Development and Validation of a Machine Learning-Based Radiomics Model Using Ultrasound Image Features for Prostate Cancer Risk Stratification.

Authors

Abstract

This study aimed to construct a risk stratification model for prostate cancer (PCa) ultrasound imaging data and machine learning algorithms, with the goal of providing an effective tool for early diagnosis, personalized treatment, and clinical decision-making. A total of 211 histopathologically confirmed PCa patients were retrospectively enrolled and categorized into low-risk (<i>n</i> = 65), intermediate-risk (<i>n</i> = 55), and high-risk (<i>n</i> = 91) groups based on prostate-specific antigen levels, Gleason scores, and clinical T stage. From ultrasound images, 135 quantitative radiomic features-including morphological, texture, and edge descriptors-were extracted using the PyRadiomics toolkit. Feature dimensionality was reduced using the Pearson correlation coefficient (PCC), followed by recursive feature elimination (RFE) with 10-fold nested cross-validation to select the most informative features. Three machine learning algorithms-support vector machine (SVM), random forest (RF), and logistic regression (LR)-were trained and evaluated. Model performance was assessed using accuracy, sensitivity, specificity, and area under the curve (AUC). The RF model achieved the best performance in both training and test cohorts, with AUCs of 0.87 and 0.86, and accuracies of 90% and 88%, respectively. DeLong's test confirmed that RF significantly outperformed SVM (<i>p</i> = 0.016) and LR (<i>p</i> = 0.004) in AUC comparison. The RF model also demonstrated robust predictive ability across risk subgroups: in the high-risk group, it achieved an AUC of 0.89, accuracy of 89%, sensitivity of 88%, and specificity of 90%; in the intermediate- and low-risk groups, AUCs were 0.86 and 0.81, respectively. Feature importance analysis revealed that wavelet-transformed Gray Level Dependence Matrix (GLDM) texture features, particularly DependenceEntropy and DependenceVariance, were the most predictive, highlighting the value of intratumoral textural heterogeneity in risk classification. The RF-based ultrasound radiomics model enables accurate stratification of PCa risk, with remarkable performance in identifying high-risk patients.

Tags

Topics