Machine learning using T2-weighted radiomics for International Society of Urological Pathology (ISUP) classification in prostate cancer.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology and Diagnostic Imaging, Hospital Israelita Albert Einstein, São Paulo, São Paulo, Brazil; Department of Medicine, Albert Einstein Israeli Faculty of Health Sciences (FICSAE), São Paulo, São Paulo, Brazil. Electronic address: [email protected].
- Department of Software Engineering, Paulista Faculty of Informatics and Administration (FIAP), São Paulo, São Paulo, Brazil.
- Department of Computer Science, University of São Paulo (USP), São Paulo, São Paulo, Brazil.
- Department of Radiology and Diagnostic Imaging, Hospital Israelita Albert Einstein, São Paulo, São Paulo, Brazil; Department of Medicine, Albert Einstein Israeli Faculty of Health Sciences (FICSAE), São Paulo, São Paulo, Brazil.
- Department of Radiology and Diagnostic Imaging, Hospital Israelita Albert Einstein, São Paulo, São Paulo, Brazil; Center of Interventional Radiology, Hospital Israelita Albert Einstein, São Paulo, São Paulo, Brazil.
Abstract
To evaluate the performance of an ARARAT radiomics-based artificial intelligence model using T2-weighted magnetic resonance imaging (MRI) for posthistological stratification of prostate cancer aggressiveness. A retrospective analysis of 112 biopsy-proven prostate cancer lesions (International Society of Urological Pathology [ISUP] <3 vs ≥ 3) from the PROSTATEx dataset was performed. Fixed 5-mm spherical region of interests (ROIs) were placed on axial T2-weighted MRI at radiologist-defined lesion sites, and 107 radiomic features were extracted using PyRadiomics. Machine-learning models were trained with 10-fold cross-validation and patient-level holdout testing. Model performance and feature importance of ARARAT were assessed using standard classification metrics and SHapley Additive exPlanations (SHAP). The cohort comprised 112 lesions, including 77 ISUP < 3 (68.8%) and 35 ISUP ≥ 3 (31.2%). ISUP 1 and 2 accounted for 32.1% and 36.6% of cases, respectively, while ISUP 3, 4, and 5 represented 17.9%, 7.1%, and 6.3%, respectively. Lesions were located in the peripheral zone (44.6%), anterior stroma (40.2%), and transition zone (15.2%), with no association between anatomical zone and ISUP category (P = 0.639). The random forest achieved the highest performance on the holdout set, with an area under the curve (AUC) of 0.84 (95% confidence interval [CI]: 0.63-0.98), average precision (AP) 0.70 (95% CI: 0.35-0.97), sensitivity of 85.7%, specificity of 78.6%, accuracy of 81.0%, and an F1-score of 0.75. The Brier score was 0.18, and calibration improved after isotonic regression, reducing the expected calibration error from 0.168 to 0.099. First-order and grey-level size zone matrix features were the dominant predictors. T2-weighted MRI-based radiomics enables accurate posthistological stratification of prostate cancer aggressiveness, supporting further external validation.