Contrast-Free Ischemic Stroke Assessment: Generating CBV Maps from NCCT Scans Using Deep Learning Models

Cerebral blood volume (CBV) is a pivotal parameter in computed tomography perfusion (CTP) imaging techniques for assessing cerebral perfusion, giving insights into blood distribution across different brain regions. Traditionally, CTP involves the use of a contrast agent, which poses significant risks for vulnerable patient groups, despite the precision it offers. In the absence of contrast-based CT perfusion imaging techniques, generating CBV maps from non-contrast CT (NCCT) scans allows for the quick and precise diagnosis of ischemic stroke and other vascular disorders. To generate CBV images from NCCT scans, we perform a comparative evaluation of multiple transfer learning and diffusion models. For enhancing pixel-wise accuracy and perceptual quality, we propose a novel hybrid loss function combining mean squared error (MSE) and multiscale edge structural similarity index measure (MSE-SSIM) with adaptive weighting and wavelet similarity feature. From our extensive experiments, we observe that the latent diffusion model (LDM), trained using our hybrid loss function, exhibits superior performance in generating CBV maps from NCCT images with 85.20% MSE-SSIM, and 20.34dB PSNR.