Machine learning-based analysis of MRI radiomics in the discrimination of classical and non-classical polycystic over syndrome

Abstract

Purpose: The aim of this study is to investigate the value of radiomics analysis on T2-weighted Magnetic Resonance imaging (MRI) images in differentiating classical and non-classical polycystic ovary syndrome (PCOS). Materials and Methods: A total of 202 ovaries from 101 PCOS patients (mean age of 23±4 years) who underwent pelvic MRI between 2014 and 2022, were included in the study. Of the patients, 53 (52.5%) were phenotype A, 12 (11.9%) were phenotype B, 25 were phenotype C (25.1%), and 11 were phenotype D (10.9%). 130 (64.4%) of the ovaries were classical PCOS, 72 (35.6%) were non-classical PCOS. The ovaries were manually segmented in all axial sections using the 3D Slicer program. A total of 851 features were extracted. Python 2.3, Pycaret library was used for machine learning (ML) analysis. Datasets were randomly divided into train (70 %, 141) and test (30 %, 61) datasets. The performances of ML algorithms were compared with AUC, accuracy, recall, precision and F1 scores. Results: Accuracy and AUC values in the training set ranged from 57%-73% and 0.50-0.73, respectively. The two best ML algorithms were Random Forest (rf) (AUC:0.73, accuracy:73%) and Gradient Boosting Classifier (gbc) (AUC:0.71, accuracy:70%). AUC, accuracy, recall and precision values and F1 score of the blend model obtained from these two models were 0.70, 73 %, 56 %, 66%, 58%, respectively. Conclusion: Radiomic features obtained from T2W MRI are successful in distinguishing between classical and non-classical PCOS.