Unsupervised Learning-Based Cross-Modal Transformation of Brain MR-CT Image
Yongde Guo, Haoran Hao · 2023
Medical imaging technology serves as a crucial instrument for disease screening and medical diagnosis in clinical medicine, with computed tomography (CT) and magnetic resonance imaging (MRI) standing as prevalent and significant diagnostic imaging techniques. However, the limitations of medical imaging acquisition and patient-specific factors can hinder medical images acquisition, resulting in certain challenges for clinical diagnoses. Furthermore, MR image-guided radiotherapy necessitates the synthesis of pseudo-CT images from MR images for radiotherapy dose estimation. Consequently, there has led to a growing interest among researchers in the research of a cross-modal transformation between MR and CT images. In clinical medicine, the limited volume of aligned medical images data and the systematic errors in the alignment process, while non-aligned medical images data are more in line with the actual clinical environment, make unsupervised learning-based cross-modal conversion of medical images gradually become a research focus. In light of this, we develop a novel cycle-consistent adversarial networks (CycleGAN) model for the unsupervised learning-based cross-modality transformation task of brain MR-CT images. Within this model structure, we propose an Att-UCTransformer generator and an HDD-PatchGAN discriminator. The proposed model undergoes quantitative analysis and demonstrates more promising results than the original CycleGAN model in three evaluation metrics, namely MAE, PSNR, and SSIM, effectively enhancing the accuracy of unsupervised learning-based cross-modality conversion tasks for medical images.