MV CBCT-Based Synthetic CT Generation Using a Deep Learning Method for Rectal Cancer Adaptive Radiotherapy

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Online Megavoltage cone beam CT (MV CBCT) offers real-time patient anatomy but suffers from image quality limitations, hindering its use in adaptive radiotherapy (ART).
• Current MV CBCT image quality and Hounsfield Unit (HU) accuracy are insufficient for critical applications like ART in rectal cancer treatment.
• Rectal cancer patients treated with intensity-modulated radiation therapy (IMRT) require precise anatomical information for effective ART.

Purpose of the Study:

• To improve MV CBCT image quality and HU accuracy for rectal cancer patients using a deep learning approach.
• To generate synthetic CT (sCT) images from MV CBCT that are suitable for adaptive radiotherapy (ART).
• To evaluate the performance of the generated sCT in terms of image quality, autosegmentation, and dose calculation.

Main Methods:

• A cycle-consistent adversarial network (CycleGAN) was employed to enhance MV CBCT images.
• The model was trained using CT and MV CBCT data from 30 rectal cancer patients undergoing IMRT.
• The generated sCT images were evaluated on 10 independent patients for image quality (MAE, SSIM), autosegmentation accuracy (Dice Similarity Coefficient - DSC), and dose calculation accuracy.

Main Results:

• The CycleGAN model significantly reduced the Mean Absolute Error (MAE) from 135.84 HU (CT-CBCT) to 52.99 HU (CT-sCT).
• Structural Similarity Index (SSIM) improved from 0.44 (CT-CBCT) to 0.81 (CT-sCT), indicating superior image quality.
• Autosegmentation for femoral heads on sCT required minimal manual correction, with high DSC values (0.93 for CTV, 0.94 for bladder) for other structures.
• Dose calculations based on sCT showed smaller deviations compared to CBCT-based plans, approaching CT-based plan accuracy.

Conclusions:

• The deep learning-based CycleGAN method effectively improves MV CBCT image quality and HU accuracy for rectal cancer patients.
• The generated synthetic CT (sCT) is suitable for adaptive radiotherapy (ART), enabling treatment based on actual patient anatomy.
• This approach overcomes a key limitation for implementing ART in rectal cancer treatment using online MV CBCT data.