Multi-regional prostate segmentation tool

Multi-regional prostate segmentation tool performs automatic segmentation of the prostate into central-transition zone (CZ+TZ), peripheral zone (PZ), and seminal vesicle (SV) on T2-weighted MRI to support quantitative imaging and biomarker extraction.

Key Features:

• Automatic Segmentation: Automates segmentation of prostate T2-weighted MRI into CZ+TZ, PZ, and SV.
• Deep Learning Architecture: Employs a U-Net-based convolutional neural network (CNN) with deep supervision.
• Cyclical Learning Rate: Uses a cyclical learning rate strategy during training to optimize performance.
• Training Dataset: Trained on 243 T2-weighted prostate MRI studies from an intercontinental cohort spanning seven countries and ten machines from three vendors.
• Ground Truth Annotation: Ground truth established by manual delineation performed by two experienced radiologists.
• Performance Evaluation: Evaluated using the dice similarity coefficient (DSC) with reported DSCs of 0.88 ± 0.01 for the prostate gland, 0.85 ± 0.02 for CZ-TZ, and 0.72 ± 0.02 for PZ and SV on a test set of 120 studies.
• Generalization and Robustness: Validated across different manufacturers and geographic locations with no statistically significant differences in performance.

Scientific Applications:

• Prostate Cancer Evaluation: Facilitates calculation of automated PSA density and other imaging biomarkers for prostate cancer assessment.
• Clinical Guidelines Compliance: Supports evaluation according to PI-RADS v2.1 guidelines for identifying clinically significant cancers.
• Diagnostic Aid: Assists diagnosis and monitoring by providing precise segmentation outputs for radiologists and clinicians.

Methodology:

The model was trained using a U-Net-based CNN with deep supervision and a cyclical learning rate on 243 T2-weighted MRI studies (seven countries, ten machines, three vendors); ground truth was manual delineation by two experienced radiologists; evaluation used the dice similarity coefficient (DSC) on a test set of 120 studies yielding DSCs of 0.88 ± 0.01 (prostate), 0.85 ± 0.02 (CZ-TZ), and 0.72 ± 0.02 (PZ and SV).