Comparative analysis of deep learning models for the generation of synthetic CT from MRI for radiotherapy dose planning

Abstract

Introduction: Radiotherapy aims to precisely target tumors while sparing healthy tissue, traditionally relying on CT imaging for accurate dose planning. However, CT has limitations in soft tissue contrast and exposes patients to ionizing radiation. MRI offers superior soft tissue contrast without radiation but lacks electron density information, restricting its use in dose planning. This study addresses this gap by developing deep learning models to generate pseudo-CT images from MRI, enabling MRI-only workflows in radiotherapy.

Methodology: Paired MRI and CT scans from 12 subjects were processed using normalization, alignment, and masking. Four deep learning architectures (U-Net, Pix2Pix, CycleGAN, and conditional GAN (cGAN)) were trained to generate synthetic CT images from MRI data. Model performance was evaluated using metrics including mean absolute error (MAE), mean squared error (MSE), peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and Pearson correlation coefficient (PCC).

Results: Pix2Pix achieved the highest SSIM and PSNR, indicating strong structural preservation and reduced noise. It also had the lowest MAE and MSE, showing high accuracy in synthetic-CT generation. The cGAN model scored highest in PCC, highlighting its effective intensity alignment with real CT data. Statistical tests confirmed Pix2Pix’s superior performance, though CycleGAN and cGAN also showed notable results in alignment accuracy.

Conclusion: Deep learning models, particularly Pix2Pix, can generate reliable pseudo-CT images from MRI, supporting MRI-only radiotherapy planning. This approach reduces radiation exposure and may streamline radiotherapy workflows, offering a promising advance for patient-centered cancer