A systematic review of deep learning techniques for generating synthetic CT images from MRI data

Introduction: This systematic review evaluates various studies on deep learning algorithms for generating synthetic CT images from MRI data, focusing on challenges in image quality and accuracy in current synthetic CT generation methods. Magnetic resonance imaging (MRI) is increasingly important in clinical settings due to its detailed visualization and noninvasive nature, making it a valuable tool for advancing patient care and identifying new areas for research. Materials and Methods: In this study we conducted a thorough search across several databases to identify studies published between January 2009 and January 2024 on using deep learning to generate synthetic CT (sCT) images from MRI for radiotherapy. The review focused on peer-reviewed, English-language studies and excluded unpublished, non-English, and irrelevant studies. Data on deep learning methods, input modalities, and anatomical sites were extracted and analyzed using a result-based synthesis approach. The review categorized 84 studies by anatomical site, following PRISMA guidelines for summarizing the findings. Results: The U-Net model is the most frequently used deep learning model for generating synthetic CT images from MRI data, with 34 articles highlighting its effectiveness in capturing fine details, Conditional GANs are also widely used, while Cycle-GANs and Pix2pix are effective in image translation tasks. Significant differences in performance metrics, such as MAE and PSNR, were observed across anatomical regions and models, highlighting the variability in accuracy among different deep learning approaches. Conclusion: This review underscores the need for continued refinement and standardization in deep learning approaches for medical imaging to address variability in performance metrics across anatomical regions and models.