CT Brain Image Synthesization from MRI Brain Images Using CycleGAN

Kazuki Takamiya, Yutaro Iwamoto, Masahiro Nonaka, Yen-Wei Chen

Abstract: Medical images from multiple modalities are essential for diagnosis and effective treatment of various diseases. However, obtaining these images separately is a time and cost consuming task for both patients and physicians. Whereas CT scanning also increases the radiation exposure for patients. Therefore, generating CT scans from radiation-free MR images is a very desirable task. Synthesized brain CT scans are useful for obtaining cranial information that is difficult to obtain from MR imaging but gives crucial information for brain surgery and treatment. In this study, we propose a deep learning-based method to synthesize CT image from MR image. We present a cycle GAN-based method to synthesize CT brain image from MR brain image. We also propose a novel normalization method called as range-of-interest (ROI) normalization to emphasize the tissue and bone regions. Using this method, the synthesized CT image can be treated in the same way as the actual CT image. Additionally, we evaluate the impact of intensity normalization using three different intensity normalizations for synthesizing MR-CT images, which is an important image preprocessing step. Our proposed ROI normalization for CT images and max-min normalization for MR images generate the highest-quality synthetic image results, with Mean Absolute Error at approximately 94.60 HU of the produced 3D volume of CT images.