CT-guided CBCT multi-organ segmentation using a multi-channel conditional consistency diffusion model for lung cancer
Xiaoqian Chen1, Richard L J Qiu1, Shaoyan Pan1
1Department of Radiation Oncology, Winship Cancer Institute, Emory University School of Medicine, Atlanta, GA, 30322, United States of America.
Biomedical Physics & Engineering Express
|May 20, 2025
Summary
This study introduces a new AI model, the multi-channel conditional consistency diffusion model (MCCDM), to improve the accuracy of organ-at-risk segmentation in thoracic cone beam CT images. The model enhances precision for adaptive radiotherapy planning, even with image artifacts.
Area of Science:
- Medical Imaging
- Radiotherapy
- Artificial Intelligence
Background:
- Accurate segmentation of organs-at-risk (OARs) is critical for adaptive radiotherapy, especially in thoracic cases using cone beam computed tomography (CBCT).
- CBCT image quality is often degraded by artifacts, complicating precise OAR segmentation compared to planning CT scans.
Purpose of the Study:
- To develop and evaluate a novel multi-channel conditional consistency diffusion model (MCCDM) for improved OAR segmentation in thoracic CBCT images.
- To leverage domain transfer capabilities of the model to enhance segmentation accuracy across different imaging modalities.
Main Methods:
- A multi-channel conditional consistency diffusion model (CBCT-MCCDM) was developed, trained jointly with CT images and their corresponding masks for end-to-end learning.
- The model was applied to segment esophagus, heart, left lung, right lung, and spinal cord in CBCT images from lung cancer patients undergoing stereotactic body radiation therapy (SBRT).
- Quantitative evaluation involved comparing model-generated contours with ground truth using Dice similarity coefficients (DSC), sensitivity, specificity, HD95, and MSD.
Main Results:
- The CBCT-MCCDM achieved high average Dice similarity coefficients: 0.82 (esophagus), 0.88 (heart), 0.95 (left lung), 0.96 (right lung), and 0.96 (spinal cord).
- Excellent sensitivity and specificity values were recorded across all segmented OARs, indicating robust performance.
- The proposed method outperformed two state-of-the-art methods (CBCT-only and U-Net) in segmentation accuracy.
Conclusions:
- The proposed CBCT-MCCDM effectively enhances OAR segmentation accuracy in thoracic CBCT images, overcoming common imaging artifacts.
- This model shows significant potential for improving dose verification and online replanning in adaptive radiotherapy.
- The domain transfer capability of MCCDM offers a promising approach for cross-modality segmentation tasks in medical imaging.
