Abstract
Multiparametric MRI (mpMRI) is a key, non-invasive imaging approach used to identify and localize prostate cancer (PCa). When integrated with radiomic feature extraction, mpMRI data can assist in estimating tumor aggressiveness. Although T2 mapping delivers quantitative metrics useful for PCa assessment, it is not yet routinely implemented in clinical imaging workflows. Our team previously introduced a deep learning framework capable of generating estimated T2 maps from standard T1- and T2-weighted sequences. This study explores the incremental diagnostic benefit of incorporating those estimated T2 maps with conventional T2-weighted scans for identifying clinically significant prostate cancer (csPCa). A total of 76 peripheral zone lesions, covering both clinically significant and insignificant PCa cases, were retrospectively assessed. Radiomic features were obtained from standard T2-weighted scans and the AI-generated T2 maps. Feature selection and model development were performed using five-fold cross-validation. Logistic regression and Gaussian Process classifiers were applied, and diagnostic performance was evaluated through area under the curve (AUC) and accuracy metrics. The combined approach using both T2-weighted and estimated T2 map features achieved an AUC of 0.803, which was significantly higher than that obtained from models using only T2-weighted features (AUC = 0.700, p = 0.048). Deep learning-based T2 map features offer complementary quantitative data that enhance the prediction accuracy for peripheral zone csPCa, supporting improved risk evaluation in non-invasive prostate cancer diagnostics.
