Research on Ultrasound Image-Assisted Diagnosis of Prostate Cancer Based on Machine Learning.

Abstract

Background: This study aimed to explore the diagnostic value of a machine learning model based on ultrasound image features for prostate cancer.

Methods: 600 patients with prostate tumours detected by transrectal ultrasound (TRUS) at Quzhou People's Hospital between July 2023 and June 2025, and with corresponding pathological results, were selected. Based on these results, the patients were divided into prostate cancer and benign lesion groups. Stratified random sampling was then used to divide these groups into a training group (n = 420) and a validation group (n = 180) at a ratio of 7:3. Regions of interest were manually delineated by sonographers and radiomics features were extracted using Pyradiomics software. Using pathological results as the gold standard, the radiomics features with the highest diagnostic value in differentiating between benign and malignant lesions were selected. Based on the selected features, a support vector machine (SVM) algorithm model was then constructed, and the efficacy of the SVM and fusion models in diagnosing prostate cancer was evaluated using receiver operating characteristic curves.

Results: In the validation group, the SVM model based solely on ultrasound features achieved an accuracy of 73.06%, a sensitivity of 82.59%, a specificity of 65.50% and the area under the curve (AUC) of 0.729 [95% confidence interval (CI): 0.666-0.792] in diagnosing prostate cancer. After incorporating clinical features such as age, total prostate-specific antigen and prostate volume, the combined model improved the accuracy to 85.00%, sensitivity to 86.61%, specificity to 82.35%, and the AUC to 0.824 (95% CI: 0.785-0.863). Calibration and decision curve analyses further confirmed the model's good calibrability and net clinical benefit.

Conclusions: A machine learning fusion model based on whole-gland TRUS radiomics features and combined with clinical indicators demonstrates good performance in estimating patient-level risk for prostate cancer. It may serve as a potential non-invasive decision support tool for risk stratification, but its clinical utility requires further external validation.