Fetal Weight Percentile Classification Across Gestational Weeks by Comparing Machine-Learning Algorithms Using Ultrasound Images

Abstract

Background: Accurate assessment of the fetal weight percentile is an important component of prenatal care, as it enables clinicians to monitor fetal growth and identify fetuses at risk of growth restriction or macrosomia. However, conventional ultrasound-based assessment may be affected by measurement variability and operator dependence. Radiomics-based machine-learning approaches may provide a more objective and reproducible framework for classifying fetal weight percentiles. Objectives: This study aimed to compare the performance of several machine-learning algorithms for classifying fetal weight percentile categories using ultrasound images. Methods: This analytical retrospective study was conducted at Kermanshah University of Medical Sciences using archived ultrasound data from 200 pregnant women collected during 1401 - 1402. Ultrasound images were preprocessed and denoised in MATLAB using four filters: Butterworth, Ideal, Median, and Wavelet. Regions of interest corresponding to the fetal head, abdomen, and femur were identified, and 1715 radiomics features per case were extracted using 3D Slicer. To reduce dimensionality and improve model robustness, principal component analysis was applied before classification. Machine-learning models were developed in the MATLAB Classification Learner Toolbox, including an ensemble model, a support vector machine, k-nearest neighbors, and an artificial neural network. Model performance was evaluated using 10-fold cross-validation. Evaluation metrics included accuracy, precision, sensitivity, specificity, F1-score, and the area under the receiver operating characteristic curve. Results: Among the evaluated models, the ensemble model demonstrated the highest internal performance, with an accuracy of 0.95 and an area under the receiver operating characteristic curve of 0.93. The accuracies of the k-nearest neighbors, support vector machine, and neural network models were 0.93, 0.92, and 0.91, respectively, with corresponding area under the receiver operating characteristic curve values of 0.92, 0.91, and 0.90, respectively. Conclusions: These findings indicate that radiomics-based machine-learning models, particularly the ensemble model, show promising performance for classifying fetal weight percentile categories from ultrasound images. Nevertheless, because this study was retrospective and relied solely on internal validation, further prospective and external validation is required before clinical application. Accordingly, the proposed model should be considered a decision-support tool rather than a replacement for clinical judgment.