reported. If neither external nor internal validation was performed, the performance of the training set was reported, and the study was excluded from the meta‐analysis.

Statistical analyses were conducted using R statistical software (version 4.2.1), utilizing the Meta and Metaplus packages.

Sensitivity and specificity, along with their corresponding 95% confidence intervals (CIs), were calculated. A random effects model was applied to account for heterogeneity arising from variations in clinical settings. Heterogeneity between studies was assessed using Cochrane's Q test and the I ² index, with p‐values <0.05 considered indicative of significant heterogeneity. Pooled estimates and their 95% CIs were visualized using forest plots.

2.6. Quality assessment

The overall quality of the studies included in the meta‐analysis was assessed using the Quality Assessment Tool for Artificial Intelligence‐Centered Diagnostic Test Accuracy Studies (QUADAS‐AI) criteria. ¹¹ The details are provided in Table S1. These criteria are based on the revised and extended QUADAS‐2 ¹² and QUADAS‐C ¹³ guidelines and include four domains: patient selection, index test, reference standard, and flow and timing, all in relation to the risk of bias. This updated tool evaluates each domain, offering a robust framework for conducting reviews of AI‐centered diagnostic studies.

3. RESULTS

3.1. General characteristics

Twelve studies ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ , ²¹ , ²² , ²³ , ²⁴ , ²⁵ developed ML models that included only radiomics features extracted from ultrasound images (Table 1). Five of them lacked an independent internal or external validation set ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ and one lacked the information required for the analysis ¹⁹ (i.e., number of patients/images with benign and malignant histology in the validation set) and was therefore finally excluded for the meta‐analysis (Figure 1) (Table 2).

TABLE 1.

Data source for the 12 studies that developed machine learning models using only radiomic features extracted from ultrasound images.

First Author, year	Country	Assessment time	Sample size	Number of images	Type of ML	Family of radiomics features	Performance (AUC)	Sensitivity	Specificity	Accuracy	Type of validation
Acharya U. R. 2014	NA	NA	20	2600	KNN	Textural, statistical	NA	1	1	1	Cross‐validation
Pathak H. 2015	NA	NA	NA	120	SVM	Wavelet	NA	0.92	0.92	0.92	Internal
Martinez‐Mas J. 2019	Belgium	NA	NA	384	Extreme Learning Machine with linear‐Sigmoid‐Gaussian kernel	Geometric (Fast Fourier Transform) features	0.88	0.93	0.77	0.87	Cross‐validation
Al‐karawi D. 2021	UK	2005–2013	232	150	SVM	Statistical, fractal, Gabor filter, Uniform Local Binary Pattern, Histograms of Oriented Gradient	NA	0.90	0.90	0.90	Internal
Chiappa V. 2021	Italy	2017–2019	241	NA	Ensemble of SVM (TRACE4©)	Morphology, intensity‐based statistics, texture	Solid: 0.87 Cystic:0.88 Mixed:0.89	Solid: 0.78 Cystic:0.75 Mixed: 0.81	Solid: 0.83 Cystic:0.90 Mixed: 0.81	Solid: 0.80 Cystic:0.87 Mixed: 0.81	Cross‐validation
Hussein I. J. 2021	NA	NA	NA	125	SVM	Local Binary Pattern	NA	0.97	0.90	0.95	Cross‐validation
Qi L. 2021	China, Tianjin	2013–2016	265	279	Multivariate logistic regression	Textural, wavelet	0.88	0.76	0.90	0.84	Internal
Yin Q. 2022	China; Guangzhou	2008–2019	137	NA	Upgrade Logistic regression	Local Binary Pattern, Laws' Texture Energy	NA	NA	NA	0.97	No validation
Barcroft J. F. 2024	Multicenter (UK; Italy)	2017–2022	761	1920	ODS radiomics model	Fractal, wavelet	0.90	1	0.73	NA	External
Du Y. 2024	China	2014–2022	849	NA	LightGBM	Statistical, shape, textural	0.75	0.60	0.89	0.72	Internal
Moro F, 2024	Italy	2014–2021	326	775	Random Forest	Statistical and textural	0.80	0.78	0.76	0.78	Internal
Liu L. 2024	Multicenter (China)	2021–2023	1080	NA	Not specified	Statistical, shape, textural	0.87	0.84	0.80	0.81	External

Note: Summary of the studies included in the paper listed according to different parameters (first author, publication year, study period, country, population) and AI models characteristics (specific type of ML model being assessed, model input, performance, validation type). NA (Not Specified); AUC (Area Under the receiver operating characteristic Curve); ML (Machine Learning); LightGBM (Light Gradient Boosting Machine); KNN (k‐Nearest Neighbor); ODS (Ovarian Diagnostic Score); SVM (Support Vector Machine). The performance has been reported in “AUC” when available, otherwise in “sensitivity/specificity”, then in “accuracy”. The indicated performance refers to the external or internal validation set. Studies subsequently included in the meta‐analysis are highlighted in gray.

