Machine Learning Algorithms Incorporating Radiomic Ultrasound Features (R-IOTA)

Epithelial ovarian cancer and other tubo-ovarian malignancies represent a major clinical challenge among gynecological cancers, particularly in developed countries, where they are associated with substantial morbidity and mortality. Despite advances in diagnosis and treatment, the prognosis of advanced tubo-ovarian cancer remains poor, with overall 5-year survival rates ranging approximately between 30% and 45%. Early and accurate characterization of adnexal masses is therefore essential to optimize clinical management, ensure appropriate referral of patients with suspected malignancy to specialized centers, and avoid unnecessary surgical procedures in patients with benign disease.

International scientific societies, including the European Society of Gynaecological Oncology (ESGO), International Society of Ultrasound in Obstetrics and Gynecology (ISUOG), International Ovarian Tumor Analysis (IOTA) group, and European Society of Gynaecological Endoscopy (ESGE), recommend ultrasound examination as the first-line imaging modality for the evaluation and risk stratification of adnexal masses. Expert ultrasound assessment and validated prediction models, such as the Assessment of Different NEoplasias in the adneXa (ADNEX) model developed by the IOTA group, have demonstrated high diagnostic accuracy for the discrimination between benign and malignant adnexal lesions and have shown the ability to further classify malignant masses according to different tumor categories.

However, ultrasound-based characterization of adnexal masses remains dependent on the availability of expertise and may be influenced by interobserver variability. The development of objective and reproducible diagnostic tools capable of extracting additional information from ultrasound images may therefore represent an important opportunity to improve diagnostic consistency and support clinical decision-making.

In recent years, radiomics has emerged as an innovative approach for the quantitative analysis of medical images. Radiomic techniques allow the extraction of high-dimensional imaging features that may capture subtle patterns and characteristics not readily identifiable through conventional visual interpretation. When combined with machine learning (ML) algorithms, radiomic features derived from ultrasound images may provide new opportunities for automated lesion characterization and improved prediction of tumor behavior.

Preliminary studies investigating ultrasound radiomics and artificial intelligence approaches in adnexal masses have shown promising results; however, currently available evidence is limited by small sample sizes, single-center designs, heterogeneous methodologies, and insufficient external validation. Large multicenter international studies are needed to evaluate the reproducibility, generalizability, and clinical applicability of AI-based approaches in this setting.

The present multicenter international IOTA study aims to develop and validate ultrasound-based machine learning models integrating radiomic features for the characterization of adnexal masses. The study will collect and analyze ultrasound imaging data from participating centers and will evaluate the ability of AI-based models to provide accurate and reproducible classification of adnexal lesions.

The primary objective of the study is to assess the performance of ultrasound-based machine learning models incorporating radiomic features in distinguishing benign from malignant adnexal masses.

Secondary objectives include the evaluation of additional AI-based models for the characterization of malignant adnexal masses, including the discrimination between borderline tumors, primary invasive malignancies, and metastatic lesions among masses classified as malignant by the primary model. Furthermore, the study aims to evaluate the ability of AI-based approaches to differentiate primary epithelial ovarian carcinoma from non-epithelial ovarian malignancies among lesions classified as primary ovarian cancer.

The study will also investigate the potential clinical utility of the developed models through decision curve analysis, evaluating the potential net benefit of AI-assisted prediction strategies across different clinical decision thresholds.

An independent prospective external validation cohort will be used to further evaluate the generalizability of the developed models and to explore their potential impact in a real-world clinical setting. This analysis will be performed retrospectively and post hoc, comparing AI model outputs with actual clinical decisions made during routine patient care. Importantly, AI predictions will not be available to clinicians and will not influence diagnostic pathways, treatment decisions, surgical planning, or patient management.

The retrospective clinical utility assessment will include evaluation of the agreement between AI predictions and clinicians' preoperative assessment of malignancy based on available clinical and imaging information. In addition, the study will explore the potential proportion of surgical procedures that could have been avoided according to AI model predictions, while maintaining patient safety. Surgical and follow-up outcomes, including intraoperative and postoperative complications and complications occurring during follow-up, will also be evaluated, including specific analyses of cases classified retrospectively as potentially unnecessary procedures.

A health economic evaluation will be performed to compare healthcare resource utilization and costs associated with standard clinical management with a reconstructed scenario based on AI model outputs. This analysis aims to explore the potential cost-effectiveness of AI-supported strategies for the management of adnexal masses.

Patient-reported outcomes will also be assessed to evaluate aspects related to the patient experience during standard clinical management, including satisfaction with diagnostic communication, clarity of information provided, and perceived quality of care.

Overall, this study aims to provide robust multicenter international evidence regarding the performance, reproducibility, and potential clinical utility of ultrasound-based machine learning models incorporating radiomic features for the evaluation of adnexal masses. The ultimate goal is to investigate whether AI-assisted approaches may contribute to a more standardized, accurate, and personalized diagnostic pathway while supporting appropriate allocation of healthcare resources.