Deep learning based direct segmentation assisted by deformable image registration for cone-beam CT based...

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Accurate auto-segmentation is crucial for efficient cone-beam CT (CBCT)-based online adaptive radiotherapy.
• Deep learning (DL) segmentation of CBCT faces challenges due to poor image quality and limited labeled data.
• Deformable image registration (DIR) is commonly used to transfer contours from planning CT (pCT) to CBCT.

Purpose of the Study:

• To improve DL-based direct segmentation of CBCT images by leveraging DIR techniques.
• To develop a robust method for auto-segmentation in adaptive radiotherapy, reducing physician workload.

Main Methods:

• Utilized deformed pCT contours from multiple DIR methods as pseudo-labels for initial DL model training.
• Employed DIR-derived contours as influencer volumes to guide the region of interest for DL segmentation.
• Fine-tuned the DL model using a smaller set of manually verified labels.

Main Results:

• DL-based direct CBCT segmentation without influencer volumes performed poorly compared to DIR.
• Incorporating influencer volumes significantly improved DL segmentation accuracy to DIR-based levels.
• Fine-tuning further enhanced the DL model, achieving a mean Dice score of 0.86.

Conclusions:

• Deformed pCT contours as pseudo-labels and influencer volumes enhance DL-based CBCT segmentation.
• This hybrid approach, combining DIR and DL, can outperform traditional DIR methods.
• Fine-tuning with true labels further refines DL model performance for adaptive radiotherapy applications.