Multi-constraint generative adversarial network for dose prediction in radiotherapy

Area of Science:

• Medical Physics
• Artificial Intelligence in Oncology
• Radiotherapy Planning

Background:

• Radiation therapy (RT) is a cornerstone of cancer treatment, necessitating precise dose delivery to target volumes while sparing organs at risk.
• Deep learning models are increasingly used to predict radiation dose distributions, aiming to enhance treatment planning efficiency and accuracy.

Purpose of the Study:

• To introduce Mc-GAN, a novel multi-constraint generative adversarial network for automated dose distribution prediction.
• To improve the accuracy and reliability of dose prediction in radiation therapy planning.

Main Methods:

• Developed a generative adversarial network (Mc-GAN) incorporating a UNet-like generator with dilated convolutions for comprehensive feature extraction.
• Integrated a dual attention module (DAM) to enhance focus on semantic relevance during feature extraction.
• Introduced locality-constrained loss (LCL) and self-supervised perceptual loss (SPL) alongside traditional losses to refine prediction accuracy.

Main Results:

• Mc-GAN demonstrated superior performance compared to state-of-the-art methods on two in-house datasets.
• The model achieved significant improvements across planning target volume (PTV) and organs at risk (OARs) criteria.
• Evaluations confirmed the effectiveness of the novel loss functions in dose prediction accuracy.

Conclusions:

• The proposed Mc-GAN model offers a robust and accurate solution for automated dose distribution prediction in radiation therapy.
• This deep learning approach holds potential for optimizing clinical treatment planning and improving patient outcomes.
• Mc-GAN's multi-constraint strategy effectively addresses the complexities of dose prediction in radiotherapy.