Automated prostate multi-regional segmentation in magnetic resonance using fully convolutional neural networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate prostate MRI segmentation is crucial for evaluating prostate cancer, aiding in PSA density calculation and PI-RADS v2.1 assessments.
• Automated segmentation of central-transition zone (CZ-TZ), peripheral zone (PZ), and seminal vesicles (SV) supports the identification of clinically significant cancer.

Purpose of the Study:

• To develop a robust and reproducible Convolutional Neural Network (CNN)-based model for automatic multi-regional prostate MRI segmentation.
• To validate the model's generalizability using an intercontinental cohort of prostate MRI data.

Main Methods:

• A U-Net-based CNN model with deep supervision and cyclical learning rates was trained on T2-weighted prostate MRI data.
• The dataset comprised 243 studies from 7 countries, with manual delineations by radiologists serving as ground truth.
• Model performance was assessed using the Dice Similarity Coefficient (DSC), with DSC > 0.7 considered accurate.

Main Results:

• The model achieved high DSC scores: 0.88 ± 0.01 for the prostate gland, 0.85 ± 0.02 for CZ-TZ, 0.72 ± 0.02 for PZ, and 0.72 ± 0.02 for SV.
• No statistically significant performance differences were observed across different MRI machine vendors or continents.
• The segmentation accuracy met the predefined threshold for clinical utility.

Conclusions:

• CNN-based automatic segmentation of multi-regional prostate T2-weighted MR images is accurate and reproducible.
• The developed model demonstrates strong generalizability in varied clinical settings, including diverse equipment and populations.
• This automated segmentation technique can significantly aid in the diagnosis and follow-up of prostate cancer patients.