FIGURE 1.

Flow diagram.

TABLE 2.

Data source for the 6 studies that developed and validated machine learning models using only radiomic features extracted from ultrasound images included in the meta‐analysis.

First Author, year	Country	Assessment time	Sample size	Sample size validation	Sample size validation malignant	Sample size validation benign	Number of images	Images validation	Images validation malignant	Images validation benign	Type of ML	Family of radiomics features	Performance (AUC)	AUC (CI)	Sensitivity	Sensitivity (CI)	Specificity	Specificity (CI)	Accuracy	Type of validation
Pathak H. 2015	NA	NA	NA	NA	NA	NA	120	50	25	25	SVM	Wavelet	NA	NA	0.92	NA	0.92	NA	0.92	Internal
Al‐karawi D. 2021	UK	2005–2013	232	NA	NA	NA	150	76	38	38	SVM	Statistical, fractal, Gabor filter, Uniform Local Binary Pattern, Histograms of Oriented Gradient	NA	NA	0.90	NA	0.90	NA	0.90	Internal
Qi L. 2021	China, Tianjin	2013–2016	265	NA	NA	NA	279	83	53	30	Multivariate logistic regression	Textural, wavelet	0.88	(0.798–0.957)	0.76	(0.574–0.883)	0.90	(0.774–0.963)	0.84	Internal
Du Y. 2024	China	2014–2022	849	169	60	109	NA	NA	NA	NA	LightGBM	Statistical, shape, textural	0.75	NA	0.60	NA	0.89	NA	0.72	Internal
Moro F. 2024	Italy	2014–2021	326	98	73	25	775	NA	NA	NA	Random Forest	Statistical and textural	0.80	(0.70, 0.90)	0.78	(0.69, 0.88)	0.76	(0.69, 0.88)	0.78	Internal
Liu L. 2024	China	2021–2023	1080	397	107	290	NA	NA	NA	NA	Not specified	Statistical, shape, textural	0.87	(0.827–0.903)	0.84	NA	0.80	NA	0.81	External

Note: Summary of the studies included in the paper listed according to different parameters (first author, publication year, study period, country, study population, validation set population) and AI models characteristics (specific type of ML model being assessed, model input, performance, validation type). NA (Not Specified); AUC (Area Under the receiver operating characteristic Curve); CI (Confidence Interval); ML (Machine Learning); LightGBM (Light Gradient Boosting Machine); SVM (Support Vector Machine). The performance has been reported in “AUC” when available, otherwise in “sensitivity/specificity”, then in “accuracy”. The indicated performance refers to the external or internal validation set.

The results of the quality assessment for the studies included in the meta‐analysis, as evaluated using the QUADAS‐AI tool, are reported in Figure 2. All studies were found to have a high risk of bias in the subject selection domain (i.e., source was not specified, image preprocessing was not performed, and the scanner model used for image acquisition was not reported), and all but one study exhibited a high risk in the index test domain (i.e., the radiomic model was not validated in an external cohort). Generally, there was a low risk of bias in the reference standard domain (i.e., the reference standard used in most studies accurately classified the target condition), as histology was specified as the reference standard in the majority of studies (4/6), and in the workflow domain (i.e., the time interval between the index test and the reference standard was deemed reasonable in most studies).

FIGURE 2.

Risk of bias for each domain in included studies in the meta‐analysis using QUADAS‐AI. “High” = high risk of bias; “Low” = low risk of bias; “Uncertain” = insufficient data to permit the definition of bias whether high or low.

3.2. Individual characteristics of the included studies

All studies but one were conducted in a single center, most in China. All studies were published between 2015 and 2024, and the sample size ranged from 232 to 1080 patients.

The summary sensitivity and specificity of radiomics‐based models in discriminating between benign and malignant adnexal masses was 0.80 (95% CI 0.71–0.87) and 0.86 (95% CI 0.80–0.90), respectively (Figure 3A,B).

FIGURE 3.

Forest plots of summary sensitivity (A) and specificity (B) of studies included in the meta‐analysis. CI, confidence interval.

The study by Liu et al. ²⁵ included the largest number of patients (total number = 1080) and was the only study with an external validation set (training = 683, validation = 397). However, it was the only study that used the Ovarian‐Adnexal Reporting and Data System (O‐RADS) rather than histology as the reference standard and did not specify the type of ML model. Specifically, O‐RADS risk classes 1–2‐3 were classified as benign (low risk of malignancy) and O‐RADS risk classes 4–5 were classified as malignant (intermediate/high risk of malignancy). The developed ML model showed good performance (AUC 0.87, sensitivity 0.84 and specificity 0.80) in the external validation set. The performance of the model was compared with that of O‐RADS assessed by junior examiners and was similar (AUC 0.88, sensitivity 0.84, specificity 0.92).

