An Unsupervised Learning-Based Regional Deformable Model for Automated Multi-Organ Contour Propagation

Area of Science:

• Medical Physics
• Radiotherapy
• Medical Imaging

Background:

• Adaptive radiation therapy requires accurate and timely contour propagation.
• Cone-beam computed tomography (CBCT) is used for image guidance in radiotherapy.
• Traditional image registration methods can be affected by CBCT artifacts.

Purpose of the Study:

• To evaluate a deep unsupervised learning-based regional deformable model for automatic contour propagation.
• To assess the model's performance in mapping organs at risk and target volumes from planning CT to CBCT.
• To mitigate the impact of CBCT artifacts on contour accuracy.

Main Methods:

• A deep unsupervised learning model was developed for contour mapping.
• A regional deformable framework utilizing narrow-band mapping was employed.
• The model was trained, validated, and tested on 373 retrospective CBCT volumes from 111 breast cancer patients.

Main Results:

• The model achieved good agreement with manual reference segmentations.
• Mean Dice scores for key structures ranged from 0.78 (tumor bed) to 0.95 (right lung).
• The proposed method demonstrated robustness against CBCT image artifacts.

Conclusions:

• The deep learning-based regional deformable model effectively automates contour propagation for breast cancer adaptive radiotherapy.
• This technique shows significant promise for enhancing radiotherapy planning and delivery.
• Further research is warranted to fully explore the potential of deep learning in contour propagation.