Inter-translation Between CT and MRI Brain Scans Based on Cycle Consistent Adversarial Networks
Haoran Li · Applied and Computational Engineering · 2023
Since Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) are often used together in diagnosing brain diseases, it is rather time-consuming and expensive for patients to obtain two scans simultaneously. In this paper, the author proposed a new idea for CT-MRI inter-translation, using dataset of human brain scans to achieve unpaired image-to-image translation between CT scans and MRI scans. More specifically, the author proposed using Cycle Consistent Adversarial Networks (CycleGAN) to realize the idea of style transfer between CT and MRI. Briefly, this paper had trained two sets of generator and discriminator to form a “cycle”. This model can retain the main characteristics of a scan while transferring the scan’s style. To keep this translation process “cycle-consistent”, the 〖Loss〗_cyc is used to keep the main content. Furthermore, the 〖Loss〗_GAN ensures that the generated images exhibit a close stylistic resemblance to the target domain and the 〖Loss〗_identity guarantees that the hue of the generated images is retained during the translation process. In addition, the author has added weights to the 〖Loss〗_cyc and the 〖Loss〗_identity to help the model perform better. Finally, this paper visualize the training process by presenting the original images and generated images together. Experimental results indicates that CycleGAN achieves a competitive performance can hopefully be a good auxiliary in brain diseases diagnosis.