A Comprehensive Evaluation of Bidirectional MRI-CT Generation Methods Considering Three Factors: Dataset, Model, and Training Strategy

Precise imaging is essential in medical applications, including disease diagnosis, treatment planning, and patient monitoring. Magnetic Resonance Imaging (MRI) and computed tomography (CT) are common imaging modalities in these applications, but each has limitations. MRI excels in distinguishing between soft tissues but lacks electron density information, hindering accurate diagnosis. Conversely, CT provides electron density information, but its poor contrast makes distinguishing soft tissues challenging. However, acquiring both MRI and CT scans from a patient is redundant and exposes unnecessary radiation. To address this challenge, we propose a bidirectional MRI-CT generation method determined through a comprehensive evaluation considering the dataset type, model architecture, and training strategy. Our novel semi-supervised dual-training strategy enables simultaneous training on unpaired and paired images. Based on our findings, we suggest pre-training the UVCGAN 2 model on the large-scale RadImageNet dataset, followed by traditional fine-tuning and applying the proposed dual-training strategy, which has been shown to improve training stability and generation quality. The proposed method achieves the best results with metrics: MAE \(_{sMRI}\) = 0.04, MAE \(_{sCT}\) = 0.01, PSNR \(_{sMRI}\) = 23.87, PSNR \(_{sCT}\) = 28.97, SSIM \(_{sMRI}\) = 0.78, and SSIM \(_{sCT}\) = 0.87. Furthermore, we developed a graphical user interface (GUI) to analyze registered image pairs efficiently. The GUI and code are publicly available at https://github.com/HC94/bidirectional_mri_ct_generation .