A feasibility study on an automated method to generate patient-specific dose distributions for radiotherapy using...

Area of Science:

• Medical Physics
• Radiotherapy
• Deep Learning

Background:

• Intensity-modulated radiation therapy (IMRT) requires precise dose distributions for effective cancer treatment.
• Optimizing dose distributions is complex and time-consuming.
• Deep learning offers potential for automating and improving treatment planning.

Purpose of the Study:

• To develop a deep learning method for predicting optimal radiation therapy dose distributions.
• To utilize a database of existing IMRT treatment plans for model training.
• To predict patient-specific dose distributions based on planning images and segmented anatomy.

Main Methods:

• A dataset of 80 early-stage nasopharyngeal cancer (NPC) cases was used, with 70 for training and 10 for testing.
• ResNet101 deep learning network was employed to predict dose maps (coarse and fine) from input images.
• Two input types were tested: raw images with structures and images with added radiation beam geometry information.

Main Results:

• The deep learning model accurately predicted patient-specific dose distributions using both input methods.
• The model incorporating radiation geometry showed improved performance in predicting out-of-field dose distributions.
• Gamma analysis indicated comparable accuracy for most organs at risk, with minor differences for optic pathways.

Conclusions:

• The proposed deep learning system, enhanced with radiation geometry, shows promise for generating patient-specific dose distributions.
• This method can be applied to radiotherapy quality assurance and automated treatment planning.
• The approach facilitates slice-by-slice dose distribution generation for improved planning workflows.