The Use of Artificial Intelligence for Ultrasound Screening of Ovarian Cancer in Postmenopausal Women

Overview

The ASTOM study is a monocentric prospective observational pilot study conducted at the Gynecologic Oncology Unit of the IRCCS Azienda Ospedaliero-Universitaria of Bologna. This study is aimed at evaluating the application of artificial intelligence to ultrasound screening for ovarian cancer in postmenopausal women, a population at increased risk in which early diagnosis remains a major clinical challenge due to the absence of effective screening methods. Ovarian cancer accounts for a significant proportion of gynecological malignancies and is the leading cause of death among gynecologic cancers in developed countries. Although gynecologic ultrasound is currently the first-line imaging modality for the characterization of adnexal masses, its diagnostic performance is strongly dependent on operator expertise, leading to variability in interpretation and potential misclassification of lesions, particularly in complex cases such as multilocular or multilocular-solid ovarian cysts, which may represent either benign conditions such as cystadenomas or malignant lesions including primary ovarian carcinomas or metastases from gastrointestinal tumors. In this context, the ASTOM study seeks to develop an integrated predictive model combining clinical data, ultrasound imaging, and radiomic features extracted from images, with the goal of improving preoperative oncological risk stratification and supporting clinical decision-making, thereby contributing to a precision medicine approach that could reduce unnecessary surgical interventions in patients with low-risk lesions while ensuring appropriate management of high-risk cases. The study will enroll approximately 100 menopausal women aged between 18 and 90 years presenting with ultrasound evidence of multilocular or multilocular-solid ovarian cysts, either awaiting surgery or undergoing follow-up for stable adnexal masses. All participants will undergo standard clinical and ultrasound evaluations as part of routine care, with additional collection of anonymized clinical, imaging, and radiomic data for research purposes. Ultrasound images will be acquired using a dedicated machine and standardized protocols, and volumes of interest will be delineated by expert sonographers, after which radiomic features will be extracted using validated software tools and integrated into a centralized database. The predictive model will be developed using advanced machine learning techniques, including convolutional neural networks, to automatically or semi-automatically classify lesions according to their risk of malignancy and, in high-risk cases, to differentiate primary ovarian tumors from metastatic lesions, particularly those originating from the gastrointestinal tract, which often present with overlapping imaging characteristics. The primary endpoint of the study is the diagnostic performance of the integrated model in accurately stratifying ovarian lesions into different risk categories, measured through metrics such as sensitivity, specificity, accuracy, positive and negative predictive values, and area under the ROC curve, with comparison to the performance of expert sonographers. The secondary objective focuses on the model's ability to distinguish primary ovarian neoplasms from metastases, an area where current models such as the widely used ADNEX algorithm show limitations. Statistical analysis will include descriptive analysis of patient characteristics, evaluation of model performance, subgroup analyses, multivariate regression to control for confounding factors, and sensitivity analyses to assess robustness, with appropriate handling of missing data through imputation techniques. As a pilot feasibility study, no formal sample size calculation has been performed, but the estimated cohort size is based on historical patient volumes at the study center and is considered sufficient to develop and preliminarily validate the model, generating data that may inform larger future studies. The overall duration of the study is expected to be 24 months, including 12 months for patient recruitment, followed by phases of follow-up and data analysis. The ASTOM study represents an innovative attempt to integrate artificial intelligence into routine gynecological oncology practice, addressing current limitations of operator-dependent imaging interpretation and existing predictive models, with the potential to enhance diagnostic accuracy, optimize patient management pathways, and contribute to the broader implementation of data-driven precision medicine in ovarian cancer care.