Robust Real-time Segmentation of Bio-Morphological Features in Human Cherenkov Imaging during Radiotherapy via Deep...

Area of Science:

• Medical Imaging
• Radiation Oncology
• Artificial Intelligence

Background:

• Cherenkov imaging visualizes radiation delivery in real-time.
• Patient-specific bio-morphological features (e.g., vasculature) are crucial for precise radiation therapy (RT).
• Conventional image processing for feature segmentation is slow and lacks accuracy.

Purpose of the Study:

• To develop the first deep learning framework for real-time bio-morphological feature segmentation in Cherenkov images.
• To enable patient positioning verification and motion management in RT.
• To overcome limitations of slow and inaccurate conventional segmentation methods.

Main Methods:

• A ResNet segmentation framework was pre-trained using a large fundus photography dataset with vessel annotations.
• Transfer learning was applied, fine-tuning the model on a smaller Cherenkov image dataset with annotated vasculature.
• The deep learning framework was tested on 19 breast cancer patients.

Main Results:

• The framework achieved video frame rate processing, with segmentation completed in less than 0.7 milliseconds per instance.
• High segmentation accuracy was demonstrated, with an average Dice score of 0.85.
• Consistent and rapid segmentation of features like veins, scars, and pigmented skin was achieved across diverse patient images.

Conclusions:

• The developed deep learning framework provides a foundation for online, real-time monitoring in radiation therapy.
• This approach significantly improves speed and consistency compared to manual segmentation.
• Enables enhanced precision in patient positioning and motion management during RT.