Predictive value of machine learning-based T2-weighted MRI radiomics in the diagnosis of polycystic ovary syndrome.

Objective: This study aims to explore the predictive performance of machine learning-based radiomic features extracted from T2-weighted magnetic resonance imaging (MRI) in differentiating between women with polycystic ovary syndrome (PCOS) and healthy counterparts.

Methods: The study included patients diagnosed with PCOS who had undergone pelvic MRI in the endocrine department between 2014 and 2022, along with an age-matched control group. The ovaries were manually segmented from T2-weighted images using the 3D Slicer software. Both first- and second-order features, including wavelet filters, were extracted from the images. Utilizing the Python 2.3 programming language and the Pycaret library, various machine learning algorithms were employed to identify highly correlated features. The optimal model was selected from the 15 algorithms assessed.

Results: The study involved a total of 202 ovaries from 101 patients with PCOS (mean age 23±4 years) and 78 ovaries from the control group comprising 40 individuals (mean age 24±5 years). In the training set, the machine learning models displayed accuracy and area under the curve (AUC) values ranging from 72% to 83% and 0.50 to 0.81%, respectively. Notably, the Light Gradient Boosting Machine (LightGBM) model emerged as the most effective model among the various machine learning algorithms, exhibiting an AUC of 0.81 and an accuracy of 83%. When evaluated on the test set, the AUC, accuracy, recall, precision and F1 values of the LightGBM model were 0.80, 82%, 91%, 86%, 88%, respectively.

Conclusion: Machine learning-based T2-weighted MRI radiomics seems viable in differentiating between individuals with and without PCOS.