Establishment of a predictive model for benign and malignant prostatic peripheral lesions by combined TRUS and CEUS
SUN Ya, SHI Yanping, LI Huarong, WANG Jiajun, XU Jingxu, LIANG Lei
2023, 43(9):
1432-1438.
doi:10.16352/j.issn.1001-6325.2023.09.1432
Asbtract
(
87 )
PDF (4457KB)
(
85
)
References |
Related Articles |
Metrics
Objective To construct a machine learning model based on radiomics features by combing conventional B-mode transrectal ultrasound (TRUS) and contrast-enhanced ultrasound (CEUS)in order to improve prostate cancer (PCa) detection in the peripheral zone, and to explore its clinical application value. Methods A retrospective study of 109 cases (45 benign, 64 malignant lesions) with targeted biopsy-confirmed pathology who underwent B-mode and CEUS examinations was performed. Time-intensity curves were obtained using Sonoliver software forall lesions in regions of interest. The following parameters were collected: the maximum intensity (IMAX), the rise time (RT), the time to peak (TTP), and the mean transit time (MTT). Clinical risk factors, including age, serum total prostate-specific antigen (tPSA), free PSA(fPSA), f/t PSA, prostate volume and prostate-specific antigen density (PSAD), were collected. Feature selection and model construction were carried out by feature correlation analysis and multivariable Logistic regression analysis, respectively. Results Finally, 32 features were screened out by the radiomics models. Three radiomics models (B model, CEUS model and B-CEUS combined model) were established by logistic regression, and radiomics scores were obtained at the same time. The risk factors like age, PSAD, and RT were significant independent predictors (P<0.05). The diagnostic accuracy of B model, CEUS model, B-CEUS combined model, risk factor model and risk factor-radiomics combined model were 0.75, 0.71, 0.73, 0.70 and 0.84, respectively, and the area under the curves(AUC) were 0.79, 0.75, 0.84, 0.79 and 0.91, respectively. The risk factor-radiomics combined model had a significant significance in the validation cohort (P<0.05). Conclusions The machine learning model combined with conventional B-mode TRUS, CEUS, and relevant clinical factors can better predict the benign and malignant prostate peripheral lesions, which may significantly support clinical decision-making.