An unsupervised dual contrastive learning framework for scatter correction in cone-beam CT image

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Machine Learning in Healthcare

Background:

• Cone-beam computed tomography (CBCT) is essential in radiotherapy but suffers from scatter artifacts, degrading image quality.
• These artifacts limit the diagnostic and therapeutic applications of CBCT images.
• Effective scatter correction is vital for enhancing CBCT utility.

Purpose of the Study:

• To develop and evaluate an unsupervised contrastive learning method for scatter correction in CBCT.
• To improve the image quality of CBCT by reducing scatter artifacts.
• To assess the potential for clinical adoption of the proposed scatter correction technique.

Main Methods:

• Proposed an unsupervised contrastive learning framework for CBCT scatter correction.
• Generated synthetic planning CT (spCT) images from low-quality CBCT.
• Extracted scatter artifacts by analyzing projection differences and applying low-pass filtering.
• Reconstructed corrected CBCT (cCBCT) images using the FDK algorithm.

Main Results:

• The corrected CBCT (cCBCT) images demonstrated improved anatomical consistency, CT number accuracy, and spatial homogeneity compared to original CBCT.
• Significant reduction in Mean Absolute Error (MAE) for overall image data, fat ROIs, and muscle ROIs.
• The proposed method outperformed conventional unsupervised synthetic image generation techniques.

Conclusions:

• The unsupervised contrastive learning approach effectively enhances CBCT image quality by correcting scatter artifacts.
• The developed method shows promising potential for integration into clinical radiotherapy workflows.
• Further validation and clinical studies are warranted to confirm its utility.