Artificial intelligence-based machine learning models for preoperative diagnosis and staging of ovarian tumors

BACKGROUND: Ovarian cancer remains the most lethal gynecological malignancy, necessitating precise diagnostic strategies to improve patient outcomes. This study aims to develop and evaluate machine learning models that utilize patient history, imaging, and blood test data to differentiate between benign and malignant ovarian tumors and predict the stage of malignant cases. METHODS: A total of 357 individuals diagnosed with ovarian tumors participated in the study. Among these, 139 tumors were identified as benign, 40 as borderline, and 178 as malignant. The analysis employed four machine learning classifiers support vector machine (SVM), random forest (RF), logistic regression (LR), and decision tree utilizing 39 features derived from blood tests, imaging, and the patient's background to generate diagnostic outcomes. The study focused on assessing the significance of these features in predicting malignancy and determining the stage of the disease. RESULTS: The RF algorithm demonstrated the highest accuracy, reaching 94% based on imaging and tumor markers, with an AUC of 0.9. Key features contributing to this success include Human Epididymis Protein 4 (HE4) and Cancer Antigen 125 (CA125). In terms of staging malignant tumors, the SVM exhibited lower error rates, particularly in predicting advanced-stage disease (AUC: 0.77). Notably, CA125 and the presence of ascites emerged as the most influential factors for accurately staging the disease. CONCLUSION: The utilization of AI models proves effective in accurately classifying both malignant and benign ovarian tumors, showcasing promising advancements in diagnostic capabilities.