Learning-based CT Perfusion Image Denoising with Only Noisy Training Data

Computed tomography perfusion (CTP) imaging measures the hemodynamics of the brain parenchyma, which provides important information for acute ischemic stroke diagnosis. CTP images are very noisy due to the low dose in the sequential scans. Conventional denoising methods significantly reduce the spatial resolution, which leads to low sensitivity to small infarct cores. Despite of the success of deep neural networks in medical image denoising and reconstruction, they require high quality reference images for training, which are not available in CTP. In this work, we proposed a network-based method for CTP denoising which only required noisy data. Rather than mapping a noisy image to its clean version, the network was learned by mapping the noisy image to its estimation with independent noise. The noise-to-noise training was equivalent to learning noise-to-clean mapping with independent noises. The method was trained and tested on ISLES 2018 dataset and demonstrated improved spatial resolution, noise level and contrast to noise ratio compared to existing denoising methods.