Du et al., ²³ including 849 patients (training = 680, validation = 169), developed and internally validated a LightGBM radiomics model that showed AUC of 0.75, sensitivity of 0.60, and specificity of 0.89 in the internal validation set.

Moro et al., ²⁴ in a population of 326 patients (training = 228 and validation = 98), developed and internally validated a random forest model with AUC 0.80, sensitivity 0.78, and specificity 0.76 in the validation set. The performance of the model was lower than the subjective assessment of an expert (sensitivity 0.99, specificity 0.72) and the IOTA‐ADNEX (AUC 0.88, sensitivity 0.99 and specificity 0.64). In this study, the authors included only adnexal masses with a solid ultrasound morphology.

Qi et al. ²² developed and internally tested a multivariate logistic regression radiomics model on a total of 279 images (training = 196, validation = 83). The model showed higher performance than subjective assessment of both junior and senior sonographers in terms of AUC (0.88 vs. 0.70 and 0.79), while sensitivity and specificity were shown to be higher than the junior sonographer (0.76 vs. 0.57 and 0.90 vs. 0.80) and similar to the senior sonographer (0.76 vs. 0.70 and 0.90 vs. 0.86) in the validation set.

Al‐karawi et al. ²¹ collected 150 images of ovarian masses (training = 74, validation = 76). Their Support Vector Machine model fusing the top 5 best‐performing textural features showed great performance in distinguishing between benign and malignant adnexal masses, with sensitivity 0.90, specificity 0.90, and accuracy 0.90. Similar results were obtained by Pathak et al., ²⁰ who developed and internally tested a Support Vector Machine radiomics model on a total number of 120 images (training = 70, validation = 50). Their radiomics model showed sensitivity 0.92, specificity 0.92, and accuracy 0.92 in classifying adnexal masses using the 6 more relevant features of the 14 extracted features.

4. DISCUSSION

In the present study, we summarized the results of articles that developed and validated ML models using only radiomics features extracted from ultrasound images to distinguish between benign and malignant adnexal masses. We found that the summary accuracy of the models in the validation set was good and similar to that of an expert examiner. Most studies were at high risk of bias for subjective selection and index testing due to small sample size, lack of external validation, and single‐center setting.

To the best of our knowledge, this is the first meta‐analysis specifically focused on the application of radiomics to ultrasound imaging in the context of adnexal masses. We conducted a literature search across multiple databases to ensure the robustness of the study. Data such as sample size, number of images, type of ML model developed, and the families of radiomic features included were reported. Additionally, the quality of the studies was assessed using the QUADAS‐AI tool, which has been specifically adapted for AI research, representing a key strength of this study.

Finally, we only included models constructed with radiomic features, excluding mixed models developed with both radiomic and other variables (i.e., ultrasound data, clinical features). This was because we wanted to know the performance of radiomic analysis without other information in characterizing adnexal masses. This could be a strength, but also a limitation. In fact, this strategy prevented us from verifying whether the performance of the models would be higher if other data were added.

Currently, various AI approaches are employed in ultrasound imaging to enhance diagnostic accuracy, with primary applications in classification and automatic segmentation. In particular, ML techniques play a crucial role in radiomics, where quantitative features extracted from ultrasound images are analyzed to classify lesions, predict malignancy, and support clinical decision‐making by detecting patterns that are not discernible to the human eye. Supervised learning models, such as Support Vector Machines, Logistic Regression, and Random Forests, are trained on labeled datasets to predict outcomes based on input features and are widely used in radiomic analysis. Support Vector Machines identify optimal decision boundaries between classes, Logistic Regression estimates the probability of specific outcomes, and Random Forests aggregate predictions from multiple decision trees to enhance classification robustness. Additional algorithms, including k‐Nearest Neighbors (kNN), Extreme Gradient Boosting (XGBoost), and LightGBM, further refine predictions by comparing data point similarities or iteratively minimizing prediction errors. In contrast, unsupervised learning models are less commonly used than supervised models to predict outcomes, such as distinguishing between benign and malignant ovarian masses. Instead, they are primarily employed to uncover inherent patterns in unlabeled data, typically through clustering techniques, to reveal hidden structures or groupings without predefined labels. For instance, fuzzy c‐means clustering assigns each pixel a degree of membership to multiple clusters based on its features. This approach is particularly useful for assessing echogenicity and segmenting regions of interest. All these unsupervised models should be further explored to better understand their potential role in radiomic feature selection and the diagnosis of ovarian masses.

