Role of machine learning algorithms in predicting the treatment outcome of uterine fibroids using high-intensity focused ultrasound ablation with an immediate nonperfused volume ratio of at least 90%

Abstract

OBJECTIVE: This study aimed to investigate the role of machine learning (ML) classifiers to determine the most informative multiparametric (mp) magnetic resonance im-aging (MRI) features in predicting the treatment outcome of high-intensity focused ultrasound (HIFU) ablation with an immediate nonperfused volume (NPV) ratio of at least 90%.PATIENTS AND METHODS: Seventy-three women who underwent HIFU treatment were di-vided into groups A (n=47) and B (n=26), com-prising patients with an NPV ratio of at least 90% and <90%, respectively. An ensemble fea-ture ranking model was introduced based on the score values assigned to the features by five dif-ferent ML classifiers to determine the most in-formative mpMRI features. The relationship be-tween the mpMRI features and the immediate NPV ratio of 90% was evaluated using Pear-son's correlation coefficients. The diagnostic ability of the ML classifiers was evaluated using standard performance metrics, including the ar-ea under the receiver operating characteristic curve, accuracy, sensitivity, and specificity in eight folds cross-validation.RESULTS: For all the 12 most informative fea-tures, the area under receiver operating char-acteristic curve (AUROC), accuracy, specifici-ty, and sensitivity ranged from 0.5 to 0.97, 0.34 to 0.97, 0.56 to 1.0, and 0.87 to 1.0, respectively. The gradient boosting (GBM) classifier demon-strated the best predictive performance with an AUROC of 0.95 and accuracy of 0.92, followed by the random forest, AdaBoost, logistic re-gression, and support vector classifiers, which yielded an AUROC of 0.92, 0.92, 0.83, and 0.78 and accuracy of 0.96, 0.88, 0.84, and 0.84, re-spectively. GBM had the best classifier perfor-mance with the best performing features from each mpMRI group, Ktrans ratio of the fibroid to the myometrium, the ratio of area under the curve of the fibroid to the myometrium, subcuta-neous fat thickness, the ratio of apparent diffu-sion coefficient value of fibroid to the myometri-um, and T2-signal intensity of the fibroid.CONCLUSIONS: The preliminary findings of this study suggest that the most informative and best performing features from each mpMRI group should be considered for predicting the treatment outcome of HIFU ablation to achieve an immediate NPV ratio of 90%.