Synthesis of ADC and DWI stroke sequences from onset NCCT studies following a structural cycle representation
摘要
Timely characterization of ischemic stroke is crucial for providing endovascular or pharmacological treatment. Current stroke care protocols include NCCT as first imaging support but exhibit low differentiation between healthy and affected tissue. Complementary, some protocols include MRI ADC/DWI finding observations, but this modality has relatively low availability.
Purpose:To generate synthetic ADC and DWI sequences from NCCT studies, maintaining textural characteristics of ischemic lesions.
Methods:From a cohort of ischemic stroke patients, we introduced a multi-domain generative representation to approximate ADC/DWI sequences from NCCT onset studies. The proposed approach explores a cycle consistency loss over an architecture of stacked U-Net blocks that guarantee textural stroke lesion preservation.
Results:The proposed approach was validated in a retrospective study of 96 patients, recovering synthetic MRI sequences with an average SSIM/PSNR score of 0.81/26.19. Also, a specific validation around stroke lesion achieved an average SSIM/PSNR score of 0.51/28.8. The proposed approach outperformed a standard CycleGAN and a StarGAN architecture, evidencing significant statistical differences (
The proposed method is able to recover an approximation to MRI (ADC/DWI) sequences from NCCT scans, while potentially avoiding the current delay in clinical practice and providing a first approximation of lesion patterns observed from ADC and DWI studies.