Registration-guided deep learning image segmentation for cone beam CT-based online adaptive radiotherapy.

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Online adaptive radiotherapy (ART) requires accurate organ at risk (OAR) segmentation on low-quality cone beam computed tomography (CBCT) images.
• Deep learning (DL) segmentation models struggle with CBCT's image quality and limited training data.

Purpose of the Study:

• To propose and evaluate a registration-guided deep learning (RgDL) framework for improved online CBCT segmentation.
• To overcome challenges of low image quality and insufficient training data in DL-based CBCT segmentation.

Main Methods:

• Developed a RgDL framework integrating image registration with DL segmentation (U-Net).
• Implemented two versions: Rig-RgDL (rigid registration) and Def-RgDL (deformable image registration).
• Trained and evaluated on a head-and-neck dataset, comparing against registration-only and DL-only methods.

Main Results:

• RgDL frameworks achieved higher segmentation accuracy (Dice Similarity Coefficient) than baseline methods.
• Def-RgDL yielded an average DSC of 86.5%, outperforming deformable image registration (DIR) and DL alone.
• RgDL demonstrated robustness to limited training data, with significantly smaller performance drops compared to DL alone.

Conclusions:

• The RgDL framework effectively integrates patient-specific registration guidance with DL models for superior online CBCT segmentation.
• RgDL overcomes low image quality and data limitations, showing promise for enhancing online adaptive radiotherapy accuracy and efficiency.