Learning-based CBCT correction using alternating random forest based on auto-context model.

Area of Science:

• Medical Imaging
• Radiotherapy Technology
• Machine Learning in Healthcare

Background:

• Quantitative Cone Beam CT (CBCT) is crucial for image-guided radiotherapy, enabling accurate treatment setup and dose calculations.
• Current CBCT image quality limitations restrict its use primarily to patient setup.
• Advanced applications require significant improvements in CBCT image fidelity.

Purpose of the Study:

• To develop and evaluate a learning-based approach for enhancing CBCT image quality.
• To extend the clinical utility of CBCT beyond patient setup in radiotherapy.

Main Methods:

• An auto-context model integrated into a machine learning framework was used.
• Data preprocessing involved noise reduction and image alignment.
• Random forests were trained on image patches, with iterative refinement enhancing corrected CBCT (CCBCT) quality.

Main Results:

• The algorithm was validated on brain and pelvis CBCT data from 26 patients.
• Quantitative metrics including Mean Absolute Error (MAE), Peak Signal-to-Noise Ratio (PSNR), Normalized Cross-Correlation (NCC), and Spatial Nonuniformity (SNU) demonstrated significant image quality improvement.
• Specific results showed mean MAE of 12.81 HU (brain) and 19.94 HU (pelvis), mean PSNR of 40.22 dB (brain) and 31.31 dB (pelvis), mean NCC of 0.98 (brain) and 0.95 (pelvis), and SNU of 2.07% for both.

Conclusions:

• The proposed learning-based correction method significantly improves CBCT image quality.
• This advancement holds substantial potential for enhancing CBCT's role in adaptive radiotherapy.
• The method is poised to expand clinical applications of CBCT in radiation oncology.