A high-performance method of deep learning for prostate MR-only radiotherapy planning using an optimized Pix2Pix...

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• MRI-only radiotherapy planning eliminates CT-related artifacts and reduces patient burden.
• Accurate synthetic CT (sCT) generation from MRI is crucial for dose calculation in MRI-only workflows.
• Conditional Generative Adversarial Networks (cGANs) show promise for synthesizing medical images.

Purpose of the Study:

• To optimize a Pix2Pix cGAN model for generating synthetic CT (sCT) images from MRI for prostate radiotherapy planning.
• To compare the performance of the optimized Pix2Pix model against five other sCT generation architectures.
• To evaluate image quality and dosimetric accuracy of sCT images.

Main Methods:

• T2-weighted MRI and CT images from 39 prostate cancer patients were used.
• The Pix2Pix cGAN model was tuned by optimizing generators, loss functions, and hyperparameters.
• sCT images were evaluated using Mean Absolute Error (MAE), Mean Error (ME), and 3D gamma analysis against reference CT (rCT).
• Dose-volume histogram (DVH) metrics were compared between sCT and rCT.

Main Results:

• The optimized Pix2Pix model (Perceptual loss, ResNet 9 blocks) achieved the lowest MAE across different regions (e.g., 13.4 HU for bladder).
• Highest gamma passing rates (99.4% at 1%/1mm) were obtained with the Pix2Pix model.
• Mean errors for DVH metrics were minimal (e.g., -0.2% for PTV V95%, 0.1% for rectum V70Gy, -0.1% for bladder V50Gy).

Conclusions:

• The Pix2Pix-based sCT generation method yields superior image quality compared to other deep learning approaches.
• sCT images generated using Pix2Pix demonstrate comparable dose uncertainties to reference CT, validating its use in MRI-only radiotherapy.
• This optimized Pix2Pix model offers a robust solution for MRI-only prostate radiotherapy planning.