Diagnostic value and external validation of biparametric magnetic resonance imaging radiomics in clinically significant prostate cancer.

BACKGROUND

Given the significant economic burden of prostate cancer (PCa), its diagnostic methods need to be improved. The limitations of the subjective Prostate Imaging Reporting and Data System (PI-RADS) underscore the need for generalizable radiomics in clinically significant PCa (csPCa). This study aimed to build a machine-learning model based on biparametric magnetic resonance imaging (bpMRI) to diagnose csPCa.

METHODS

Prognostic model development: This study retrospectively included the data of 445 patients from two centers, of whom 206 had csPCa and 239 had clinically non-significant PCa (ncsPCa). The training set comprised 120 csPCa patients and 141 ncsPCa patients. The test set comprised 52 csPCa patients and 61 ncsPCa patients. The external validation comprised 34 csPCa patients and 37 ncsPCa patients.

RESULTS

Features were extracted from T2-weighted imaging (T2WI) sequences and apparent diffusion coefficient (ADC) maps based on bpMRI radiomics. From 3662 radiomics features, 10 stable radiomics features were selected for model construction based on intraclass correlation coefficients (ICCs). Three diagnostic models for csPCa were constructed. The area under the curve (AUC) values for the PI-RADS-scoring model, which was based on visual assessments by radiologists, were 0.8271, 0.7905, and 0.8331 in the training, test, and external validation sets, respectively; while those for the clinical scoring model were 0.9236, 0.8846, and 0.8378, respectively; and those for the radiomics model were 0.9790, 0.9584, and 0.9523, respectively. There were significant differences between the radiomics model and the PI-RADS-scoring model (P0.001) in both the training and test sets. The P value for the radiomics model and clinical scoring model in the training set was <0.001, while that in the validation set was 0.056. Overall, the AUC values for the three models indicated that the diagnostic performance of the bpMRI radiomics model, which was based on T2WI sequences and ADC images, for csPCa was better than that of both the PI-RADS-scoring and clinical scoring models.

CONCLUSIONS

The radiomics model can reliably detect and classify csPCa, and is a very powerful non-invasive auxiliary tool that could be used as an alternative method for diagnosing csPCa in personalized medicine.