U-Net Architecture for Prostate Segmentation: The Impact of Loss Function on System Performance

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Prostate segmentation from MRI is crucial for radiotherapy planning.
• Automating segmentation improves accuracy and efficiency.
• Deep learning model performance depends on hyper-parameter tuning, including loss functions.

Purpose of the Study:

• To investigate the impact of various loss functions on deep learning-based prostate segmentation.
• To compare the performance of nine different loss functions using a U-Net model.

Main Methods:

• A U-Net model was trained for prostate segmentation on T2-weighted MRI data.
• Nine loss functions were evaluated: BCE, IoU, Dice, BCE+Dice, W(BCE+Dice), Focal, Tversky, Focal Tversky, and Surface loss.
• Model performance was assessed using a five-fold cross-validation and metrics like Dice Similarity Coefficient (DSC) and 95% Hausdorff Distance (95HD).

Main Results:

• Performance ranking varied by metric, but W(BCE+Dice) and Focal Tversky generally showed strong results (e.g., DSC 0.71-0.74).
• Surface loss consistently ranked lowest (e.g., DSC 0.40).
• Segmentation accuracy was lower for the prostate apex and base compared to the mid-gland.

Conclusions:

• The choice of loss function significantly influences the performance of deep learning models for prostate segmentation.
• Compound loss functions tend to outperform single loss functions for this task.
• Further research may optimize loss functions for specific regions like the prostate apex and base.