Improving Image Quality of Cone-Beam CT Using Alternating Regression Forest

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Machine Learning in Medicine

Background:

• Cone-Beam Computed Tomography (CBCT) is widely used in image-guided radiotherapy but suffers from lower image quality compared to planning CT.
• Improving CBCT image quality is crucial for accurate dose calculation and adaptive radiotherapy workflows.
• Current methods for CBCT correction often require manual intervention or are limited in their ability to capture complex anatomical variations.

Purpose of the Study:

• To develop and evaluate a novel learning-based method for improving CBCT image quality.
• To assess the accuracy of the proposed method in correcting CBCT images to a level comparable to planning CT.
• To demonstrate the potential of the enhanced CBCT images for quantitative use in adaptive radiotherapy.

Main Methods:

• A CBCT image quality improvement method utilizing anatomic signatures and an auto-context alternating regression forest was proposed.
• Patient-specific anatomical features were extracted as voxel signatures from aligned training images.
• A regression forest was trained using the most relevant features to correct CBCT images of new patients.

Main Results:

• The algorithm was evaluated on 10 patients' CBCT and CT datasets.
• Quantitative metrics including Mean Absolute Error (MAE), Peak Signal-to-Noise Ratio (PSNR), and Normalized Cross-Correlation (NCC) were used.
• The corrected CBCT images achieved a mean MAE of 16.66 HU, PSNR of 37.28 dB, and NCC of 0.98 compared to ground truth CT.

Conclusions:

• The developed learning-based method significantly improves CBCT image quality.
• The proposed technique demonstrates high accuracy in CBCT correction, approaching the quality of planning CT.
• This method holds great potential for enabling quantitative applications of CBCT in image-guided adaptive radiotherapy.