Development and Validation of Interpretable Machine Learning Models for Clinically Significant Prostate Cancer Diagnosis in Patients With Lesions of PI‐RADS v2.1 Score ≥3

• Published in: Journal of magnetic resonance imaging Vol. 60; no. 5; pp. 2130 - 2141
• Main Authors: Ruan, Mingjian, Liu, Yi, Yao, Kaifeng, Wang, Kexin, Fan, Yu, Wu, Shiliang, Wang, Xiaoying
• Format: Journal Article
• Language: English
• Published: Hoboken, USA John Wiley & Sons, Inc 01.11.2024; Wiley Subscription Services, Inc
• Subjects: Aged; Biopsy; Clinical significance; Decision trees; Diffusion Magnetic Resonance Imaging - methods; Diffusion rate; Echo-Planar Imaging - methods; Field strength; Humans; Image Interpretation, Computer-Assisted - methods; Learning algorithms; Lesions; Machine Learning; Magnetic Resonance Imaging - methods; Male; Medical imaging; Middle Aged; Parameter identification; Prostate - diagnostic imaging; Prostate - pathology; prostate biopsy; Prostate cancer; Prostate-Specific Antigen - blood; Prostatic Neoplasms - diagnostic imaging; Prostatic Neoplasms - pathology; random forest; Regression analysis; Reproducibility of Results; Retrospective Studies; Risk Assessment; ROC Curve; Sensitivity analysis; Statistical analysis; Statistical tests; Support vector machines; prostate biopsy; prostate cancer; random forest; machine learning
• DOI: 10.1002/jmri.29275
• ISSN: 1053-1807; 1522-2586

Background For patients with PI‐RADS v2.1 ≥ 3, prostate biopsy is strongly recommended. Due to the unsatisfactory positive rate of biopsy, improvements in clinically significant prostate cancer (csPCa) risk assessments are required. Purpose To develop and validate machine learning (ML) models based on clinical and imaging parameters for csPCa detection in patients with PI‐RADS v2.1 ≥ 3. Study Type Retrospective. Subjects One thousand eighty‐three patients with PI‐RADS v2.1 ≥ 3, randomly split into training (70%, N = 759) and validation (30%, N = 324) datasets, and 147 patients enrolled prospectively for testing. Field Strength/Sequence 3.0 T scanners/T2‐weighted fast spin echo sequence and DWI with diffusion‐weighted single‐shot gradient echo planar imaging sequence. Assessment The factors evaluated for csPCa detection were age, prostate specific antigen, prostate volume, and the diameter and location of the index lesion, PI‐RADSv2.1. Five ML models for csPCa detection were developed: logistic regression (LR), extreme gradient boosting, random forest (RF), decision tree, and support vector machines. The csPCa was defined as Gleason grade ≥2. Statistical Tests Univariable and multivariable LR analyses to identify parameters associated with csPCa. Area under the receiver operating characteristic curve (AUC), Brier score, and DeLong test were used to assess and compare the csPCa diagnostic performance with the LR model. The significance level was defined as 0.05. Results The RF model exhibited the highest AUC (0.880–0.904) and lowest Brier score (0.125–0.133) among the ML models in the validation and testing cohorts, however, there was no difference when compared to the LR model (P = 0.453 and 0.548). The sensitivity and negative predictive values in the validation and testing cohorts were 93.8%–97.6% and 82.7%–95.1%, respectively, at a threshold of 0.450 (99% sensitivity of the RF model). Data Conclusion The RF model might help for assessing the risk of csPCa and preventing overdiagnosis and unnecessary biopsy for men with PI‐RADSv2.1 ≥ 3. Evidence Level 3 Technical Efficacy Stage 2