Multiparametric MRI and Machine Learning Based Radiomic Models for Preoperative Prediction of Multiple Biological Characteristics in Prostate Cancer

ObjectivesThis study aims to develop and evaluate multiparametric MRI (MP-MRI)-based radiomic models as a noninvasive diagnostic method to predict several biological characteristics of prostate cancer. MethodsA total of 252 patients were retrospectively included who underwent radical prostatectomy and MP-MRI examinations. The prediction characteristics of this study were as follows: Ki67, S100, extracapsular extension (ECE), perineural invasion (PNI), and surgical margin (SM). Patients were divided into training cohorts and validation cohorts in the ratio of 4:1 for each group. After lesion segmentation manually, radiomic features were extracted from MP-MRI images and some clinical factors were also included. Max relevance min redundancy (mRMR) and recursive feature elimination (RFE) based on random forest (RF) were adopted to select features. Six classifiers were included (SVM, KNN, RF, decision tree, logistic regression, XGBOOST) to find the best diagnostic performance among them. The diagnostic efficiency of the construction models was evaluated by ROC curves and quantified by AUC. ResultsRF performed best among the six classifiers for the four groups according to AUC values (Ki67 = 0.87, S100 = 0.80, ECE = 0.85, PNI = 0.82). The performance of SVM was relatively the best for SM (AUC = 0.77). The number and importance of DCE features ranked first in the models of each group. The combined models of MP-MRI and clinical characteristics showed no significant difference compared with MP-MRI models according to Delong's tests. ConclusionsRadiomics models based on MP-MRI have the potential to predict biological characteristics and are expected to be a noninvasive method to evaluate the risk stratification of prostate cancer.

Automated summary

This study developed and evaluated radiomic models based on multiparametric MRI as a noninvasive diagnostic method to predict various biological characteristics of prostate cancer. The results showed that the RF-based model performed best among the six classifiers for all the groups.