Deep learning-based multi-modality image conversion for evaluating dose calculation and position correction accuracy...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiation Oncology

Background:

• Cone-beam CT (CBCT) is crucial for image-guided radiation therapy (IGRT) but suffers from limited image quality.
• Poor CBCT image quality can compromise dose calculation and positional accuracy, hindering adaptive radiation therapy (ART).

Purpose of the Study:

• To develop a deep learning framework for multi-modality image synthesis to improve ART and IGRT feasibility.
• To generate synthetic CT (sCT) from CBCT and synthetic MRI (sMRI) from CT using conditional generative adversarial networks (CGANs).

Main Methods:

• Two CGAN models were developed: CBCT-to-CT for sCT generation and CT-to-MRI for sMRI generation.
• Models were trained on 80 cases and validated on 20 cases, evaluating image quality with SSIM and PSNR.
• Dose distributions were recalculated using sCT, and positional accuracy of sMRI was assessed via registration to original CT.

Main Results:

• The CBCT-to-CT model significantly improved SSIM (0.15 to 0.82) and PSNR (11.3 to 19.9 dB) for sCT.
• The CT-to-MRI model enhanced SSIM (0.09 to 0.63) and PSNR (10.3 to 23.9 dB) for sMRI.
• sCT-based dose distribution achieved >95% gamma pass rate, and sMRI positional corrections were within 1 mm.

Conclusions:

• Conditional generative adversarial networks (CGANs) enable high-fidelity multi-modality image synthesis from CBCT.
• This deep learning approach improves both dose calculation and position correction accuracy, enhancing adaptive radiation therapy (ART) and image-guided radiation therapy (IGRT).