Synthetic CT generation from CBCT using deep learning for adaptive radiotherapy in prostate cancer

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Conventional Cone Beam Computed Tomography (CBCT) provides images during radiotherapy but lacks the accuracy of planning CT (pCT) for dose calculations.
• Generating synthetic CT (sCT) from CBCT can bridge this gap, enabling adaptive radiotherapy.
• Accurate sCT generation is crucial for improving treatment planning and patient outcomes in prostate cancer management.

Purpose of the Study:

• To evaluate the accuracy of deep learning models (U-Net, ResU-Net) for synthetic CT (sCT) generation from CBCT in prostate cancer patients.
• To assess the clinical applicability of generated sCTs for treatment planning.
• To investigate the potential of sCTs to support adaptive radiotherapy decision-making.

Main Methods:

• Retrospective analysis of 50 CBCT-CT mappings from 10 prostate cancer patients.
• Training U-Net and ResU-Net models using PyTorch on preprocessed and anatomically matched CBCT images.
• Quantitative comparison of generated sCTs against planning CTs using metrics like SSIM, PSNR, and MAE.

Main Results:

• Both U-Net and ResU-Net significantly improved image similarity (SSIM) and signal-to-noise ratio (PSNR) compared to CBCT.
• ResU-Net achieved a statistically significant higher PSNR than U-Net.
• Mean Absolute Error (MAE) was significantly reduced by both models, with ResU-Net showing a slightly lower error (61.8 HU) compared to U-Net (65.3 HU).

Conclusions:

• Deep learning models, specifically U-Net and ResU-Net, offer effective and feasible solutions for generating clinically relevant sCTs from CBCT.
• The generated sCTs support accurate dose calculations, crucial for effective radiotherapy.
• These models facilitate adaptive radiotherapy workflows in prostate cancer management.