딥러닝을 이용한 정상인과 알츠하이머 환자의 MRI-to-FDG PET 영상 변환 연구

Abstract

Alzheimer's disease is a subtype of dementia, representing a prevalent condition in an aging society. As of now, there is no cure for this condition, emphasizing the importance of early intervention to impede its progression. In the pursuit of proactive treatment at the initial stages, it is crucial to identify various biomarkers associated with Alzheimer's disease. Prior to MRI imaging, FDG PET stands out as a biomarker capable of early detection. Therefore, the objective of this study is to transform MRI images into FDG PET images. This research aims to transform paired images of MRI and FDG PET, among various biomarkers associated with Alzheimer's disease, with a particular focus on optimizing the generation of FDG PET images from MRI. To achieve this, image transformation experiments were conducted using the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset, utilizing CycleGAN when the MRI and FDG PET images were captured within 180 days of each other. The quantitative metrics, including MAE, MSE, SSIM, and PSNR, indicate successful image generation compared to the baseline model. The MAE for the generated images versus the original images across all groups was consistently 0.011 ± 0.008, with AD, MCI, and CN recording values of 0.011 ± 0.007, 0.011 ± 0.008, and 0.011 ± 0.007, respectively. For MSE, the results were 0.002 ± 0.002 for all groups, with AD, MCI, and CN showing values of 0.002 ± 0.002 each. SSIM values were consistent across all groups at 0.987 ± 0.004, with AD at 0.987 ± 0.003, MCI at 0.986 ± 0.004, and CN at 0.988 ± 0.004. Notably, PSNR was 33.837 ± 3.951 across all groups, with values of 33.606 ± 3.929 for AD, 34.415 ± 4.044 for MCI, and 34.415 ± 4.044 for CN. These outcomes suggest that our study can generate images with features akin to FDG PET images, reflecting the functional aspects of the brain using MRI images. However, it is essential to acknowledge limitations, such as the inclusion of only cognitively intact individuals in the image generation experiment and the absence of qualitative assessments conducted by specialists. Nevertheless, despite these limitations, this study paves the way for future research, anticipating further exploration into the validity of transforming FDG PET images through MRI. Additionally, there is potential for research into the transformation of images related to the accumulation of amyloid. Furthermore, the application of this technique to neurodegenerative disorders detectable by FDG PET holds promising prospects.