Deep Learning-Based Methods for Prostate Segmentation in Magnetic Resonance Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate prostate segmentation on MRI is crucial for adaptive radiotherapy and radiomics.
• Manual segmentation is time-consuming and prone to variability.
• Deep learning offers automated solutions for segmentation tasks.

Purpose of the Study:

• To evaluate three deep learning models (UNet, ENet, ERFNet) for automated 3D prostate segmentation on T2-weighted MRI.
• To compare their accuracy, speed, and performance on limited datasets.
• To assess the feasibility of using these models in clinical settings with potential hardware limitations.

Main Methods:

• Three deep learning models (UNet, ENet, ERFNet) were applied to 85 manual prostate MRI segmentations.
• K-fold cross-validation and Tversky loss function were utilized for model training and evaluation.
• Segmentation accuracy (Dice Similarity Coefficient) and time were measured on CPU hardware.

Main Results:

• ENet and UNet demonstrated higher accuracy than ERFNet.
• ENet achieved a Dice Similarity Coefficient of 90.89% with a segmentation time of approximately 6 seconds on CPU.
• ENet was significantly faster than UNet while maintaining high accuracy.

Conclusions:

• ENet is a highly accurate and efficient deep learning model for automated prostate MRI segmentation.
• Its speed and performance on limited datasets make it suitable for clinical applications, potentially personalizing patient management.
• This automated segmentation can support adaptive radiotherapy and radiomics research.