Predictive value of magnetic resonance imaging diffusion parameters using artificial intelligence in low-and intermediate-risk prostate cancer patients treated with stereotactic ablative radiotherapy: A pilot study

Abstract

Introduction: To investigate the predictive value of the pre-treatment diffusion parameters of diffusion-weighted magnetic resonance imaging (DW-MRI) using artificial intelligence (AI) for prostate-specific antigen (PSA) response in patients with low- and intermediate-risk prostate cancer (PCa) treated with stereotactic ablative radiotherapy (SABR). Methods: Retrospective evaluation was performed for 30 patients using pre-treatment multi-parametric MR image datasets between 2017 and 2021. MR-based mean- and minimum apparent diffusion coefficients (ADCmean, ADCmin) were calculated for the intraprostatic dominant lesion. Therapeutic response was assessed using PSA levels. Predictive performance was assessed by the receiver operating characteristic (ROC) analysis. Statistics performed with a significance level of p ≤ 0.05. Results: No biochemical relapse was detected after a median follow-up of twenty-three months (range: 3–50), with a median PSA of 0.01 ng/ml (range: 0.006–2.8) at the last examination. Significant differences were observed between the pre-treatment ADCmean, ADCmin parameters, and the group averages of patients with low and high 1-year-PSA measurements (p < 0.0001, p < 0.0001). In prediction, the random forest (RF) model outperformed the decision tree (DT) and support vector machine (SVM) models by yielding area under the curves (AUC), with 0.722, 0.685, and 0.5, respectively. Conclusion: Our findings suggest that pre-treatment MR diffusion data may predict therapeutic response using the novel approach of machine learning in PCa patients treated with SABR. Implications for practice: Clinicians shall measure and implement the evaluation of the suggested parameters (ADCmin, ADCmean) to provide the most accurate therapy for the patient.

Keywords ADC, Machine learning, Multiparametric, Prediction models, Predictive