MRI-Only Radiotherapy Planning for Nasopharyngeal Carcinoma Using Deep Learning

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Radical radiotherapy is standard for nasopharyngeal carcinoma (NPC).
• MRI offers advantages over CT for soft-tissue visualization but lacks electron density (ED) data crucial for radiotherapy planning.
• Current methods require CT, introducing additional radiation exposure.

Purpose of the Study:

• To develop and evaluate a pseudo-CT (pCT) generation method using deep learning for MRI-only radiotherapy planning in NPC.
• To provide essential electron density (ED) information from MRI data.
• To enable precise dose calculations without relying on conventional CT scans.

Main Methods:

• A U-Net deep learning model and a pix2pix generative adversarial network (GAN) were trained using 1433 paired MR and CT images from 20 NPC patients.
• Rigid registration and preprocessing were applied to the image datasets.
• Target volumes and organs at risk contours were transferred to the generated pCT for dose calculation and comparison with original CT-based plans.

Main Results:

• The pCT generation process was rapid, averaging 7.90 ± 0.47 seconds per patient.
• The U-Net and pix2pix GAN models showed comparable pCT quality, with no significant statistical difference.
• Dose distributions calculated on pCT were highly consistent with original CT, evidenced by a 99.1% ± 0.3% mean gamma pass rate and minimal differences in critical dose metrics.

Conclusions:

• The proposed deep learning model accurately generates pCT from MRI, suitable for precise dose calculations in NPC radiotherapy.
• Implementing MRI-only planning can enhance structure delineation and significantly reduce radiation dose.
• This approach is particularly beneficial for MR-guided linear accelerators in NPC treatment.