Sharp loss: a new loss function for radiotherapy dose prediction based on fully convolutional networks.

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Neural networks are common for radiotherapy dose prediction.
• Existing methods struggle with dose imbalance, reducing accuracy.
• A new loss function is proposed to overcome this limitation.

Purpose of the Study:

• To introduce and evaluate a novel 'sharp loss' function for improved radiotherapy dose prediction.
• To mitigate the impact of dose imbalance on prediction accuracy.
• To compare the performance of the sharp loss function against mean square error (MSE) loss.

Main Methods:

• Utilized the U-Net architecture for dose prediction modeling.
• Developed a novel 'sharp loss' function with an adjustable parameter gamma (γ).
• Trained and tested the model on a dataset of 110 left-breast cancer patients treated with volumetric-modulated arc radiotherapy.

Main Results:

• The sharp loss function demonstrated superior dose prediction accuracy over MSE loss.
• With γ=100, sharp loss significantly reduced mean absolute differences for planning target volume and organs at risk.
• Specific improvements were noted for the planning target volume, ipsilateral lung, heart, and spinal cord.

Conclusions:

• The proposed sharp loss function effectively enhances the accuracy of radiotherapy dose prediction.
• This novel approach offers a promising solution for overcoming dose imbalance challenges in radiotherapy planning.