Efficient segmentation using domain adaptation for MRI-guided and CBCT-guided online adaptive radiotherapy

Area of Science:

• Medical Imaging
• Radiotherapy
• Machine Learning

Background:

• Accurate delineation of regions of interest (ROIs) is crucial for adaptive radiotherapy (ART).
• Current manual segmentation methods are time-consuming and labor-intensive, hindering efficient ART workflows.

Purpose of the Study:

• To develop efficient and accurate segmentation methods for MRI-guided ART (MRIgART) and CBCT-guided ART (CBCTgART).
• To improve the speed and precision of ROI delineation in ART using advanced deep learning techniques.

Main Methods:

• Two domain adaptation methods were designed to transfer features from planning CT (pCT) to MRI or CBCT.
• Generalized deep learning models were trained on large datasets and then personalized for individual patients via fine-tuning.
• The proposed method was compared against deformable image registration (DIR) and standard deep learning approaches.

Main Results:

• The proposed method demonstrated superior or comparable performance across different ART modalities and patient cohorts.
• For MRIgART, the mean Dice Similarity Coefficient (DSC) reached 92.20% for the proposed method.
• For CBCTgART, the proposed method achieved mean DSCs of 91.18% and 81.56% for nasopharyngeal and pancreatic cancer patients, respectively.

Conclusions:

• Personalized modeling significantly enhances segmentation accuracy in adaptive radiotherapy.
• The developed methods offer a more efficient and precise approach to ROI delineation for ART.