Finally, AI also includes deep learning algorithms, a subset of ML, which leverage multi‐layer neural networks to model complex data and automatically learn hierarchical features directly from raw inputs. In the context of ovarian masses, DL models are primarily applied for two purposes: (1) automated segmentation of regions of interest, and (2) image‐based classification. In the pipeline of the radiomics analysis, DL models can be used for automatic segmentation of the region of interest for feature extraction.

The role of radiomics in the field of adnexal masses has been studied and summarized by other authors in other imaging modalities, including computed tomography (CT) and magnetic resonance imaging (MRI). ²⁶ In particular, in the systematic review by Huang et al., 57 articles were included dealing with radiomics applied to CT and MRI in ovarian cancer, and 16 of them aimed to classify adnexal masses. However, the outcomes assessed varied across the included studies, including discrimination between epithelial ovarian and metastatic tumors, borderline and invasive tumors, and benign and borderline, as well as malignant and benign adnexal masses. Of the 9 studies reported in the study by Huang et al. that aimed to discriminate between benign and malignant masses ²⁷ , ²⁸ , ²⁹ , ³⁰ , ³¹ , ³² , ³³ , ³⁴ , ³⁵ (Table S2), seven had an independent internal or external validation set, and the performance of the models was similar to that found for the ultrasound‐based model, with sensitivity ranging from 0.60 to 1.0 and specificity from 0.60 to 1.0. In particular, the largest study, including 1329 patients, developed and internally validated a radiomics‐based model applied to CT imaging, reporting a sensitivity of 0.84 and a specificity of 0.82 in the validation set. ³¹

As with ultrasound, most studies applying radiomics to CT and MRI for the assessment of adnexal tumors are few, include small sample sizes, and lack external validation. All these factors limit the generalizability of the results and therefore the application of radiomics analysis in clinical practice. Indeed, to date, the IOTA‐ADNEX model remains the tool of choice for preoperative diagnosis of adnexal masses. ⁵ However, we believe that increasing the sample size and adding other variables to the radiomics analysis could improve the performance of radiomics in classifying adnexal masses. We also suggest for those authors who would develop radiomics models applied to ultrasound images to compare the model with that of ADNEX to assess the true clinical impact and suitability of the radiomics model itself in discriminating between benign and malignant adnexal masses. In addition, if a radiomics model is developed that achieves similar performance to ADNEX, it should be validated in an external cohort to ensure its reproducibility. If the performance of the radiomics model will be similar or superior to that of ADNEX, radiomics and ML models could be integrated into ultrasound machines in the future to assist clinicians in the automated, consistent, and immediate evaluation of adnexal masses using the information provided by a single ultrasound image. Incorporating radiomics systems into clinical practice can improve diagnosis and patient management, reduce healthcare costs, and help reduce the workload of gynecologists by increasing their efficiency and accuracy. Other prospects include the development of models for predicting specific histology to personalize surgical treatment and patients' response to treatment.

5. CONCLUSION

Our meta‐analysis demonstrated the suitability of radiomics analysis applied to ultrasound images in discriminating between benign and malignant adnexal masses. It seems to have similar accuracy to an expert examiner. However, the comparison with the reference IOTA‐ADNEX model has not been sufficiently investigated. Moreover, the number of studies present in the literature is few, lacks external validation, and includes small sample sizes.

AUTHOR CONTRIBUTIONS

Francesca Moro: Conceptualization; data curation; investigation; project administration; supervision; visualization; writing—review and editing. Marianna Ciancia: Conceptualization; data curation; investigation; project administration; supervision; visualization; writing—review and editing. Maria Sciuto: Data curation; investigation; writing–original draft. Giulia Baldassari: Data curation; formal analysis; investigation; visualization; writing—original draft. Huong Elena Tran: Data curation; formal analysis; investigation; visualization; supervision. Antonella Carcagnì: Methodology. Anna Fagotti: Supervision. Antonia Carla Testa: Supervision; writing—review and editing.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflict of interest.

Supporting information

Appendix S1. Search strategy.

Table S1. Description of quality assessment based on QUADAS‐AI domains.

Table S2. Data source for the 9 studies that developed machine learning models using only radiomic features extracted from Computed Tomography (CT) or Magnetic Resonance Imaging (MRI) images, reported in the study by Huang et al.

Moro F, Ciancia M, Sciuto M, et al. Performance of radiomics analysis in ultrasound imaging for differentiating benign from malignant adnexal masses: A systematic review and meta‐analysis. Acta Obstet Gynecol Scand. 2025;104:1433‐1442. doi: 10.1111/aogs.15146