Reliable Delineation of Clinical Target Volumes for Cervical Cancer Radiotherapy on CT/MR Dual-Modality Images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiotherapy Planning

Background:

• Accurate clinical target volume (CTV) delineation is essential for effective radiotherapy.
• Integrating magnetic resonance (MR) images with computed tomography (CT) images aids target delineation.
• Directly acquiring MR images can be challenging, necessitating alternative approaches.

Purpose of the Study:

• To develop an AI-based method for generating synthetic MR images from CT images.
• To improve CTV delineation accuracy using AI-generated MR images.
• To enhance radiotherapy planning by reducing manual workload and inter-observer variability.

Main Methods:

• An attention-guided single-loop image generation model was proposed to create high-quality MR images from CT scans.
• The image generation model incorporated an attention mechanism and enhanced loss function for improved feature extraction.
• A CTV segmentation model fused multi-scale features using image fusion and a hollow space pyramid module.

Main Results:

• The AI image generation model significantly improved image quality metrics, including PSNR (14.87 to 16.72) and SSIM (0.58 to 0.67).
• The proposed segmentation method achieved higher accuracy compared to FCN, with improved Intersection over Union (0.8360 to 0.9043) and Dice coefficient (0.8998 to 0.9473).
• Segmentation accuracy metrics, including Hausdorff distance and mean surface distance, demonstrated clinically acceptable results.

Conclusions:

• AI-driven generation of synthetic MR images from CT scans can effectively enhance CTV delineation for radiotherapy.
• The proposed attention-guided model and multi-scale feature fusion segmentation method achieve high accuracy and clinical acceptability.
• This approach has the potential to streamline radiotherapy planning, reduce physician workload, and minimize inter-observer variability.