Synthetic CT reconstruction using a deep spatial pyramid convolutional framework for MR-only breast radiotherapy.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiation Oncology

Background:

• Magnetic resonance imaging (MRI) offers superior soft-tissue contrast over computed tomography (CT), driving its adoption in MRI-guided radiation therapy (MR-IGRT).
• The development of MRI-based therapy systems fuels interest in MRI-only workflows, necessitating synthetic CT (sCT) generation from MRI for treatment planning and dose calculations.

Purpose of the Study:

• To propose and evaluate a novel deep spatial pyramid convolutional framework for MRI-to-CT image translation for sCT generation.
• To compare the proposed framework's performance against the established U-Net architecture within a generative adversarial network (GAN) framework.

Main Methods:

• Utilized atrous spatial pyramid pooling (ASPP) with atrous convolution to capture multi-scale features efficiently and reduce model parameters.
• Developed a generative model with stacked encoders and decoders incorporating the ASPP module.
• Compared training time, image quality (RMSE, SSIM, PSNR), and dosimetric accuracy of the proposed framework against a conventional GAN framework.

Main Results:

• The proposed framework demonstrated significant reductions in training time and improvements in image quality across various training data set sizes compared to the conventional framework.
• Achieved excellent image quality metrics on 1042 test images: RMSE (17.7 ± 4.3 HU), SSIM (0.9995 ± 0.0003), and PSNR (71.7 ± 2.3).
• Dose distributions calculated from generated sCT achieved >98% passing rates using the 3D gamma index (2%/2 mm criterion).

Conclusions:

• The deep spatial pyramid convolutional framework offers superior performance for sCT generation compared to conventional GANs.
• This method represents a crucial step towards enabling MRI-only radiation therapy workflows.
• The proposed framework facilitates broader clinical applications of MR-IGRT, including online adaptive therapy.